Papers with computational overhead
Iterative Structured Pruning for Large Language Models with Multi-Domain Calibration (2026.eacl-industry)
Copied to clipboard
| Challenge: | Existing models with unstructured pruning often yield irregular sparsity patterns that necessitate specialized hardware or software support. |
| Approach: | They propose a structured pruning framework that eliminates entire architectural components and maintains compatibility with standard hardware accelerators. |
| Outcome: | The proposed model pruning framework achieves significant compression with minimal performance degradation on multiple models across diverse downstream tasks. |
MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing (2026.acl-industry)
Copied to clipboard
Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, Linke Ouyang, Zhiyuan Zhao, Tao Chu, Tianyao He, Fan Wu, Qintong Zhang, Zhenjiang Jin, Guang Liang, Rui Zhang, Wenzheng Zhang, Yuan Qu, Zhifei Ren, Yuefeng Sun, Zirui Tang, Boyu Niu, Yuanhong Zheng, Dongsheng Ma, Ziyang Miao, Hejun Dong, Siyi Qian, Junyuan Zhang, Fangdong Wang, Jingzhou Chen, Xiaomeng Zhao, Liqun Wei, Wei Li, Shasha Wang, RuiLiang Xu, Yuanyuan Cao, Lu Chen, Qianqian Wu, Huaiyu Gu, Lindong Lu, Dechen Lin, null Shenguanlin, Xuanhe Zhou, Linfeng Zhang, Yuhang Zang, Xiaoyi Dong, Jiaqi Wang, Bo Zhang, Lei Bai, Pei Chu, Weijia Li, Jiang Wu, Lijun Wu, Zhenxiang Li, Guangyu Wang, Zhongying Tu, Chao Xu, Kai Chen, Bowen Zhou, Dahua Lin, Wentao Zhang, Conghui He
| Challenge: | Document images are characterized by higher resolutions, denser content, and more complex structural layouts. |
| Approach: | They propose a 1.2B-parameter document parsing vision-language model that decouples layout analysis from local content recognition. |
| Outcome: | The proposed model surpasses general-purpose and domain-specific models on multiple benchmarks while maintaining significantly lower computational overhead. |
Automatic Rule Induction for Efficient Semi-Supervised Learning (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to generalize from labeled and unlabeled data are difficult to explain and behave unreliably. |
| Approach: | They propose a framework for automatic discovery and integration of symbolic rules into pretrained transformer models by using an attention mechanism. |
| Outcome: | The proposed framework can improve state-of-the-art methods with no manual effort and minimal computational overhead. |
Rethinking KenLM: Good and Bad Model Ensembles for Efficient Text Quality Filtering in Large Web Corpora (2025.acl-short)
Copied to clipboard
| Challenge: | Existing methods to efficiently filter large web corpora require GPU resources. |
| Approach: | They propose an ensemble approach that leverages two contrasting KenLMs to filter large web corpora. |
| Outcome: | The proposed method significantly reduces noisy content while preserving high-quality content compared to the traditional KenLM training method. |
Fusion Makes Perfection: An Efficient Multi-Grained Matching Approach for Zero-Shot Relation Extraction (2024.naacl-short)
Copied to clipboard
| Challenge: | Existing methods to extract unseen relations require laborious manual annotation . a new approach uses fine-grained matching to reduce manual annotation cost . |
| Approach: | They propose an efficient multi-grained matching approach that uses virtual entity matching to reduce manual annotation cost. |
| Outcome: | The proposed approach outperforms the state-of-the-art methods and achieves inference efficiency and accuracy in zero-shot relation extraction tasks. |
Unifying Language Agent Algorithms with Graph-based Orchestration Engine for Reproducible Agent Research (2025.acl-demo)
Copied to clipboard
Qianqian Zhang, Jiajia Liao, Heting Ying, Yibo Ma, Haozhan Shen, Jingcheng Li, Peng Liu, Lu Zhang, Chunxin Fang, Kyusong Lee, Ruochen Xu, Tiancheng Zhao
| Challenge: | Language agents powered by large language models (LLMs) have demonstrated remarkable capabilities in understanding, reasoning, and executing complex tasks. |
| Approach: | They propose a flexible framework that addresses engineering overhead and insufficient evaluation frameworks for fair comparison. |
| Outcome: | The proposed framework simplifies language agent development and establishes a foundation for reproducible agent research. |
Derailer-Rerailer: Adaptive Verification for Efficient and Reliable Language Model Reasoning (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing prompting methods struggle with complex tasks and reasoning stability, limiting their practical deployment. |
| Approach: | They propose a framework that adaptively balances reasoning accuracy and computational efficiency by employing a lightweight Derailer mechanism to assess reasoning stability and selectively triggers an advanced Rerailer verification process only when necessary. |
| Outcome: | The proposed framework achieves significant accuracy improvements (8-11%) while maintaining 2-3 times better efficiency than existing verification methods. |
Resource-Efficient Anonymization of Textual Data via Knowledge Distillation from Large Language Models (2025.coling-industry)
Copied to clipboard
| Challenge: | Existing approaches to anonymize textual data from large language models pose privacy risks due to their API-based access. |
| Approach: | They propose a method to distill large language models into smaller encoder-only models via named entity recognition coupled with regular expressions to create a lightweight model capable of effective anonymization. |
| Outcome: | The proposed approach reduces computational overhead while maintaining semantic integrity of data. |
HyperEdit: Unlocking Instruction-based Text Editing in LLMs via Hypernetworks (2026.findings-acl)
Copied to clipboard
Yiming Zeng, Jinghan Cao, Zexin Li, Wanhao Yu, Zhankai Ye, Dawei Xiang, Ting Hua, Xin Liu, Shangqian Gao, Tingting Yu
| Challenge: | Existing approaches treat instruction-based text editing as a generic text generation problem. Existing methods either over-edit or fail to apply modifications consistently. |
| Approach: | They propose a framework that processes each editing request to best align with it. |
| Outcome: | The proposed framework achieves 9% improvement over the state-of-the-art model. |
Towards Hierarchical Multi-Step Reward Models for Enhanced Reasoning in Large Language Models (2026.findings-acl)
Copied to clipboard
Teng Wang, Jiang Zhangyi, Zhenqi He, Hailei Gong, Shenyang Tong, Wenhan Yang, Zeyu Li, Yanan Zheng, Zifan He, Zewen Ye, Shengjie Ma, Jianping Zhang
| Challenge: | Existing Process Reward Models (PRMs) are vulnerable to reward hacking and require expensive, large-scale annotation of reasoning steps. |
| Approach: | They propose a reward model approach which evaluates both individual and consecutive reasoning steps from fine-grained and coarse-grounded level. |
| Outcome: | Empirical results show that the proposed model performs better than existing PRMs and is more robust than existing models. |
On-device System of Compositional Multi-tasking in Large Language Models (2025.emnlp-industry)
Copied to clipboard
Ondrej Bohdal, Konstantinos Theodosiadis, Asterios Mpatziakas, Dimitrios Filippidis, Iro Spyrou, Christos Zonios, Anastasios Drosou, Dimosthenis Ioannidis, Kyenghun Lee, Jijoong Moon, Hyeonmok Ko, Mete Ozay, Umberto Michieli
| Challenge: | Existing approaches to generative AI for large language models struggle when executing complex tasks simultaneously. |
| Approach: | They propose a novel approach tailored specifically for compositional multi-tasking scenarios . they add a learnable projection layer on top of the combined summarization and translation adapters. |
| Outcome: | The proposed approach performs well and is fast in both cloud-based and on-device implementations. |
Adversarial Text Normalization (2022.naacl-industry)
Copied to clipboard
| Challenge: | Text-based adversarial attacks are becoming more commonplace and accessible to general internet users. |
| Approach: | They propose a method that restores baseline performance on attacked content with low computational overhead. |
| Outcome: | The proposed method restores baseline performance on attacked content with low computational overhead. |
NEST: Nested Evidence Survival for Retrieval (2026.acl-industry)
Copied to clipboard
| Challenge: | Existing approaches to retrieval-augmented generation (RAG) rely on rigid heuristics or computational overhead. |
| Approach: | They propose a lightweight, training-free RAG framework that separates recall amplification from precision selection. |
| Outcome: | Evaluated on WebQuestions, HotpotQA and internalQA benchmarks, NEST outperforms strong adaptive RAG baselines. |
AdaBERT-CTC: Leveraging BERT-CTC for Text-Only Domain Adaptation in ASR (2023.emnlp-industry)
Copied to clipboard
| Challenge: | End-to-end (E2E) automatic speech recognition models struggle to recognize out-of-domain words such as proper nouns and domain-specific terms. |
| Approach: | They propose a domain adaptation technique that relies solely on textual data to adapt to out-of-domain words. |
| Outcome: | The proposed method outperforms the base model by up to 14% relative word error rate improvement on several out-of-domain, publicly available datasets. |
SFAL: Semantic-Functional Alignment Scores for Distributional Evaluation of Auto-Interpretability in Sparse Autoencoders (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Interpreting the internal representations of large language models (LLMs) is crucial for their deployment in real-world applications, impacting areas such as AI safety, debugging, and compliance. |
| Approach: | They propose an alternative evaluation strategy that assesses the alignment between the semantic neighbourhoods of features and their functional neighbourhoods by using co-occurrence statistics. |
| Outcome: | The proposed evaluation strategy reduces reliance on scoring on large-scale models and improves efficiency and cost-effectiveness. |
ALToolbox: A Set of Tools for Active Learning Annotation of Natural Language Texts (2022.emnlp-demos)
Copied to clipboard
Akim Tsvigun, Leonid Sanochkin, Daniil Larionov, Gleb Kuzmin, Artem Vazhentsev, Ivan Lazichny, Nikita Khromov, Danil Kireev, Aleksandr Rubashevskii, Olga Shahmatova, Dmitry V. Dylov, Igor Galitskiy, Artem Shelmanov
| Challenge: | Currently, the framework supports text classification, sequence tagging, and seq2seq tasks. |
| Approach: | They propose an open-source framework for active learning annotation in natural language processing that provides an easy-to-deploy GUI annotation tool directly in the Jupyter IDE. |
| Outcome: | The proposed framework reduces computational overhead and duration of AL iterations and increases annotated data reusability. |
A Unified Encoder-Decoder Framework with Entity Memory (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to index, retrieve, and read documents as evidence suffer from large computational overheads. |
| Approach: | They propose an encoder-decoder framework with an entity memory that stores entity knowledge as latent representations and pre-trained on Wikipedia along with encoder parameters. |
| Outcome: | The proposed framework outperforms memory-based and non-memory encoder-decoder models on various entity-intensive question answering and generation tasks. |
Surface-Based Retrieval Reduces Perplexity of Retrieval-Augmented Language Models (2023.acl-short)
Copied to clipboard
| Challenge: | Augmenting language models with a retrieval mechanism has been shown to improve performance while keeping the number of parameters low. |
| Approach: | They propose to augment language models with a retrieval mechanism by replacing semantic retrieval with BM25 . they find that the model's performance is better explained by surface-level similarities, they say . |
| Outcome: | The proposed method reduces perplexity and lowers the number of parameters while keeping the number low. |
MergeIT: From Selection to Merging for Efficient Instruction Tuning (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for instruction tuning rely on LLMs to score instruction quality . existing methods rely only on Llms to rank instruction quality, but this approach is expensive and time-consuming . |
| Approach: | They propose a novel LLM-based Merging strategy for better Instruction Tuning that shifts the focus from selection to synthesis. |
| Outcome: | The proposed method reduces time and computational cost while preserving diversity and reducing redundancy. |
Assessing Factual Reliability of Large Language Model Knowledge (2024.naacl-long)
Copied to clipboard
| Challenge: | Factual knowledge of LLMs is typically evaluated using accuracy, yet this metric does not capture the vulnerability of LRMs to hallucination-inducing factors like prompt and context variability. |
| Approach: | They propose a metric designed to measure LLMs’ factual reliability by comparing the distance between the probability distributions of a valid output and its counterparts produced by the same LLM probing the same fact using different styles of prompts and contexts. |
| Outcome: | The proposed metric measures the distance between the probability distributions of a valid output and its counterparts produced by the same LLM probing the same fact using different styles of prompts and contexts. |
A Frustratingly Simple Decoding Method for Neural Text Generation (2024.lrec-main)
Copied to clipboard
| Challenge: | Neural text generation is notorious for repetitive loops and tedious outputs. |
| Approach: | They propose a method that penalizes future generation of repetitive content . they construct an anti-LM based on previously generated text . |
| Outcome: | The proposed method outperforms established baselines in terms of generation quality, decoding speed, and universality. |
Beyond Fine-tuning: Few-Sample Sentence Embedding Transfer (2020.aacl-main)
Copied to clipboard
| Challenge: | Fine-tuning (FT) pre-trained sentence embedding models on small datasets has been shown to have limitations. |
| Approach: | They propose to combine embeddings from a pre-trained model with a simple sentence embeddable model. |
| Outcome: | The proposed approach outperforms FT on small datasets with negligible computational overhead. |
Semantic Token Clustering for Efficient Uncertainty Quantification in Large Language Models (2026.eacl-short)
Copied to clipboard
| Challenge: | Large language models have limited truthfulness and tendency toward overconfidence constrain reliability in factual tasks. |
| Approach: | They propose an efficient method that leverages semantic information encoded in LLMs to quantify uncertainty. |
| Outcome: | The proposed method achieves comparable performance to baselines while significantly reducing computational overhead. |
CoLeM: A framework for semantic interpretation of Russian-language tables based on contrastive learning (2025.acl-srw)
Copied to clipboard
| Challenge: | Existing methods for semantic interpretation of tables in Russian lack explicit semantics . tabular data are one of the key formats for presenting structured information in various domains ranging from scientific research to business analytics. |
| Approach: | They propose a contrastive learning approach to minimize reliance on manual labeling and enhance the accuracy of column annotation for rare semantic types. |
| Outcome: | The proposed method minimizes reliance on manual labeling and improves column annotation accuracy for rare semantic types. |
SEA: Supervised Embedding Alignment for Token-Level Visual-Textual Integration in MLLMs (2025.emnlp-main)
Copied to clipboard
Yuanyang Yin, Yaqi Zhao, Yajie Zhang, Yuanxing Zhang, Ke Lin, Jiahao Wang, Xin Tao, Pengfei Wan, Wentao Zhang, Feng Zhao
| Challenge: | Multimodal Large Language Models (MLLMs) integrate visual and textual inputs, yet modality alignment remains one of the most challenging aspects. |
| Approach: | They propose a token-level supervision alignment method that enables more precise visual-text alignment during pretraining. |
| Outcome: | The proposed method improves performance across various model sizes, with smaller models benefiting the most. |
DisComp: A Two-Stage Prompt Optimization Framework Combining Task-Agnostic and Task-Aware Compression (2025.findings-naacl)
Copied to clipboard
| Challenge: | Extended prompts can lead to substantial computational overhead and increased hardware demands, limiting the scalability and efficiency of large language models. |
| Approach: | They propose a two-stage prompt compression framework that combines task-agnostic and task-based strategies to efficiently compress prompt length without compromising performance. |
| Outcome: | The proposed framework outperforms task-agnostic and task-specific compression methods on three benchmark datasets and is up to 6.56 faster at inference compared to the best token-level compression method. |
Better Late Than Never: Model-Agnostic Hallucination Post-Processing Framework Towards Clinical Text Summarization (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for generating concise and coherent summaries may include unintended text with hallucinations, causing computational costs. |
| Approach: | They propose a model-agnostic framework to post-process medical hallucinations . MEDAL integrates with any medical summarization model, requiring no additional computational overhead . |
| Outcome: | MEDAL can post-process medical hallucinations without additional computational overhead. |
PLEX: Adaptive Parameter-Efficient Fine-Tuning for Code LLMs using Lottery-Tickets (2025.naacl-industry)
Copied to clipboard
| Challenge: | PLEX is a lottery-ticket based parameter-efficient fine-tuning method that adapts large language models to well-supported and underrepresented programming languages (PLs) in pretraining. |
| Approach: | They propose a lottery-ticket based parameter-efficient fine-tuning method that adapts large language models to well-supported and underrepresented programming languages (PLs) |
| Outcome: | The proposed method achieves state-of-the-art performance among PEFT methods while maintaining competitive results with reduced computational overhead. |
LAGCL4Rec: When LLMs Activate Interactions Potential in Graph Contrastive Learning for Recommendation (2025.findings-emnlp)
Copied to clipboard
Leqi Zheng, Chaokun Wang, Canzhi Chen, Jiajun Zhang, Cheng Wu, Zixin Song, Shannan Yan, Ziyang Liu, Hongwei Li
| Challenge: | Traditional contrastive learning methods treat negative feedback as equally hard or easy, ignoring informative semantic difficulty during training. |
| Approach: | They propose a framework leveraging Large Language Models to Activate interactions in Graph Contrastive Learning for Recommendation. |
| Outcome: | The proposed framework outperforms state-of-the-art benchmarks on multiple benchmarks. |
Minimal Ranks, Maximum Confidence: Parameter-efficient Uncertainty Quantification for LoRA (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Low-Rank Adaptation (LoRA) enables parameter-efficient fine-tuning of large language models . standard LoRA lacks mechanisms for uncertainty quantification, leading to overconfident and poorly calibrated models. |
| Approach: | They propose a parameter-efficient Bayesian LoRA method that decomposes weight updates into low-rank matrices. |
| Outcome: | The proposed method achieves strong performance with improved calibration and generalization while maintaining computational efficiency. |
EFUF: Efficient Fine-Grained Unlearning Framework for Mitigating Hallucinations in Multimodal Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to eliminate hallucinations require expensive human annotation . hallucination in multimodal large language models poses unique challenges for current research . |
| Approach: | They propose a fine-grained unlearning framework that performs gradient ascent to eliminate hallucinations without paired data. |
| Outcome: | The proposed method reduces hallucinations while preserving quality with modest computational overhead. |
UniEDU: Toward Unified and Efficient Large Multimodal Models for Educational Tasks (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Existing research has focused on plain text, while real-world K-12 scenarios often involve multimodal data. |
| Approach: | They propose a unified language and vision assistant called UniEDU for educational applications . it excels across multiple educational tasks while maintaining strong generalization capabilities . authors propose to use UniEDu for industry-scale deployment . |
| Outcome: | The proposed model excels across multiple educational tasks while maintaining strong generalization capabilities. |
Iterative Refinement in the Continuous Space for Non-Autoregressive Neural Machine Translation (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing non-autoregressive inference procedures that refine in token space often require computational overhead. |
| Approach: | They propose an efficient inference procedure that iteratively refines translation purely in the continuous space using a latent variable instead of the latent variables. |
| Outcome: | The proposed procedure is twice as efficient and more effective than the existing EM-like inference procedure. |
POSITION BIAS MITIGATES POSITION BIAS: Mitigate Position Bias Through Inter-Position Knowledge Distillation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Positional bias (PB) manifests as non-uniform sensitivity across contextual locations . previous studies have addressed PB by modifying the underlying architectures or employing extensive contextual awareness training. |
| Approach: | They propose a position-to-position knowledge distillation framework that leverages position-induced disparities to counteract PB. |
| Outcome: | The proposed framework reduces positional bias and improves performance on retrieval and reasoning tasks. |
Hierarchical Policy Optimization for Simultaneous Translation of Unbounded Speech (2026.acl-long)
Copied to clipboard
| Challenge: | Existing synthesis methods cannot guarantee data quality. |
| Approach: | They propose a hierarchical reward that balances translation quality and latency objectives by combining supervised fine-tuning data with supervised inputs. |
| Outcome: | The proposed model can reuse key-value caches across both modalities and eliminate redundant feature recomputation. |
Improving Deep Transformer with Depth-Scaled Initialization and Merged Attention (D19-1)
Copied to clipboard
| Challenge: | Existing methods to improve NLP convergence and computational overhead are limited by stacking more layers. |
| Approach: | They propose a depth-scaled initialization method which reduces parameter variance at initialization and reduces output variance of residual connections to ease gradient back-propagation. |
| Outcome: | The proposed method outperforms the base model on translation tasks with five translation directions while matching the decoding speed of the baseline model. |
MemoPhishAgent: Memory-Augmented Multi-Modal LLM Agent for Phishing URL Detection (2026.acl-industry)
Copied to clipboard
| Challenge: | Traditional phishing website detection relies on static heuristics or reference lists, which lag behind rapidly evolving attacks. |
| Approach: | They propose a memory-augmented multi-modal LLM agent that leverages episodic memories to guide decisions on recurring and novel threats. |
| Outcome: | The proposed agent outperforms state-of-the-art phishing detection tools on two public datasets and improves recall by 20%. |
Towards Computationally Feasible Deep Active Learning (2022.findings-naacl)
Copied to clipboard
Akim Tsvigun, Artem Shelmanov, Gleb Kuzmin, Leonid Sanochkin, Daniil Larionov, Gleb Gusev, Manvel Avetisian, Leonid Zhukov
| Challenge: | Active learning (AL) is a technique for reducing the amount of annotation required for training machine learning models. |
| Approach: | They propose two techniques that reduce the amount of time required for AL . they use pseudo-labeling and distilled models to train a successor model . |
| Outcome: | The proposed algorithm reduces the time and computational overhead required to train an acquisition model and estimate uncertainty on instances in the unlabeled pool. |
Gaussian Process Optimization for Adaptable Multi-Objective Text Generation using Linearly-Weighted Language Models (2024.findings-naacl)
Copied to clipboard
| Challenge: | Multi-objective text generation requires a method to optimize for dynamic requirements of the overall objective. |
| Approach: | They propose a linear combination of objective-specific language models to efficiently adapt the decoding process and optimize for the desired overall objective without retraining one or more language models. |
| Outcome: | The proposed method outperforms other weighting schemes and standard baselines in a few iterations of decoding. |
Router-Tuning: A Simple and Effective Approach for Dynamic Depth (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to improve computational efficiency are under-explored and face several critical challenges. |
| Approach: | They propose a method that selectively activates only a subset of the model's layers, skipping those deemed less important. |
| Outcome: | The proposed method significantly improves performance on Attention layers and MoE layers while reducing redundant computation and memory usage. |
Retrieved Sequence Augmentation for Protein Representation Learning (2024.emnlp-main)
Copied to clipboard
Chang Ma, Haiteng Zhao, Lin Zheng, Jiayi Xin, Qintong Li, Lijun Wu, Zhihong Deng, Yang Lu, Qi Liu, Sheng Wang, Lingpeng Kong
| Challenge: | Using multiple sequence alignments (MSA) to extract evolutionary knowledge is limited. |
| Approach: | They propose to use multiple sequence alignments to augment protein representations . they propose to employ Retrieved Sequence Augmentation to enhance protein representation learning . |
| Outcome: | The proposed method surpasses MSA Transformer by 5% in structural and property prediction tasks while being 373 times faster. |
ProxyLM: Predicting Language Model Performance on Multilingual Tasks via Proxy Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Performance prediction is a method to estimate the performance of Language Models (LMs) on various Natural Language Processing (NLP) tasks. |
| Approach: | They propose a task- and language-agnostic framework to predict the performance of Language Models (LMs) using proxy models. |
| Outcome: | The proposed framework outperforms the state-of-the-art in root-mean-square error (RMSE) and other robustness tests on multilingual NLP tasks. |
Grammar-Constrained Neural Semantic Parsing with LR Parsers (2021.findings-acl)
Copied to clipboard
| Challenge: | a context-free grammar can be used to enforce syntactical constraints when predicting logical forms. |
| Approach: | They propose a model that uses an LR parser to maintain syntactically valid sequences throughout the decoding procedure. |
| Outcome: | The proposed model is conceptually simpler and adds less overhead during inference compared to other approaches . it is compared with existing grammar-guided decoding frameworks and is cost-effective . |
Multimodal Transformers are Hierarchical Modal-wise Heterogeneous Graphs (2025.acl-long)
Copied to clipboard
| Challenge: | Multimodal Sentiment Analysis (MSA) is a rapidly developing field that integrates multimodal information to recognize sentiments. |
| Approach: | They propose a multimodal fusion model that integrates multimodal information to recognize sentiments using multimodal transformers. |
| Outcome: | The proposed model achieves significantly higher performance than MulTs and the existing model is robust. |
Test-Time Steering for Lossless Text Compression via Weighted Product of Experts (2025.findings-emnlp)
Copied to clipboard
| Challenge: | gzip and neural compression models often lead to poor performance in unseen data. |
| Approach: | They propose a framework that performs Test-Time Steering via a Weighted Product of Experts (wPoE) |
| Outcome: | The proposed framework performs Test-Time Steering via a Weighted Product of Experts (wPoE) it integrates with any autoregressive language model, providing a practical solution for enhancing text compression across diverse data distributions. |
AtTGen: Attribute Tree Generation for Real-World Attribute Joint Extraction (2023.acl-long)
Copied to clipboard
| Challenge: | Attribute extraction aims to identify attribute names and the corresponding attribute values from descriptive texts. |
| Approach: | They propose a unified formulation for real-world attribute extraction application, where closed-world, open-world and semi-open attribute extraction tasks are modeled uniformly. |
| Outcome: | The proposed model outperforms existing methods on three datasets and outperformed existing methods by a large margin. |
Encouraging Neural Machine Translation to Satisfy Terminology Constraints (2021.findings-acl)
Copied to clipboard
| Challenge: | a new approach to encourage neural machine translation to satisfy lexical constraints is proposed . a BLEU score and percentage of generated constraint terms are improved by the proposed method . |
| Approach: | They propose a method that encourages neural machine translation to satisfy lexical constraints at training step . they use a simplified augmentation strategy without source factors and constraint token masking to make it easier to learn the copy behavior . |
| Outcome: | The proposed method improves on baselines in terms of BLEU score and percentage of generated constraint terms. |
MAIN-RAG: Multi-Agent Filtering Retrieval-Augmented Generation (2025.acl-long)
Copied to clipboard
Chia-Yuan Chang, Zhimeng Jiang, Vineeth Rakesh, Menghai Pan, Chin-Chia Michael Yeh, Guanchu Wang, Mingzhi Hu, Zhichao Xu, Yan Zheng, Mahashweta Das, Na Zou
| Challenge: | Existing RAG systems struggle with the quality of retrieval documents, causing performance degradation and reducing performance. |
| Approach: | They propose a training-free RAG framework that leverages multiple LLM agents to collaboratively filter and score retrieved documents. |
| Outcome: | The proposed framework outperforms existing RAG frameworks in QA benchmarks and shows superior answer consistency and answer accuracy over baseline methods. |
Reasoning in Flux: Enhancing Large Language Models Reasoning through Uncertainty-aware Adaptive Guidance (2024.acl-long)
Copied to clipboard
Zhangyue Yin, Qiushi Sun, Qipeng Guo, Zhiyuan Zeng, Xiaonan Li, Junqi Dai, Qinyuan Cheng, Xuanjing Huang, Xipeng Qiu
| Challenge: | Extensive experiments across various reasoning tasks demonstrate that UAG not only enhances the reasoning abilities of LLMs but consistently outperforms several strong baselines with minimal computational overhead. |
| Approach: | They propose an approach to guide LLMs onto an accurate and reliable trajectory by identifying and adjusting uncertainty signals within each step of the reasoning chain. |
| Outcome: | The proposed approach outperforms strong baselines and outperformed strong models with minimal computational overhead. |
In-Context Former: Lightning-fast Compressing Context for Large Language Model (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to reduce inference costs of transformer-based large language models entail quadratic complexity . et al., 2017): transformer-derived large language model performance is a major challenge. |
| Approach: | They propose a method that compresses long contexts into short soft prompts . they use the self-attention mechanism of the large model to extract and condense information . |
| Outcome: | The proposed method reduces compression costs by 68 to 112 times while achieving 90% of baseline performance. |
Training Long-Context LLMs Efficiently via Chunk-wise Optimization (2025.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in long-context large language models have demonstrated superior retrieval quality compared to retrievalaugmented generation (RAG) approaches. |
| Approach: | They propose a memory-efficient training paradigm that partitions lengthy inputs into manageable chunks. |
| Outcome: | The proposed model expands maximum sequence length from 1K to 16K tokens on a single RTX 3090 GPU, while SpaCO achieves accelerated training speed. |
Improving Chain-of-Thought for Logical Reasoning via Attention-Aware Intervention (2026.findings-eacl)
Copied to clipboard
| Challenge: | Modern logical reasoning with LLMs relies on employing complex interactive frameworks that decompose the reasoning process into subtasks solved through carefully designed prompts or requiring external components, which limit their scalability. |
| Approach: | They propose a non-interactive, end-to-end framework for reasoning tasks that enables reasoning to emerge within the model itself. |
| Outcome: | The proposed framework improves generalization while preserving analyzability without external resources. |
ASETF: A Novel Method for Jailbreak Attack on LLMs through Translate Suffix Embeddings (2024.emnlp-main)
Copied to clipboard
| Challenge: | Attaching suffixes to harmful instructions can hack the defense of Large language models (LLMs) However, due to the unreadable of adversarial suffix, it can be relatively easily penetrated by common defense methods such as perplexity filters. |
| Approach: | They propose an algorithm to embed adversarial suffixes into coherent and understandable text to attack Large language models (LLMs) using a Advbench dataset. |
| Outcome: | The proposed approach reduces the computation time of adversarial suffixes and achieves a much better attack success rate than existing techniques. |
Towards Faithful Multi-step Reasoning through Fine-Grained Causal-aware Attribution Reasoning Distillation (2025.coling-main)
Copied to clipboard
| Challenge: | Recent advances have witnessed large language models (LLMs) achieving significant milestones across various domains of natural language processing. |
| Approach: | They introduce fine-grained attribution reasoning distillation (FARD) which incorporates grounded citations to consolidate the relationships between reasoning steps. |
| Outcome: | The proposed method outperforms CoT distillation methods on mathematical and general reasoning benchmarks. |
SHIFT: Selected Helpful Informative Frame for Video-guided Machine Translation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Video-guided machine translation (VMT) aims to improve translation quality by integrating contextual information from paired short video clips. |
| Approach: | They propose a plug-and-play framework for video-guided machine translation with multimodal large language models. |
| Outcome: | The proposed framework improves performance of MLLMs while reducing computational cost. |
MiniConGTS: A Near Ultimate Minimalist Contrastive Grid Tagging Scheme for Aspect Sentiment Triplet Extraction (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches within the pretraining-finetuning paradigm tend to meticulously craft complex tagging schemes and classification heads, or incorporate external semantic enhancements to enhance performance. |
| Approach: | They propose to integrate a minimalist tagging scheme and a novel token-level contrastive learning strategy to improve pretrained representations. |
| Outcome: | The proposed framework achieves comparable or superior performance compared to state-of-the-art techniques while featuring a more compact design and reduced computational overhead. |
Debiased Orthogonal Boundary-Driven Efficient Noise Mitigation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing noise mitigation methods face limitations due to their task-specific design, model dependency, and significant computational overhead. |
| Approach: | They propose a model-agnostic noise label mitigation paradigm that uses an estimator model and a scoring function to assess the noise level of input pairs. |
| Outcome: | The proposed method is superior to existing methods in training tasks and tasks. |
AdaTP: Attention-Debiased Token Pruning for Video Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing visual token compression methods rely on attention scores but have inherent biases . global and local attention biased scores cause excessive computational overhead . |
| Approach: | They propose a token pruning pipeline that targets global and local attention biases . the pipeline is designed to reduce computational overhead of Video Large Language Models based on visual tokens compiled from multiple video frames . |
| Outcome: | The proposed method significantly reduces the computational overhead of Video Large Language Models while retaining the performance of vanilla models. |
T2: An Adaptive Test-Time Scaling Strategy for Contextual Question Answering (2025.emnlp-main)
Copied to clipboard
Zhengyi Zhao, Shubo Zhang, Zezhong Wang, Huimin Wang, Yutian Zhao, Bin Liang, Yefeng Zheng, Binyang Li, Kam-Fai Wong, Xian Wu
| Challenge: | Existing efficient test-time scaling methods introduce budget constraints or early stop mechanisms to avoid overthinking for straightforward questions but add human bias to the reasoning process. |
| Approach: | They propose a framework that dynamically adapts reasoning depth based on question complexity. |
| Outcome: | Experimental results show that the proposed framework achieves higher accuracy than baseline methods and reduces computational overhead by up to 25.2%. |
Mamba-Shedder: Post-Transformer Compression for Efficient Selective Structured State Space Models (2025.naacl-long)
Copied to clipboard
| Challenge: | Large pre-trained models have achieved outstanding results in sequence modeling . alternative architectures, such as Selective Structured State Space Models (SSMs), have been proposed to address these inefficiencies. |
| Approach: | They propose to reduce the size and computational overhead of large pre-trained models by removing selected components at different granularities. |
| Outcome: | The proposed models achieve a speedup of up to 1.4x during inference while maintaining accuracy. |
Icon2: Aligning Large Language Models Using Self-Synthetic Preference Data via Inherent Regulation (2025.emnlp-main)
Copied to clipboard
Qiyuan Chen, Hongsen Huang, Qian Shao, Jiahe Chen, Jintai Chen, Hongxia Xu, Renjie Hua, Ren Chuan, Jian Wu
| Challenge: | Large Language Models (LLMs) require high quality preference datasets to align with human preferences. |
| Approach: | They propose a framework that leverages inherent regulation of LLMs’ representation space for efficient and tailored preference dataset construction, named Icon2. |
| Outcome: | The proposed framework improves performance on benchmarks like AlpacaEval 2.0 and Arena-Hard while reducing computational costs by up to 48.1%. |
EventWeave: A Dynamic Framework for Capturing Core and Supporting Events in Dialogue Systems (2026.acl-long)
Copied to clipboard
Zhengyi Zhao, Shubo Zhang, Yiming Du, Bin Liang, Baojun Wang, Zhongyang Li, Binyang Li, Kam-Fai Wong
| Challenge: | Existing dialogue systems process conversational turns in isolation, overlooking event structures that guide natural interactions. |
| Approach: | They propose a framework that explicitly models relationships between conversational events to generate more contextually appropriate dialogue responses. |
| Outcome: | Experiments on three dialogue datasets show that the proposed approach produces more natural responses while requiring less computational overhead. |
Marking Code Without Breaking It: Code Watermarking for Detecting LLM-Generated Code (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for identifying LLM-generated code are limited by syntax-critical tokens, which can introduce syntax errors. |
| Approach: | They propose a syntax-aware watermarking method that embeds watermarks only in non-syntactic tokens and preserves code integrity. |
| Outcome: | The proposed method outperforms baseline methods on Python, C++, and Java. |
DIDS: Domain Impact-aware Data Sampling for Large Language Model Training (2025.emnlp-main)
Copied to clipboard
Weijie Shi, Jipeng Zhang, Yaguang Wu, Jingzhi Fang, Shibo Zhang, Yao Zhao, Hao Chen, Ruiyuan Zhang, Yue Cui, Jia Zhu, Sirui Han, Jiajie Xu, Xiaofang Zhou
| Challenge: | Existing approaches for optimizing domain-level sampling strategies struggle with maintaining intra-domain consistency and accurately measuring domain impact. |
| Approach: | They propose to use a Fisher-Information Matrix-guided metric to measure domain impact to ensure intra-domain consistency and accuracy. |
| Outcome: | The proposed model achieves 3.4% higher average performance while maintaining comparable training efficiency. |
SKIntern: Internalizing Symbolic Knowledge for Distilling Better CoT Capabilities into Small Language Models (2025.coling-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have high computational costs and privacy concerns due to their high computational expenses and data privacy. |
| Approach: | They propose a method that empowers SLMs to internalize symbolic knowledge and few-shot examples gradually through a progressive fine-tuning process. |
| Outcome: | The proposed approach outperforms state-of-the-art baselines by over 5% while reducing inference costs by up to 4 across a wide range of SLMs in both in-domain (ID) and out-of domain (OOD) tasks. |
Plug-and-Play Conversational Models (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Large conversational models that generate coherent and fluent responses often require large dialogue datasets. |
| Approach: | They propose and evaluate plug-and-play methods for controllable response generation . they demonstrate a high degree of control over the generated conversational responses . |
| Outcome: | The proposed method does not require further computation at decoding time and does not need fine-tuning of a large language model. |
TinyAlign: Boosting Lightweight Vision-Language Models by Mitigating Modal Alignment Bottlenecks (2026.findings-acl)
Copied to clipboard
Yuanze Hu, Xinyu Wang, Zhichao Yang, Gen Li, Ye Qiu, Zhaoxin Fan, Yifan Sun, Wenjun wu, Jin Dong, Xiaotie Deng
| Challenge: | Lightweight Vision-Language Models (VLMs) are indispensable for resource-constrained applications. |
| Approach: | They propose a framework that retrieves context from a memory bank to enhance alignment . they propose EMI-based approach to align vision and language models . |
| Outcome: | The proposed framework reduces training loss, accelerates convergence, and enhances task performance with negligible computational overhead. |
Unlocking the Effectiveness of LoRA-FP for Seamless Transfer Implantation of Fingerprints in Downstream Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | lightweight plug-and-play framework that encodes backdoor fingerprints into LoRA adapters . |
| Approach: | proposed framework encodes backdoor fingerprints into LoRA adapters via constrained fine-tuning . enables seamless fingerprint transplantation through parameter fusion, eliminating full-parameter updates while maintaining integrity. |
| Outcome: | The proposed framework achieves superior robustness against various scenarios while reducing computational overhead compared to traditional approaches. |
Taking a Deep Breath: Enhancing Language Modeling of Large Language Models with Sentinel Tokens (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies have explored compression and accumulation methods to compress contexts, but these methods lose useful context information during the compression process, leading to performance degradation. |
| Approach: | They propose a method that allows LLMs to take a deep breath and insert a special token at the end of each chunk. |
| Outcome: | Experiments on language modeling and out-of-domain tasks validate the superiority of the proposed method. |
SMART: Self-Aware Agent for Tool Overuse Mitigation (2025.findings-acl)
Copied to clipboard
Cheng Qian, Emre Can Acikgoz, Hongru Wang, Xiusi Chen, Avirup Sil, Dilek Hakkani-Tür, Gokhan Tur, Heng Ji
| Challenge: | Current Large Language Models (LLMs) lack self-awareness to balance reasoning and tool use, increasing computational overhead. |
| Approach: | They propose a paradigm that enhances an agent’s self-awareness to optimize task handling and reduce tool overuse. |
| Outcome: | The proposed model reduces tool use by 24% while improving performance by over 37%. |
LVPruning: An Effective yet Simple Language-Guided Vision Token Pruning Approach for Multi-modal Large Language Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Multi-modal Large Language Models (MLLMs) incur significant computational overhead due to the large number of vision tokens processed, limiting their practicality in resource-constrained environments. |
| Approach: | They propose a language-guided vision token pruning method that can be integrated into existing MLLMs with minimal architectural changes. |
| Outcome: | The proposed method reduces vision tokens by 90% and preserves model performance. |
Fast and Accurate Neural Machine Translation with Translation Memory (2021.acl-long)
Copied to clipboard
| Challenge: | Existing knowledge demonstrates the superiority of TM-based neural machine translation only on TM specialized tasks . |
| Approach: | They propose a translation memory-based approach to machine translation using a single bilingual sentence as its TM. |
| Outcome: | The proposed approach surpasses baselines on two general tasks and improves on the TM-specialized translation tasks. |
SALT: Step-level Advantage Assignment for Long-horizon Agents via Trajectory Graph (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable capabilities, but their application to complex, multi-step, and long-horizon tasks remains challenging. |
| Approach: | They propose a framework that provides a finer-grained advantage assignment derived solely from outcome rewards. |
| Outcome: | The proposed framework provides a finer-grained advantage assignment, derived solely from outcome rewards. |
Conditional Causal Relationships between Emotions and Causes in Texts (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies on the causal relationships between emotions and causes focus on extracting causally related clauses from documents, but none considers whether context clauses are indispensable for extracted clauses to be causally linked. |
| Approach: | They propose a task to determine whether an input pair of emotion and cause has a valid causal relationship under different contexts. |
| Outcome: | The proposed task identifies whether an input pair of emotion and cause has a valid causal relationship under different contexts and then fine-tunes the prediction results based on the characteristics of the input clauses. |
An Empirical Study of Speculative Decoding for Small Language Models (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing studies focus on 7B-70B parameters models, leaving a knowledge gap for small language models. |
| Approach: | They propose a draft-then-verify paradigm that allows for a single forward pass through a model and transfer of all model parameters to the GPU cache. |
| Outcome: | The proposed method can be used to accelerate small language models with low computational overhead. |
DART: Distilling Autoregressive Reasoning to Silent Thought (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing models that use Chain-of-Thought (CoT) have been slow to deploy in real-time applications due to its autoregressive nature. |
| Approach: | They propose a framework that replaces autoregressive CoT with non-autoregressive Silent Thought (ST) the framework uses a lightweight Reasoning Evolvement Module to align hidden states with the CoT pathway and a Reasoning Embedment Module (REM) during inference, only the ST pathway is activated, enabling the ST tokens to evolve into informative embeddings. |
| Outcome: | The proposed framework replaces autoregressive CoT with non-autoregressive Silent Thought (ST) it enables LLMs to generate answers directly from ST tokens without additional computational cost . |
TiKMiX: Efficient Semi-Dynamic Data Mixture via Data Influence for LLM Pre-training (2026.acl-long)
Copied to clipboard
Yifan Wang, null Binbinliu, Fengze Liu, Yuanfan Guo, Jiyao Deng, Xuecheng Wu, Weidong Zhou, Xiaohuan Zhou, Taifeng Wang
| Challenge: | Static data mixing strategies in large language models are often suboptimal as they fail to adapt to the model’s evolving learning states. |
| Approach: | They propose a semi-dynamic data mixing framework that uses a key observation of influence ranking invariance to reduce computational overhead by 80% . |
| Outcome: | The proposed method reduces computational overhead by 80% and achieves an average performance gain of 2% across nine downstream benchmarks, effectively mitigating data under-digestion. |
Token Dropping for Efficient BERT Pretraining (2022.acl-long)
Copied to clipboard
| Challenge: | Existing methods to accelerate pretraining of transformer-based models are computationally expensive and degrade performance on downstream tasks. |
| Approach: | They propose a "token dropping" method to accelerate the pretraining of transformer-based models by 25% . they leverage the already built-in masked language modeling loss to identify unimportant tokens with practically no computational overhead. |
| Outcome: | The proposed method reduces the pretraining cost of BERT models by 25% while achieving similar overall performance on downstream tasks. |
Revisit What You See: Revealing Visual Semantics in Vision Tokens to Guide LVLM Decoding (2026.acl-long)
Copied to clipboard
| Challenge: | Large Vision–Language Models (LVLMs) integrate visual perception with language understanding, but how vision information contributes to the model’s decoding process remains under-explored. |
| Approach: | They propose a simple training-free decoding method that guides text generation in Large Vision–Language Models by Referencing Vision Tokens. |
| Outcome: | The proposed method leverages the semantic information embedded within vision tokens by projecting it into the text token distribution. |
M-BRe: Discovering Training Samples for Relation Extraction from Unlabeled Texts with Large Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to extract training instances from unlabeled texts are expensive . sentences that contain the target relations in texts can be scarce and difficult to find . |
| Approach: | They propose a framework that can automatically extract training instances from unlabeled texts for RE. |
| Outcome: | The proposed method can extract training instances from unlabeled texts for RE. |
A little goes a long way: Improving toxic language classification despite data scarcity (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for toxic language classification have not been thoroughly explored. |
| Approach: | They propose to use data augmentation to generate new synthetic data from labeled seed datasets to improve toxic language classification. |
| Outcome: | The proposed techniques perform well on very scarce toxic language datasets while performing worse on shallower models. |
CLIP-MoE: Towards Building Mixture of Experts for CLIP with Diversified Multiplet Upcycling (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies found that CLIP can only encode one aspect of the feature space, leading to substantial information loss and indistinctive features. |
| Approach: | They propose a model-agnostic approach that fine-tunes complementary CLIP models and transforms them into a CLIP-MoE. |
| Outcome: | The proposed framework fine-tunes a series of complementary CLIP models and transforms them into a CLIP-MoE. |
ThinkSwitcher: When to Think Hard, When to Think Fast (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large reasoning models excel at solving complex tasks by leveraging long chain-of-thought (CoT) reasoning. |
| Approach: | They propose a framework that enables a single LRM to dynamically switch between short and long CoT modes based on task complexity. |
| Outcome: | The proposed framework reduces computational cost by 20-30% while maintaining high accuracy on complex tasks. |
Waste Not, Want Not; Recycled Gumbel Noise Improves Consistency in Natural Language Generation (2025.naacl-long)
Copied to clipboard
| Challenge: | Consistency in the output of language models can vary significantly in style, factual accuracy, and tone, even for similar inputs. |
| Approach: | They propose a decoding algorithm that enhances response consistency across different prompts with no degradation in response quality. |
| Outcome: | The proposed method outperforms standard sampling methods by 10% across semantic and stylistic consistency benchmarks. |
Training Neural Machine Translation to Apply Terminology Constraints (P19-1)
Copied to clipboard
| Challenge: | Existing methods to integrate domain terminology into neural machine translation (NMT) are brittle when tested in real-world situations. |
| Approach: | They propose a method to inject custom terminology into neural machine translation at run time by using the target side of terminology entries whose source side match the input as decoding-time constraints. |
| Outcome: | The proposed method is faster than state-of-the-art decoding and more efficient than constraint-free decoding. |
Efficient Multi-Agent System Training with Data Influence-Oriented Tree Search (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Model (LLM) based multi-agent systems (MAS) have high potential for tackling complex tasks through collaborative intelligence. |
| Approach: | They propose a framework that incorporates influence scores to guide tree search and data selection in data synthesis. |
| Outcome: | The proposed framework incorporates influence scores to guide tree search and data selection in data synthesis. |
Improve Transformer Models with Better Relative Position Embeddings (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for generating position embeddings are not fully utilized in NLP tasks. |
| Approach: | They propose to generalize the absolute position embedding to a generalized relative position embedded method . they also propose to use the relative embeddable method to improve the accuracy of large models . |
| Outcome: | The proposed method improves accuracy on the SQuAD1.1 dataset compared to previous methods . it can be easily adopted as a drop-in replacement for improving accuracy of large models . |
MMNMT: Modularizing Multilingual Neural Machine Translation with Flexibly Assembled MoE and Dense Blocks (2023.emnlp-main)
Copied to clipboard
| Challenge: | Mixture-of-Experts (MoE) based sparse architectures are prone to overfitting on low-resource language translation. |
| Approach: | They propose a modularized MNMT framework that flexibly assembles dense and MoE-based sparse modules to achieve the best of both worlds. |
| Outcome: | The proposed framework outperforms existing models on low-resource language translation and zero-shot translation on benchmark datasets. |
Extending Context Window of Large Language Models via Semantic Compression (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing models rely on a quadratic computation to generate long texts . current models impose limitations on the length of text inputs . |
| Approach: | They propose a semantic compression method that extends the context window of large language models . the method reduces the semantic redundancy of long inputs before passing them to the LLMs . |
| Outcome: | The proposed method extends the context window of large language models across tasks . it exhibits consistent fluency in text generation while reducing associated computational overhead. |
When to Continue Thinking: Adaptive Thinking Mode Switching for Efficient Reasoning (2025.findings-emnlp)
Copied to clipboard
Xiaoyun Zhang, Jingqing Ruan, Xing Ma, Yawen Zhu, Haodong Zhao, Hao Li, Jiansong Chen, Ke Zeng, Xunliang Cai
| Challenge: | Large reasoning models (LRMs) incur excessive computational overhead due to redundant reasoning, especially on simple tasks. |
| Approach: | They propose an Adaptive Self-Recovery Reasoning framework that suppresses unnecessary reasoning and enables implicit recovery. |
| Outcome: | The proposed framework suppresses unnecessary reasoning and enables implicit recovery. |
Think Hard Only When Needed: A Hybrid Best-of-N and Beam Search for Efficient Test-Time Compute (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) exhibit remarkable reasoning and planning capabilities, yet their substantial inference-time cost significantly impedes deployment in resourceconstrained applications. |
| Approach: | They propose a hybrid inference pipeline that combines beam search and Best-of-N . THROW generates shorter initial trajectories and evaluates them using PRMs . |
| Outcome: | THROW achieves 1.54 and 14.38 latency speedups and 35.7% and 80.4% token reductions on average compared to Best-of-N and beam search . |
TrigReason: Trigger-Based Collaboration between Small and Large Reasoning Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Reasoning Models suffer from high inference latency due to autoregressive reasoning . SpecReason adopts a polling-based design that repeatedly invokes the LRM for verification at every step . |
| Approach: | They propose a trigger-based collaborative reasoning framework that delegates most reasoning to the SRM and activates LRM intervention only when necessary. |
| Outcome: | The proposed framework reduces latency and API cost by 73.3% under edge–cloud conditions. |
VideoPro: Adaptive Program Reasoning for Long Video Understanding (2026.acl-long)
Copied to clipboard
Chenglin Li, Feng Han, Yikun Wang, Ruilin Li, Shuai Dong, Haowen Hou, Haitao Li, Qianglong Chen, Feng Tao, Jingqi Tong, Yin Zhang, Jiaqi Wang
| Challenge: | Existing methods for understanding long videos are limited due to the sparsity of visual evidence relevant to a given query. |
| Approach: | They propose a framework that enables VideoLLMs to reason over long videos and refine their predictions through executable programs. |
| Outcome: | The proposed framework outperforms existing methods across long-video understanding benchmarks. |
RMoA: Optimizing Mixture-of-Agents through Diversity Maximization and Residual Compensation (2025.findings-acl)
Copied to clipboard
| Challenge: | Multi-agent systems based on large language models are limited by high computational overhead, information loss, and robustness. |
| Approach: | They propose a Residual Mixture-of-Agents (RMoA) that integrates residual connections to optimize efficiency and reliability. |
| Outcome: | The proposed model achieves state-of-the-art performance on benchmarks of alignment, mathematical reasoning, code generation, and multitasking understanding, while significantly reducing computational overhead. |
Jailbreak Attack Initializations as Extractors of Compliance Directions (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Safety-aligned LLMs respond to prompts with compliance or refusal, each corresponding to distinct directions in the model’s activation space. |
| Approach: | They propose an initialization framework that aims to project unseen prompts further along compliance directions. |
| Outcome: | The proposed initialization framework achieves an increased attack success rate and reduced computational overhead, highlighting the fragility of safety-aligned LLMs. |
SPPO: Sequence-Level PPO for Long-Horizon Reasoning Tasks (2026.acl-long)
Copied to clipboard
Tianyi Wang, Yixia Li, Long Li, Yibiao Chen, Shaohan Huang, Yun Chen, Peng Li, Yang Liu, Guanhua Chen
| Challenge: | Proximal Policy Optimization (PPO) is central to aligning Large Language Models with verifiable rewards. |
| Approach: | They propose a scalable algorithm that harmonizes sample efficiency with stability of outcome-based updates. |
| Outcome: | The proposed algorithm outperforms standard PPO and matches the performance of computation-heavy group-based methods. |
Large Product Key Memory for Pretrained Language Models (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing product key memory (PKM) models that increase model capacity with insignificant computational overhead are limited to causal language modeling. |
| Approach: | They propose product key memory (PKM) that enables very efficient and exact nearest neighbor search in a large number of learnable memory slots. |
| Outcome: | The proposed product key memory improves model capacity and performance by replacing a feed-forward network with a model weighted model. |
SpecVLM: Enhancing Speculative Decoding of Video LLMs via Verifier-Guided Token Pruning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Video large language models (Vid-LLMs) rely on dense video token representations and require substantial memory and computational overhead in both prefilling and decoding. |
| Approach: | They propose a training-free speculative decoding framework that prunes up to 90% of video tokens to enable efficient speculation without sacrificing accuracy. |
| Outcome: | The proposed framework achieves 2.68 speedup on LLaVA-OneVision-72B and 2.11 speed up on Qwen2.5-VL-32B. |
Tiny-NewsRec: Effective and Efficient PLM-based News Recommendation (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing work fine tunes the PLM with the news recommendation task, which can cause a domain shift problem. |
| Approach: | They propose a self-supervised method to adapt general PLM to news domain with a contrastive matching task between news titles and news bodies. |
| Outcome: | The proposed method can improve both the effectiveness and efficiency of the large PLM-based news recommendation model while maintaining its performance. |
LEO-MINI: An Efficient Multimodal Large Language Model using Conditional Token Reduction and Mixture of Multi-Modal Experts (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent approaches to reduce visual tokens have been criticized for their computational efficiency and lack of visual reasoning capabilities. |
| Approach: | They propose a novel multi-modal large language model that reduces the number of visual tokens and simultaneously boosts visual reasoning capabilities. |
| Outcome: | The proposed model significantly reduces the number of visual tokens and boosts visual reasoning capabilities. |
Teach Small Models to Reason by Curriculum Distillation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Reasoning Models (LRMs) show strong System-2-style reasoning, but at the cost of significant computational overhead. |
| Approach: | They propose a two-stage curriculum distillation framework which builds a robust internal problem-solving student model and then teaches the student model to externalize this knowledge as explicit reasoning. |
| Outcome: | The proposed model outperforms single-stage baselines on mathematical benchmarks and significantly outperformed LRMs on complex tasks. |
Generative Prompt Internalization (2025.naacl-long)
Copied to clipboard
| Challenge: | Prompts used in large language model based applications are often fixed and lengthy, leading to significant computational overhead. |
| Approach: | They propose a method that internalizes complex prompts using a joint training approach and a data synthesis technique that auto-collects conversational datasets by swapping roles of agent and environment. |
| Outcome: | The proposed method internalizes complex prompts across agent-based applications and generates the content along with reasons for why it should change accordingly. |
GUARD: Glocal Uncertainty-Aware Robust Decoding for Effective and Efficient Open-Ended Text Generation (2025.findings-emnlp)
Copied to clipboard
Yuanhao Ding, Esteban Garces Arias, Meimingwei Li, Julian Rodemann, Matthias Aßenmacher, Danlu Chen, Gaojuan Fan, Christian Heumann, Chongsheng Zhang
| Challenge: | GUARD is a self-adaptive decoding method that balances coherence with diversity in open-ended text generation. |
| Approach: | They propose a self-adaptive decoding method that balances coherence and diversity . they combine global entropy estimates with local entropic deviations to integrate uncertainty . |
| Outcome: | GUARD achieves a good balance between diversity and coherence while exhibiting significant improvements in generation speed. |
Training Language Models with Memory Augmentation (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for training memory-augmented language models only introduce mem-ories at testing time or represent them using a separately trained encoder. |
| Approach: | They propose a training approach that directly takes in-batch examples as accessible memory and new methods for memory construction and data batching that are used for adapting to different sets of memories at testing time. |
| Outcome: | The proposed approach reduces perplexity from 18.70 to 15.37 on multiple language modeling and machine translation benchmarks. |
Attention-based Conditioning Methods for External Knowledge Integration (P19-1)
Copied to clipboard
| Challenge: | Existing approaches for incorporating external knowledge into deep neural networks (RNNs) lexicon features are used to concatenate external information into the input or hidden network layers. |
| Approach: | They propose a method for conditioning external knowledge into RNNs by concatenating a representation of the external information to the input or hidden network layers. |
| Outcome: | The proposed approach improves performance on six benchmark datasets. |
Invariant Language Modeling (2022.emnlp-main)
Copied to clipboard
Maxime Peyrard, Sarvjeet Ghotra, Martin Josifoski, Vidhan Agarwal, Barun Patra, Dean Carignan, Emre Kiciman, Saurabh Tiwary, Robert West
| Challenge: | Existing methods to remove spurious correlations and biases involve expensive domain alignment. |
| Approach: | They propose a framework for learning invariant representations that generalize better across environments . they adapt a game-theoretic implementation of IRM to language models . |
| Outcome: | The proposed framework can remove structured noise, ignore correlations and achieve better generalization across environments. |
Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity (2024.naacl-long)
Copied to clipboard
| Challenge: | Recent Large Language Models (LLMs) generate factually incorrect answers based on their parametric memory. |
| Approach: | They propose a retrieval-augmented large language model that can dynamically select the most suitable strategy based on query complexity. |
| Outcome: | The proposed approach improves the performance of QA systems on open-domain QA datasets. |
Exploring Graph Learning Tasks with Pure LLMs: A Comprehensive Benchmark and Investigation (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent studies focus on performance benchmarks without fully comparing LLMs to graph learning models. |
| Approach: | They evaluate off-the-shelf and instruction-tuned graph learning models across a variety of scenarios. |
| Outcome: | The proposed models outperform traditional graph learning models in few-shot settings, the authors show . their models out perform models with instruction tuning, and they show excellent generalization and robustness. |
An Experimental Design Framework for Label-Efficient Supervised Finetuning of Large Language Models (2024.findings-acl)
Copied to clipboard
Gantavya Bhatt, Yifang Chen, Arnav Das, Jifan Zhang, Sang Truong, Stephen Mussmann, Yinglun Zhu, Jeff Bilmes, Simon Du, Kevin Jamieson, Jordan Ash, Robert Nowak
| Challenge: | Supervised finetuning (SFT) on instruction datasets has shown immense potential in improving the zero-shot generalization capabilities observed in large language models (LLMs). |
| Approach: | They propose to use experimental design to minimize the computational cost of active learning by identifying useful subsets of samples to annotate from an unlabeled pool. |
| Outcome: | The proposed methods save 50% of the annotation cost compared to random sampling on generative tasks. |
XAutoLM: Efficient Fine-Tuning of Language Models via Meta-Learning and AutoML (2025.emnlp-main)
Copied to clipboard
Ernesto Luis Estevanell Valladares, Suilan Estevez-Velarde, Yoan Gutierrez, Andrés Montoyo, Ruslan Mitkov
| Challenge: | XAutoLM is a meta-learning-augmented framework that can be used to optimize discriminative and generative LM fine-tuning pipelines. |
| Approach: | They propose a meta-learning-augmented AutoML framework that reuses past experiences to optimize discriminative and generative LM fine-tuning pipelines efficiently. |
| Outcome: | XAutoLM surpasses zero-shot optimizer’s peak F1 on five of six tasks, reduces mean evaluation time of pipelines by up to 4.5x, and uncovers 50% more pipelines above zero- shot Pareto front. |
Forget for Get: A Lightweight Two-phase Gradient Method for Knowledge Editing in Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing knowledge editing methodologies often encounter parameter conflict during knowledge overwriting and excessive computational overhead. |
| Approach: | They propose a method that erases outdated knowledge and inserts new knowledge at the location that corresponds to the target knowledge. |
| Outcome: | The proposed method achieves more effective knowledge editing at a lower cost compared to previous methods across various base models. |
ConCISE: Confidence-guided Compression in Step-by-step Efficient Reasoning (2025.emnlp-main)
Copied to clipboard
Ziqing Qiao, Yongheng Deng, Jiali Zeng, Dong Wang, Lai Wei, Guanbo Wang, Fandong Meng, Jie Zhou, Ju Ren, Yaoxue Zhang
| Challenge: | Existing methods for fine-tuning-based compression suffer from verbose outputs, increasing computational overhead. |
| Approach: | They propose a framework to generate concise reasoning chains using Confidence Injection and Early Stopping. |
| Outcome: | The proposed framework reduces the length of the model by up to 50% while maintaining high task accuracy. |
Pre-training and Fine-tuning Neural Topic Model: A Simple yet Effective Approach to Incorporating External Knowledge (2022.acl-long)
Copied to clipboard
| Challenge: | Recent studies have shown that using external knowledge such as pre-trained word embeddings or pre-train language models only achieved limited performance improvements but with huge computational overhead. |
| Approach: | They propose to incorporate external knowledge into neural topic modeling by pre-trained word embeddings (PWEs) or pre-train language models (PLMs) they propose to fine-tune the neural topic model on the target dataset and reduce the huge size of training data. |
| Outcome: | The proposed approach outperforms current state-of-the-art neural topic models and some topic modeling approaches enhanced with PWEs or PLMs on three datasets and greatly reduces the huge size of training data. |
Multimodal End-to-End Sparse Model for Emotion Recognition (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing work in emotion recognition uses a two-phase pipeline, but the extracted features are fixed and cannot be fine-tuned on different tasks. |
| Approach: | They propose a two-phase pipeline for emotion recognition and personality recognition . they propose restructured datasets to enable fully end-to-end training . |
| Outcome: | The proposed model outperforms the current state-of-the-art models on emotion recognition and personality recognition tasks with half less computation in the feature extraction part. |
Language-Guided Temporal Token Pruning for Efficient VideoLLM Processing (2025.emnlp-main)
Copied to clipboard
| Challenge: | Current models struggle with long-form videos due to the quadratic complexity of attention mechanisms. |
| Approach: | They propose a model-agnostic framework that leverages temporal cues from queries to prune video tokens. |
| Outcome: | The proposed framework reduces computation by 65% while preserving 97-99% of original performance. |
Staged Knowledge Distillation Through Least-to-Most Prompting: Optimizing Teacher Guidance via Difficulty-Aware Training (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) enables the compression of large language models (LLMs) conventional methods suffer from training-inference mismatches and suboptimal performance due to expensive student-generated outputs. |
| Approach: | They propose a method that combines a CL strategy and adaptive loss design to reduce training mismatches and suboptimal performance. |
| Outcome: | L2M-KD outperforms existing white-box KD methods on instruction-following tasks . it outperformed existing methods, achieving superior student model performance with reduced overhead . |
IIET: Efficient Numerical Transformer via Implicit Iterative Euler Method (2025.emnlp-main)
Copied to clipboard
Xinyu Liu, Bei Li, Jiahao Liu, Junhao Ruan, Kechen Jiao, Hongyin Tang, Jingang Wang, Tong Xiao, JingBo Zhu
| Challenge: | High-order numerical methods enhance performance in tasks like NLP but introduce a performance-efficiency trade-off due to increased computational overhead. |
| Approach: | They propose an iterative implicit Euler Transformer which simplifies high-order numerical methods by iterating implicit Eule. |
| Outcome: | The proposed method improves accuracy and reduces inference overhead by 55% while retaining 99.4% of the original task accuracy. |
DistillMIKE: Editing Distillation of Massive In-Context Knowledge Editing in Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | In-context knowledge editing has shown respectable abilities on knowledge editing in terms of generalization and specificity. |
| Approach: | They propose a novel extension of in-context knowledge editing (IKE) that allows for massive edits to be injected into large language models. |
| Outcome: | The proposed method shows state-of-the-art perfomrances and comparable performance with MIKE. |
AdaSwitch: Adaptive Switching between Small and Large Agents for Effective Cloud-Local Collaborative Learning (2024.emnlp-main)
Copied to clipboard
Hao Sun, Jiayi Wu, Hengyi Cai, Xiaochi Wei, Yue Feng, Bo Wang, Shuaiqiang Wang, Yan Zhang, Dawei Yin
| Challenge: | Recent advances in large language models have been remarkable . users face a choice between using cloud-based LLMs for generation quality or local-based ones for lower computational cost . |
| Approach: | They propose a new LLM utilization paradigm that facilitates collaborative operation . they evaluate AdaSwitch across 7 benchmarks and compare it to other LLMs . |
| Outcome: | The proposed model improves performance of local and cloud agents across 7 benchmarks . it achieves competitive results compared to the cloud agent while utilizing less computational overhead. |
CoBa: Convergence Balancer for Multitask Finetuning of Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing multi-task learning approaches for large language models fall short due to computational intensive or lack of simultaneous task convergence. |
| Approach: | They propose a new multi-task learning approach that dynamically adjusts task weights during the training process, ensuring that the validation loss of all tasks progresses towards convergence at an even pace. |
| Outcome: | The proposed approach improves the performance of large language models by up to 13% compared to the second-best approaches. |
Faster Machine Translation Ensembling with Reinforcement Learning and Competitive Correction (2025.findings-naacl)
Copied to clipboard
| Challenge: | Recent approaches to ensembling neural machine translation models require inference across all candidate models, leading to significant computational overhead. |
| Approach: | They propose a reinforcement learning-based strategy that improves the CSB by selecting a small, fixed number of candidates and identifying optimal groups to pass to the fusion block for each input sentence. |
| Outcome: | The proposed approach improves the CSB by selecting a small, fixed number of candidates and identifying optimal groups to pass to the fusion block for each input sentence. |
Better Call SAUL: Fluent and Consistent Language Model Editing with Generation Regularization (2024.findings-emnlp)
Copied to clipboard
| Challenge: | State-of-the-art methods for updating large language models require computational overhead and lack theoretical validation. |
| Approach: | They propose a model editing method that uses sentence concatenation with augmented random facts for generation regularization. |
| Outcome: | The proposed method outperforms state-of-the-art methods while maintaining generation quality and reducing computational overhead. |
Laziness Is a Virtue When It Comes to Compositionality in Neural Semantic Parsing (2023.acl-long)
Copied to clipboard
Maxwell Crouse, Pavan Kapanipathi, Subhajit Chaudhury, Tahira Naseem, Ramon Fernandez Astudillo, Achille Fokoue, Tim Klinger
| Challenge: | Compositional generalization is a key feature of human intelligence and has been identified as a major point of weakness in neural methods for semantic parsing. |
| Approach: | They propose a neural parsing generation method that constructs logical forms from the bottom up, beginning from the logical form’s leaves. |
| Outcome: | The proposed method outperforms general-purpose parsers on a CFQ dataset and two other Text-to-SQL datasets while also being competitive with parser that have been tailored to each task. |
Alignment-Based Decoding Policy for Low-Latency and Anticipation-Free Neural Japanese Input Method Editors (2024.findings-acl)
Copied to clipboard
| Challenge: | Japanese input method editors (IMEs) allow users to input Japanese text using a limited set of characters such as the kana syllabary. |
| Approach: | They propose a simple decoding policy to enable simultaneous kana-kanji conversion in Japanese IMEs inspired by simultaneous machine translation. |
| Outcome: | The proposed approach achieves a better quality-latency trade-off than baselines while being more practical due to its ability to directly handle streaming input. |
ReKG-MCTS: Reinforcing LLM Reasoning on Knowledge Graphs via Training-Free Monte Carlo Tree Search (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to combining knowledge graphs with large language models face limitations in path exploration strategies or excessive computational overhead. |
| Approach: | They propose a training-free framework that synergizes Monte Carlo Tree Search with LLM capabilities to enable dynamic reasoning over KGs. |
| Outcome: | The proposed framework outperforms existing training-free methods and achieves competitive performance compared to fine-tuned baselines. |
StableMoE: Stable Routing Strategy for Mixture of Experts (2022.acl-long)
Copied to clipboard
| Challenge: | Existing learning-to-route methods suffer from the routing fluctuation issue . with the model scale growing, training speed will go slower and memory requirements are heavy . |
| Approach: | They propose a Mixture-of-Experts technique that can scale up the model size of Transformers with an affordable computational overhead. |
| Outcome: | The proposed method outperforms existing learning-to-route methods on language modeling and multilingual machine translation. |
Stop Looking for “Important Tokens” in Multimodal Language Models: Duplication Matters More (2025.emnlp-main)
Copied to clipboard
Zichen Wen, Yifeng Gao, Shaobo Wang, Junyuan Zhang, Qintong Zhang, Weijia Li, Conghui He, Linfeng Zhang
| Challenge: | Vision tokens in multimodal large language models often dominate computational overhead due to excessive length compared to linguistic modality. |
| Approach: | They propose a token pruning method which defines an importance criterion for vision tokens and prunes the unimportant vision token during inference. |
| Outcome: | The proposed method can prune 88.9% of vision tokens while maintaining comparable performance. |
Learning from Evolving Training Dynamics: An Entropy-Maximizing Data Curation Strategy for LLM Supervised Post-Training (2026.acl-long)
Copied to clipboard
| Challenge: | EVO-Curate is a dynamic data curation framework that synchronizes sample complexity with the maturing capacity of the Large Language Model (LLM). |
| Approach: | They propose a dynamic data curation framework that synchronizes sample complexity with maturing capacity of the Large Language Model (LLM) they use an Adaptive Dynamics Measurer to synthesize instantaneous difficulty and historical variability into a multidimensional utility score. |
| Outcome: | The proposed framework outperforms standard training baselines and traditional CL methods on instruction following, mathematical reasoning, and code generation architectures while maintaining manageable computational overhead. |
Propulsion: Steering LLM with Tiny Fine-Tuning (2025.coling-main)
Copied to clipboard
| Challenge: | Propulsion is a parameter-efficient fine-tuning method that selectively re-scales specific dimensions of a pre-trained model without modifying the model’s parameters. |
| Approach: | They propose a parameter-efficient fine-tuning method that selectively re-scales specific dimensions of a pre-trained model without modifying the parameters. |
| Outcome: | The proposed method reduces parameter count from 355.3 million to 0.086 million while maintaining competitive performance across benchmarks. |
Less is More: A Simple yet Effective Token Reduction Method for Efficient Multi-modal LLMs (2025.coling-main)
Copied to clipboard
| Challenge: | Recent advances in Multimodal Large Language Models have led to a significant surge in the resource consumption of these models. |
| Approach: | They propose a method to reduce image tokens using visual query data by using CLIP metrics to reduce computational overhead and maintain consistent performance. |
| Outcome: | The proposed method has been extensively tested across 12 datasets and shows a significant reduction in computational overhead while maintaining a consistent level of performance. |
Sharpness-Aware Minimization Improves Language Model Generalization (2022.acl-long)
Copied to clipboard
| Challenge: | Comparatively little work has been done to improve the generalization of language models . recent work shows that Sharpness-Aware Minimization (SAM) can improve generalization without much computational overhead. |
| Approach: | They propose a Sharpness-Aware Minimization procedure that encourages convergence to flatter minima to improve generalization of language models without much computational overhead. |
| Outcome: | The proposed Sharpness-Aware Minimization procedure can improve language models without much computational overhead. |
Uplift-RAG: Uplift-Driven Knowledge Preference Alignment for Retrieval-Augmented Generation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing efforts to estimate document utility rely on downstream generation performance, which conflates the influence of external documents with the intrinsic knowledge of the LLM. |
| Approach: | They propose an uplift-based definition of document utility that quantifies each document’s marginal benefit over the LLM’s internal knowledge. |
| Outcome: | The proposed framework improves the performance of the LLM by incorporating external retrieved documents into the model. |
CleanGen: Mitigating Backdoor Attacks for Generation Tasks in Large Language Models (2024.emnlp-main)
Copied to clipboard
Yuetai Li, Zhangchen Xu, Fengqing Jiang, Luyao Niu, Dinuka Sahabandu, Bhaskar Ramasubramanian, Radha Poovendran
| Challenge: | Generative large language models (LLMs) have remarkable performance in generation tasks, but datasets used to train or fine-tune these models are often not disclosed to users. |
| Approach: | They develop an inference time defense called CleanGen to mitigate backdoor attacks for generation tasks in large language models. |
| Outcome: | The proposed inference time defense achieves lower attack success rates (ASR) compared to baseline defenses for all five backdoor attacks. |
Enhancing Court View Generation with Knowledge Injection and Guidance (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods of natural language generation (NLG) rely on the extensive parameters of pretrained language models (PLMs) but their effectiveness may be compromised by insufficient domain-specific knowledge. |
| Approach: | They propose a knowledge-injected prompt encoder to incorporate domain knowledge during the training stage, thereby reducing computational overhead. |
| Outcome: | The proposed approach outperforms established baselines on real-world data in responsivity of claims and in the ability to transfer domain knowledge. |
Doubling Your Data in Minutes: Ultra-fast Tabular Data Generation via LLM-Induced Dependency Graphs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Tabular data is critical across diverse domains, yet high-quality tabular datasets remain scarce due to privacy concerns and the cost of collection. |
| Approach: | They propose a lightweight generative framework that captures sparse dependencies via an LLM-induced graph. |
| Outcome: | The proposed framework reduces constraint violations by 4% and accelerates generation by nearly 9,500 over diffusion-based methods. |
Mitigating Tail Narrowing in LLM Self-Improvement via Socratic-Guided Sampling (2025.naacl-long)
Copied to clipboard
Yiwen Ding, Zhiheng Xi, Wei He, Lizhuoyuan Lizhuoyuan, Yitao Zhai, Shi Xiaowei, Xunliang Cai, Tao Gui, Qi Zhang, Xuanjing Huang
| Challenge: | Large language models (LLMs) generate solutions themselves and iteratively train on filtered, high-quality rationales, but performance reaches a ceiling after a few iterations. |
| Approach: | They propose a strategy to improve the efficiency of sampling heavy-tailed data by using Socratic-style guidance signals to help LLMs reasoning with complex queries. |
| Outcome: | The proposed approach is effective on difficult queries and on held-out tasks, while requiring human supervision. |
CIFLEX: Contextual Instruction Flow for Sub-task Execution in Multi-Turn Interactions with a Single On-Device LLM (2025.emnlp-main)
Copied to clipboard
| Challenge: | Experiments show that CIFLEX significantly reduces computational costs without degrading task performance. |
| Approach: | They propose a new execution system for efficient sub-task handling with a single large language model. |
| Outcome: | Experiments show that CIFLEX significantly reduces computational costs without degrading task performance. |
Revisiting Transformer-based Models for Long Document Classification (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Recent literature in text classification is biased towards short text sequences . multi-page multi-paragraph documents cannot be efficiently encoded by vanilla transformers based on short text. |
| Approach: | They compare different Transformer-based Long Document Classification approaches to mitigate the computational overhead of vanilla transformers to encode much longer text. |
| Outcome: | The proposed models can process longer text and provide practical advice for long document classification tasks. |
Chain-of-Memory: Lightweight Memory Construction with Dynamic Evolution for LLM Agents (2026.acl-long)
Copied to clipboard
| Challenge: | Existing paradigms for large language model (LLM) agents use memory construction and retrieval-augmented generation. |
| Approach: | They propose a framework that advocates for a paradigm shift toward lightweight construction paired with sophisticated utilization. |
| Outcome: | Experiments show that CoM outperforms baselines with accuracy gains of 7.5%–10.4% while reducing computational overhead to approximately 2.7% of token consumption and 6.0% of latency compared to complex memory architectures. |
TriFine: A Large-Scale Dataset of Vision-Audio-Subtitle for Tri-Modal Machine Translation and Benchmark with Fine-Grained Annotated Tags (2025.coling-main)
Copied to clipboard
| Challenge: | Existing video-guided machine translation approaches use coarse-grained visual information, resulting in information redundancy and high computational overhead. |
| Approach: | They propose a fine-grained approach to video-guided machine translation using visual information . they use a large-scale dataset with annotated multimodal fine-grain tags . |
| Outcome: | The proposed approach achieves superior performance with lower computational overhead compared to coarse-grained methods and text-only models. |
Accelerating LLM Reasoning via Early Rejection with Partial Reward Modeling (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly relied upon for solving complex reasoning tasks. |
| Approach: | They propose to use Process Reward Models to scale inference time compute by generating in parallel . they propose to provide early signals that enable the rejection of suboptimal candidates before full generation of step is complete. |
| Outcome: | The proposed method achieves 1.4 – 9 reduction in inference FLOPs without degrading final performance. |
Investigating Cross-Modal Skill Injection: Scenarios, Methods, and Hyperparameters (2026.acl-long)
Copied to clipboard
| Challenge: | Existing research lacks systematic analysis of the applicability and methodology of cross-modal skill injection. |
| Approach: | They investigate the applicability and methodology of cross-modal skill injection by integrating a domain-expert LLM into a VLM. |
| Outcome: | The proposed method enables transfer of domain-specific expertise from Large Language Models (LLMs) to VLMs without incurring additional training data requirements or significant computational overhead. |
UORA: Uniform Orthogonal Reinitialization Adaptation in Parameter Efficient Fine-Tuning of Large Models (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods such as LoRA and VeRA use a low-rank approximation method that reduces the number of trainable parameters without compromising performance. |
| Approach: | They propose a parameter-efficient fine-tuning approach that leverages a low-rank approximation method that reduces the number of trainable parameters without compromising performance. |
| Outcome: | The proposed approach outperforms existing methods on GLUE and E2E benchmarks and is effective in instruction-tuning large language models and image classification models. |
FastGAS: Fast Graph-based Annotation Selection for In-Context Learning (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to select unlabeled examples for annotation require a long time due to their complexity, hindering their practical viability. |
| Approach: | They propose a graph-based selection method to efficiently identify high-quality instances while minimizing computational overhead. |
| Outcome: | The proposed method significantly reduces selection time and improves performance on different tasks. |
FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steering (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing prompt-based debiasing methods exhibit instability due to sensitivity to prompt changes . fine-tuning-based techniques incur substantial computational overhead and catastrophic forgetting . |
| Approach: | They propose a debiasing framework that encodes fairness-related features into separable directions in the hidden activation space. |
| Outcome: | The proposed framework performs inference-time debiasing without requiring retraining or prompt design . it detects bias signatures in activations and then computes debiased steering vectors . the proposed framework is available to download in the u.s. |
IntroLM: Introspective Language Models via Prefilling-Time Self-Evaluation (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to predict output quality of large language models rely on external classifiers with limited context windows and constrained representational capacity. |
| Approach: | They propose a method that enables causal language models to predict their own output quality during the prefilling phase without affecting generation using [CPX] tokens. |
| Outcome: | The proposed method outperforms existing classifiers on Qwen3-8B and DeBERTa-v3-Large models by 14% on question-answering benchmarks. |
WavRAG: Audio-Integrated Retrieval Augmented Generation for Spoken Dialogue Models (2025.acl-long)
Copied to clipboard
| Challenge: | Existing RAG frameworks rely on Automatic Speech Recognition to process speech input, which discards crucial audio information and increases computational overhead. |
| Approach: | They propose a retrieval augmented generation framework with native, end-to-end audio support that integrates audio and text into a unified knowledge representation. |
| Outcome: | The proposed framework can perform 10x faster than current pipelines while delivering 10x acceleration. |
Nudging: Inference-time Alignment of LLMs via Guided Decoding (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) require alignment to effectively and safely follow user instructions. |
| Approach: | They propose a simple, training-free algorithm that aligns any base model at inference time using a small aligned model. |
| Outcome: | The proposed algorithm outperforms large aligned models on open-instruction tasks without training. |
AdapterDrop: On the Efficiency of Adapters in Transformers (2021.emnlp-main)
Copied to clipboard
Andreas Rücklé, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, Iryna Gurevych
| Challenge: | Recent approaches to transformer models are expensive to fine-tune, slow for inference, and have large storage requirements. |
| Approach: | They propose a method to remove adapters from transformer layers during training and inference . they show that AdapterDrop can dynamically reduce computational overhead . |
| Outcome: | The proposed approach reduces computational overhead while maintaining performance over multiple tasks with minimal loss of performance. |
CFSP: An Efficient Structured Pruning Framework for LLMs with Coarse-to-Fine Activation Information (2025.coling-main)
Copied to clipboard
Yuxin Wang, MingHua Ma, Zekun Wang, Jingchang Chen, Shan Liping, Qing Yang, Dongliang Xu, Ming Liu, Bing Qin
| Challenge: | Existing LLM pruning works focus on unstructured pruning, which typically requires special hardware support for a practical speed-up. |
| Approach: | They propose a network pruning framework that leverages both coarse and fine-grained activation information as an importance criterion to guide pruning. |
| Outcome: | The proposed framework outperforms existing pruning methods on diverse models across sparsity budgets. |
MapNav: A Novel Memory Representation via Annotated Semantic Maps for VLM-based Vision-and-Language Navigation (2025.acl-long)
Copied to clipboard
Lingfeng Zhang, Xiaoshuai Hao, Qinwen Xu, Qiang Zhang, Xinyao Zhang, Pengwei Wang, Jing Zhang, Zhongyuan Wang, Shanghang Zhang, Renjing Xu
| Challenge: | Vision-language navigation (VLN) is a key task in Embodied AI . traditional approaches rely on historical observations as spatio-temporal contexts for decision making . |
| Approach: | They propose a vision-language navigation model that leverages an annotation system to replace historical frames. |
| Outcome: | The proposed model can be used as a new memory representation method in vision-language navigation . it can be applied to simulated and real-world environments, and it is validated by experiments . |
Can LLM Watermarks Robustly Prevent Unauthorized Knowledge Distillation? (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Model (LLM) watermarking is radioactive and enables the detection of watermarks inherited by student models when trained on the outputs of watermarked teacher models. |
| Approach: | They propose two types of watermark removal attacks that allow student models to perform untraceable knowledge distillation while avoiding watermark inheritance. |
| Outcome: | The proposed attacks eliminate inherited watermarks while maintaining knowledge transfer efficiency and low computational overhead. |
DiffuSeq-v2: Bridging Discrete and Continuous Text Spaces for Accelerated Seq2Seq Diffusion Models (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to text generation use discrete text within a continuous diffusion space, which incurs substantial computational overhead during training and results in slower sampling speeds. |
| Approach: | They propose a soft absorbing state that facilitates diffusion models in learning to reconstruct discrete mutations based on the underlying Gaussian space. |
| Outcome: | The proposed method accelerates training convergence by 4x and generates samples of similar quality 800x faster, rendering it closer to practical application. |
MuTIS: Enhancing Reasoning Efficiency through Multi Turn Intervention Sampling in Reinforcement Learning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for training large reasoning models with long chain-of-thought (CoT) are limited by the number of parameters and the complexity of the model. |
| Approach: | They propose a framework that leverages multi-turn interventions to produce concise reasoning chains and demonstrates strong scalability. |
| Outcome: | The proposed framework breaks the accuracy-efficiency trade-off by producing concise reasoning chains and demonstrating strong scalability on 7B models. |
BRIEF-Pro: Universal Context Compression with Short-to-Long Synthesis for Fast and Accurate Multi-Hop Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Experiments show that BRIEF-Pro generates more concise and relevant summaries, enhancing performance across small, large, and proprietary language models. |
| Approach: | They propose a universal, lightweight compressor that distills relevant evidence from retrieved documents into a concise summary for seamless integration into in-context RAG. |
| Outcome: | Experiments on four open-domain multi-hop question-answering datasets show that BRIEF-Pro generates more concise and relevant summaries, enhancing performance across small, large, and proprietary language models. |
Sensitivity-LoRA : Low-Load Sensitivity-Based Fine-Tuning for Large Language Models (2025.findings-emnlp)
Copied to clipboard
Hao Zhang, Bo Huang, Zhenjia Li, Xi Xiao, Hui Yi Leong, Zumeng Zhang, Xinwei Long, Tianyang Wang, Hao Xu
| Challenge: | Low-Rank Adaptation (LoRA) is a promising approach to adapting LLMs to specialized tasks . existing rank allocation techniques remain computationally inefficient and unstable . |
| Approach: | They propose a low-rank adapted model that approximates model weight updates using low-ranked decomposition. |
| Outcome: | The proposed method is limited by its uniform rank allocation to each incremental matrix . it leverages the second-order derivatives of the loss function to capture weight sensitivity . |
Efficient Transformer Parameter Reuse via Zero-Token Mechanism (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to scaling up parameter counts are impractical for users with limited computational resources. |
| Approach: | They propose a decoupled parameter cycling strategy that employs a head-tail decoupling strategy to decouple the first (head) and last (tail) layers from the parameter cycling process. |
| Outcome: | The proposed approach achieves superior performance under strict parameter constraints and significantly reduces computational overhead via early exits. |
Reward-Augmented Decoding: Efficient Controlled Text Generation With a Unidirectional Reward Model (2023.emnlp-main)
Copied to clipboard
| Challenge: | Large language models generate text that is problematic or lacks a desired attribute. |
| Approach: | They propose a text generation procedure that uses a small unidirectional reward model to encourage a language model to generate text with certain properties. |
| Outcome: | The proposed procedure outperforms prior weighted decoding methods and matches state-of-the-art techniques that require additional training. |
CODER: An efficient framework for improving retrieval through COntextual Document Embedding Reranking (2022.emnlp-main)
Copied to clipboard
| Challenge: | Contextual document embedding reranking is an efficient and efficient retrieval framework. |
| Approach: | They propose a highly efficient retrieval framework that uses contextual document embedding reranking to incorporate ranking context into training. |
| Outcome: | The proposed framework reduces the computational overhead of a first-stage method and can be used as stand-alone retrieval models. |
FreeChunker: A Cross-Granularity Chunking Framework (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing chunking paradigms rely on static boundary identification, limiting performance . Existing methods rely only on static knowledge, resulting in hallucinated content . |
| Approach: | They propose a Cross-Granularity Encoding Framework that treats sentences as atomic units and shifts from static chunk segmentation to flexible retrieval supporting arbitrary sentence combinations. |
| Outcome: | The proposed framework avoids the computational overhead required for semantic boundary detection and enhances adaptability to complex queries. |
FedLEKE: Federated Locate-then-Edit Knowledge Editing for Multi-Client Collaboration (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for updating large language models are inefficient in multi-client scenarios . Existing approaches assume a single-user setting and are ineffective in multiclient scenarios. |
| Approach: | They propose a new task that enables multiple clients to perform LEKE while preserving privacy and reducing computational overhead. |
| Outcome: | The proposed framework outperforms existing LEKE frameworks on two benchmark datasets and retains 96% of performance. |
RAP-ID: Mechanistic Prompt Injection Detection via Impostor Behavior Analysis (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing defenses rely on externally deployed guardrail models or response inspection . current defenses depend on external guardrails or response inspecting . |
| Approach: | They propose a mechanistic, train-free detection framework that operates exclusively on internal state dynamics during the initial forward pass. |
| Outcome: | The proposed framework achieves competitive performance with significant overall improvements compared to heuristic methods. |
Revealing and Enhancing Core Visual Regions: Harnessing Internal Attention Dynamics for Hallucination Mitigation in LVLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing training-free methods are vulnerable to the attention sink phenomenon . Existing methods include contrastive decoding and auxiliary expert models . |
| Approach: | They propose a training-free attention intervention that constructs a PAD map to identify semantically core visual regions and applies per-head Median Absolute Deviation Scaling to adaptively control the intervention strength. |
| Outcome: | The proposed intervention improves visual grounding and reduces hallucinations on multiple LVLMs and benchmarks. |
Monitoring Decoding: Mitigating Hallucination via Evaluating the Factuality of Partial Response during Generation (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to mitiga hallucinations rely on sampling multiple full-length generations, which introduces significant response latency and becomes ineffective when the model consistently produces hallucines. |
| Approach: | They propose a framework that dynamically monitors the generation process and selectively applies in-process interventions to revise hallucination-prone tokens. |
| Outcome: | The proposed framework outperforms self-consistency-based approaches in both effectiveness and efficiency, achieving higher factual accuracy while significantly reducing computational overhead. |
Analyzing and Internalizing Complex Policy Documents for LLM Agents (2026.acl-long)
Copied to clipboard
Jiateng Liu, Zhenhailong Wang, Xiaojiang Huang, Yingjie Li, Xiang Li, Chenlei Guo, Xing Fan, Ruhi Sarikaya, Heng Ji
| Challenge: | Large language model agents rely on in-context policy documents to act as effective user assistants. |
| Approach: | They propose an agentic benchmark generator with Controllable Complexity in agent policy across four levels to evaluate agents under increasing complexity. |
| Outcome: | The proposed method outperforms the baseline in data-sparse and high-complexity settings. |
Improving Efficiency in Large Language Models via Extendable Block Floating Point Representation (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are becoming more and more resource-intensive as their size increases. |
| Approach: | They propose a block floating-point (BFP) arithmetic representation that extends the exponent bit width to capture a wider dynamic range. |
| Outcome: | Extendable Exponent Sharing (EES) outperforms representative baselines in accuracy and computational efficiency. |
CURE: Controlled Unlearning for Robust Embeddings — Mitigating Conceptual Shortcuts in Pre-Trained Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Pre-trained language models are susceptible to spurious, concept-driven correlations that impair robustness and fairness. |
| Approach: | They propose a framework that disentangles and suppresses conceptual shortcuts while preserving essential content information. |
| Outcome: | The proposed framework improves on IMDB and Yelp datasets with minimal computational overhead. |
DoCIA: An Online Document-Level Context Incorporation Agent for Speech Translation (2025.findings-acl)
Copied to clipboard
Xinglin Lyu, Wei Tang, Yuang Li, Xiaofeng Zhao, Ming Zhu, Junhui Li, Yunfei Lu, Min Zhang, Daimeng Wei, Hao Yang, Min Zhang
| Challenge: | Document-level context is crucial for speech translation due to noise from ASR . incorporating document-level contextual information into ST remains a challenge . |
| Approach: | They develop an online framework that integrates document-level context into machine translation . they use document-based modules to integrate document- level context into ST . |
| Outcome: | The proposed framework outperforms baselines in sentence and discourse metrics . it can correct ASR transcription errors and improve translation performance . |
LLM-empowered Dynamic Prompt Routing for Vision-Language Models Tuning under Long-Tailed Distributions (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Pre-trained vision-language models (VLMs) often suffer from bias in class-imbalanced scenes. |
| Approach: | They propose a multi-dimensional dynamic prompt routing framework that integrates a knowledge base for classes spanning multiple visual-semantic dimensions. |
| Outcome: | The proposed framework achieves comparable results with current SOTA methods on long-tailed benchmarks, including CIFAR-LT, ImageNet-LT and Places-LT. |
Federated Data-Efficient Instruction Tuning for Large Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing federated learning (FL) uses all local data, causing excessive computational overhead and overfitting to local data. |
| Approach: | They propose a federated data-efficient instruction tuning approach which utilizes a representative subset of edge-side data to tune LLMs. |
| Outcome: | The proposed method improves Rouge-L on unseen tasks by 10.72% over the SOTA full-data instruction tuning methods while using less than 1.5% of the data samples. |
Towards Efficient and Effective Diffusion Language Model Inference via Semantic-Aware Adaptive Denoising (2026.acl-long)
Copied to clipboard
| Challenge: | Existing acceleration works cannot accurately detect semantically stabilized tokens and then skip computation, leading to sub-optimal speedup in practice. |
| Approach: | They propose a semantic-aware adaptive denoising framework that encodes scalar confidence scores into an evolution-awful feature vector and clusters vectors proactively and adaptively identify semantically converged tokens. |
| Outcome: | The proposed framework outperforms the SOTA competitor in speed and quality . it can detect semantically stabilized tokens and skip computation, resulting in sub-optimal speedup . |
Threshold Differential Attention for Sink-Free, Ultra-Sparse, and Non-Dispersive Language Modeling (2026.acl-long)
Copied to clipboard
| Challenge: | a strict sum-to-one constraint forces attention sinks on irrelevant tokens, while probability mass disperses as sequence lengths increase. |
| Approach: | They propose a sink-free attention mechanism that achieves ultra-sparsity and improved robustness at longer sequence lengths without the computational overhead of projection methods. |
| Outcome: | The proposed mechanism produces >99 % exact zeros and eliminates attention sinks while maintaining competitive performance on standard and long-context benchmarks. |
Length Controlled Generation for Black-box LLMs (2025.acl-long)
Copied to clipboard
Yuxuan Gu, Wenjie Wang, Xiaocheng Feng, Weihong Zhong, Kun Zhu, Lei Huang, Ting Liu, Bing Qin, Tat-Seng Chua
| Challenge: | Existing length control methods involve fine-tuning the parameters of LLMs, which is inefficient and suboptimal for practical use. |
| Approach: | They propose an iterative sampling framework that regulates LLMs to generate length-constrained text without modifying the underlying parameters. |
| Outcome: | The proposed method achieves 100% success rates on Llama3.1 tasks with minimal additional computational overhead. |
PermuteFormer: Efficient Relative Position Encoding for Long Sequences (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing Transformers that scale to long sequences are not compatible with relative position encoding. |
| Approach: | They propose a Performer-based model with relative position encoding that scales linearly on long sequences. |
| Outcome: | The proposed model outperforms performer on long sequences with no computational overhead and outperformed vanilla Transformer on most of the tasks. |
Resource-Friendly Dynamic Enhancement Chain for Multi-Hop Question Answering (2025.findings-acl)
Copied to clipboard
Binquan Ji, Haibo Luo, YifeiLu YifeiLu, Lei Hei, Jiaqi Wang, Tingjing Liao, Wang Lingyu, Shichao Wang, Feiliang Ren
| Challenge: | Existing approaches to solve multi-hop question answering challenges require multiple rounds of retrieval and iterative generation. |
| Approach: | They propose a framework that decomposes complex questions into coherent subquestions . it then iteratively refines these subquests through context-aware rewriting to generate effective query formulations. |
| Outcome: | The proposed framework performs on par with or surpasses state-of-the-art benchmarks while significantly reducing token consumption. |
Efficient Overshadowed Entity Disambiguation by Mitigating Shortcut Learning (2024.emnlp-main)
Copied to clipboard
Panuthep Tasawong, Peerat Limkonchotiwat, Potsawee Manakul, Can Udomcharoenchaikit, Ekapol Chuangsuwanich, Sarana Nutanong
| Challenge: | Entity disambiguation (ED) is crucial in natural language processing tasks such as question-answering and information extraction. |
| Approach: | They propose a method to reduce computational overhead on overshadowed entities by addressing shortcut learning. |
| Outcome: | The proposed method achieves state-of-the-art performance without compromising inference speed. |
Fooling the Textual Fooler via Randomizing Latent Representations (2024.findings-acl)
Copied to clipboard
| Challenge: | Several adversarial attacks can compromise the model without accessing the model architecture or model parameters (i.e., a blackbox setting) Several studies have revealed that deep NLP models are vulnerable to adversarials that slightly perturb the input to cause the models to misbehave. |
| Approach: | They propose a lightweight and attack-agnostic defense that perplexes the process of generating an adversarial example in query-based black-box attacks. |
| Outcome: | The proposed defense is lightweight and attack-agnostic and does not necessitate additional computational overhead during training nor does it rely on assumptions about the potential adversarial perturbation set while having a negligible impact on the model’s accuracy. |
From Tokens to Steps: Verification-Aware Speculative Decoding for Efficient Multi-Step Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Speculative decoding (SD) allows a lightweight draft model to propose outputs that a stronger target model verifies. |
| Approach: | They propose a verification-aware speculative decoding framework that performs step-level verification using only model-internal signals. |
| Outcome: | Experiments show that SpecGuard outperforms both SD and reward-guided SD in accuracy and reliability tests. |
Consistency is Key: On Data-Efficient Modality Transfer in Speech Translation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | End-to-end approaches to speech translation suffer from data scarcity compared to machine translation (MT). |
| Approach: | They propose a method which combines knowledge distillation and consistency learning to break the dilemma of learning-forgetting. |
| Outcome: | The proposed method outperforms the previous methods on a MuST-C dataset even without additional data. |
Agent-Dice: Disentangling Knowledge Updates via Geometric Consensus for Agent Continual Learning (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Model (LLM)-based agents extend the utility of LLMs by interacting with dynamic environments. |
| Approach: | They propose a parameter fusion framework based on directional consensus evaluation that disentangles knowledge updates through a two-stage process. |
| Outcome: | The proposed framework disentangles knowledge updates through a two-stage process with minimal computational overhead and parameter updates. |
A Survey on Proactive Defense Strategies Against Misinformation in Large Language Models (2025.findings-acl)
Copied to clipboard
Shuliang Liu, Hongyi Liu, Aiwei Liu, Duan Bingchen, Zheng Qi, Yibo Yan, He Geng, Peijie Jiang, Jia Liu, Xuming Hu
| Challenge: | Existing methods for detection of misinformation generated by large language models fail to mitigate societal risks . authors propose a paradigm shift from passive detection to anticipatory mitigation strategies . existing defenses remain reactionary in an era demanding proactive defense, authors say . |
| Approach: | They propose a three-pillar approach to prevent misinformation by fortifying integrity of training data and inference reliability by embedding self-corrective mechanisms during reasoning. |
| Outcome: | The proposed framework improves existing methods in misinformation prevention by 63% . it demonstrates that existing methods exhibit false negative rates against misinformation . |
Segment-Based Attention Masking for GPTs (2025.acl-long)
Copied to clipboard
| Challenge: | Causal masking is a fundamental component in Generative Pre-Trained Transformers (GPT) models, playing a crucial role during training. |
| Approach: | They propose to apply causal masking to all input tokens step-by-step, mimicking the generation process. |
| Outcome: | The proposed model can process the entire user prompt at once, but it is applied to all input tokens step-by-step, mimicking the generation process. |
DAC: A Dynamic Attention-aware Approach for Task-Agnostic Prompt Compression (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods rely on information entropy as the metric to compress lexical units, but ignore attention-critical tokens and information . recent advent of In-Context Learning (ICL), Chain-of-Thought (CoT), and Retrieval Augmented Generation (RAG) technologies has significantly invigorated the landscape of applications based on Large Language Models (LLMs). |
| Approach: | They propose a dynamic attention-aware approach to task-agnostic prompt compression . they integrate entropy and attention information to achieve fine-grained prompt compression. |
| Outcome: | Experiments show that the proposed approach improves across tasks and LLMs. |
MAFMO: Multi-modal Adaptive Fusion with Meta-template Optimization for Vision-Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches focus on single-modality adjustments, leading to suboptimal alignment and limited generalization. |
| Approach: | They propose a plug-and-play framework for visual recognition that integrates a Harmonic Cross-Modal Adapter and a Meta-Template Optimization module. |
| Outcome: | Extensive experiments across multiple fine-grained visual recognition benchmarks show that MAFMO consistently improves existing methods’ performance on both novel classes and harmonic mean while maintaining robustness under various challenging conditions with minimal computational overhead. |
VisiPruner: Decoding Discontinuous Cross-Modal Dynamics for Efficient Multimodal LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Multimodal Large Language Models (MLLMs) suffer from significant computational overhead due to the quadratic growth of attention computations with the number of multimodal tokens. |
| Approach: | They propose a training-free pruning framework that prunes multimodal tokens without a trained pruning method. |
| Outcome: | The proposed pruning framework outperforms existing token pruning methods and generalizes across diverse MLLMs. |
Generation-Augmented Retrieval: Rethinking the Role of Large Language Models in Zero-Shot Relation Extraction (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Recent advances in Relation Extraction (RE) emphasize Zero-Shot methodologies, aiming to recognize unseen relations between entities with no annotated data. |
| Approach: | They propose a plug-in retrieval adjuster that allows rapid fine-tuning without accessing LLMs’ parameters. |
| Outcome: | The proposed model demonstrates comparable performance on multiple benchmarks. |
Beyond Prompting: An Efficient Embedding Framework for Open-Domain Question Answering (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have recently pushed open-domain question answering (ODQA) to new heights. |
| Approach: | They propose an embedding-level framework that enhances both the retriever and the reader by reordering query representations via lightweight linear layers under an unsupervised contrastive learning objective. |
| Outcome: | The proposed framework outperforms baselines in accuracy and efficiency across three open-source LLMs, three retrieval methods, and four ODQA benchmarks. |
Rethinking Jailbreak Detection of Large Vision Language Models with Representational Contrastive Scoring (2026.acl-long)
Copied to clipboard
| Challenge: | Large Vision-Language Models (LVLMs) are vulnerable to a growing array of multimodal jailbreak attacks, necessitating a generalizable defense that is efficient for practical deployment. |
| Approach: | They propose a framework that uses a lightweight projection to separate benign and malicious inputs in safety-critical layers. |
| Outcome: | The proposed framework enables a simple yet powerful contrastive score that differentiates true malicious intent from mere distribution shift. |
The Digital Dunning-Kruger Effect: Decoupling Hallucinations via Geometric Hidden-state Observation for Semantic Truthfulness (2026.acl-long)
Copied to clipboard
Yueheng Mao, Min Yu, Gengwang Li, Jianguo Jiang, Gang Li, Meng Zhang, Zhen Xu, Weiqing Huang, Ming Liu
| Challenge: | Large Language Models (LLMs) often generate overconfident yet factually incorrect hallucinations. |
| Approach: | They propose a black-box-based framework that captures stubborn hallucinations by integrating internal geometric dynamics with output probability distributions. |
| Outcome: | The proposed framework outperforms white-box methods and reduces computational overhead by over 90%. |
Synergistic Weak-Strong Collaboration by Aligning Preferences (2025.acl-long)
Copied to clipboard
Yizhu Jiao, Xuchao Zhang, Zhaoyang Wang, Yubo Ma, Zhun Deng, Rujia Wang, Chetan Bansal, Saravan Rajmohan, Jiawei Han, Huaxiu Yao
| Challenge: | Current Large Language Models excel in general reasoning yet struggle with specialized tasks requiring proprietary or domain-specific knowledge. |
| Approach: | They propose a collaborative framework that pairs a specialized weak model with a general strong model to optimize collaboration. |
| Outcome: | The proposed framework outperforms each model alone by leveraging complementary strengths. |
GAMIC: Graph-Aligned Molecular In-context Learning for Molecule Analysis via LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Current methods for retrieving large language models rely on molecule feature similarity, such as Morgan fingerprints, which do not adequately capture the global molecular and atom-binding relationships. |
| Approach: | They propose a self-supervised learning technique that embeds demonstration examples into the input prompt. |
| Outcome: | The proposed technique outperforms simple Morgan-based retrieval methods across tasks by up to 45%. |
KVPR: Efficient LLM Inference with I/O-Aware KV Cache Partial Recomputation (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to inference large language models are limited by CPU capabilities and memory constraints. |
| Approach: | They propose an efficient I/O-aware LLM inference method that overlaps GPU computation with KV cache transfer to minimize idle GPU time. |
| Outcome: | The proposed method reduces the cost of auto-regressive decoding by 35.8% . it also achieves 46.2% higher throughput during decoding compared to state-of-the-art methods. |
DSMoE: Matrix-Partitioned Experts with Dynamic Routing for Computation-Efficient Dense LLMs (2025.emnlp-main)
Copied to clipboard
Minxuan Lv, Zhenpeng Su, Leiyu Pan, Yizhe Xiong, Zijia Lin, Hui Chen, Wei Zhou, Jungong Han, Guiguang Ding, Wenwu Ou, Di Zhang, Kun Gai, Songlin Hu
| Challenge: | Existing sparsification methods like pruning can lose model knowledge through parameter removal. |
| Approach: | They propose a novel approach that achieves sparsification by partitioning pre-trained FFN layers into computational blocks. |
| Outcome: | The proposed approach achieves superior performance across language modeling and downstream tasks under equivalent computational constraints. |
ShortGPT: Layers in Large Language Models are More Redundant Than You Expect (2025.findings-acl)
Copied to clipboard
Xin Men, Mingyu Xu, Qingyu Zhang, Qianhao Yuan, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, Weipeng Chen
| Challenge: | Recent studies have identified significant redundancy in large language models . quantization and pruning are two methods that reduce computational resources . |
| Approach: | They propose simple pruning methods that prune redundant layers based on their BI scores. |
| Outcome: | The proposed pruning methods demonstrate superior performance over previous pruning methods. |
CHESS: Optimizing LLM Inference via Channel-Wise Thresholding and Selective Sparsification (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for activation sparsification do not capture the relationship between activation and model performance. |
| Approach: | They propose a general activation sparsification approach using channel-wise thresholding and selective sparsifying to capture the relationship between activation and model performance. |
| Outcome: | The proposed approach reduces the number of activated neurons during inference by 1.27x over eight downstream tasks while activating fewer parameters than existing methods. |
Causal2Vec: Improving Decoder-only LLMs as Embedding Models through a Contextual Token (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods modify attention mechanism to be bidirectional, undermining LLMs’ ability to extract semantic information acquired during pre-training. |
| Approach: | They propose a general-purpose embedding model that pre-encodes input text into a single Contextual token and then prepends it to the LLM's input sequence. |
| Outcome: | The proposed model improves performance of decoder-only large language models without altering their architectures or introducing significant computational overhead. |
LLMVoX: Autoregressive Streaming Text-to-Speech Model for Any LLM (2025.findings-acl)
Copied to clipboard
Sambal Shikhar, Mohammed Irfan Kurpath, Sahal Shaji Mullappilly, Jean Lahoud, Fahad Shahbaz Khan, Rao Muhammad Anwer, Salman Khan, Hisham Cholakkal
| Challenge: | Existing speech-enabled LLMs degrade conversational quality by modifying the LLM, compromising its linguistic capabilities. |
| Approach: | They propose a lightweight 30M-parameter, LLM-agnostic, autoregressive streaming TTS system that generates high-quality speech with low latency. |
| Outcome: | The proposed system achieves a significantly lower word error rate compared to speech-enabled LLMs while operating at comparable latency. |
GrACE: A Generative Approach to Better Confidence Elicitation and Efficient Test-Time Scaling in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for assessing the reliability of Large Language Models (LLMs) by confidence elicitation require expensive computational overhead or suffer from poor calibration, making them unreliable for real-world deployment. |
| Approach: | They propose a Generative Approach to Confidence Elicitation that enables reliable confidence elicitation for Large Language Models. |
| Outcome: | The proposed method achieves the best discriminative capacity and calibration on open-ended tasks without resorting to additional sampling or an auxiliary model. |
Reinforcement Learning for Large Language Models via Group Preference Reward Shaping (2025.emnlp-main)
Copied to clipboard
Huaisheng Zhu, Siyuan Xu, Hangfan Zhang, Teng Xiao, Zhimeng Guo, Shijie Zhou, Shuyue Hu, Vasant G. Honavar
| Challenge: | Existing methods for fine-tuning Large Language Models (LLMs) are expensive and sensitive to reward model quality. |
| Approach: | They propose a method that leverages preference-based comparisons rather than precise numerical rewards. |
| Outcome: | Experiments show that GPRS outperforms critic-model-free RL algorithms on RLHF and reasoning tasks. |
CENTAUR: Bridging the Impossible Trinity of Privacy, Efficiency, and Performance in Privacy-Preserving Transformer Inference (2025.acl-long)
Copied to clipboard
Jinglong Luo, Guanzhong Chen, Yehong Zhang, Shiyu Liu, Hui Wang, Yue Yu, Xun Zhou, Yuan Qi, Zenglin Xu
| Challenge: | Existing privacy-preserving Transformer Inference frameworks suffer from high computational overhead and performance losses. |
| Approach: | They propose a framework that integrates random permutations and SMPC to address the "impossible trinity" CENTAUR resists diverse data reconstruction attacks and boosts inference speed by 5.030.4 times . |
| Outcome: | CENTAUR achieves an unprecedented balance between privacy, efficiency, and performance. |
GRIT: Guided Relational Integration for Efficient Multi-Table Understanding (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for encoding table structures are limited to single-table settings . end users are increasingly turning to LLMs with natural language queries . |
| Approach: | They propose a method that converts relational schemas into LLM-friendly textual representations. |
| Outcome: | The proposed method improves table-column retrieval performance across multiple tables while reducing memory and computational overhead. |
SecDecoding: Steerable Decoding for Safer LLM Generation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing decoding-time defense methods suffer from limited generalization, high computational overhead, or significant utility degradation. |
| Approach: | They propose a decoding-time defense framework that leverages a pair of small contrastive models to estimate token-level safety signals by measuring divergence in their output distributions. |
| Outcome: | The proposed framework achieves near-zero attack success rates against a wide spectrum of advanced jailbreak attacks while maintaining the model’s helpfulness with minimal degradation. |
Outcome-Grounded Advantage Reshaping for Fine-Grained Credit Assignment in Mathematical Reasoning (2026.acl-long)
Copied to clipboard
Ziheng Li, Liu Kang, Feng Xiao, Luxi Xing, Qingyi Si, Zhuoran Li, Weikang Gong, Deqing Yang, Yanghua Xiao, Hongcheng Guo
| Challenge: | Group Relative Policy Optimization (GRPO) uses a coarse-grained credit assignment mechanism that propagates group-level rewards uniformly to to every token in a sequence, neglecting the varying contribution of individual reasoning steps. |
| Approach: | They introduce Outcome-grounded Advantage Reshaping (OAR) which redistributes advantages based on how much each token influences the model’s final answer. |
| Outcome: | Empirical results show that OAR-G outperforms GRPO on a high-fidelity attribution signal and suppresses low-impact tokens while preserving the advantage mass. |
TreeBoN: Enhancing Inference-Time Alignment with Speculative Tree-Search and Best-of-N Sampling (2025.findings-emnlp)
Copied to clipboard
Jiahao Qiu, Yifu Lu, Yifan Zeng, Jiacheng Guo, Jiayi Geng, Chenhao Zhu, Xinzhe Juan, Ling Yang, Huazheng Wang, Kaixuan Huang, Yue Wu, Mengdi Wang
| Challenge: | Best-of-N (BoN) sampling generates multiple responses and selects the best one, achieving improved performance but with a high computational cost. |
| Approach: | They propose a framework that integrates a speculative tree-search strategy into Best-of-N (BoN) Sampling. |
| Outcome: | The proposed framework outperforms Best-of-N (BoN) sampling but has high computational cost . tree-search strategy reduces computational overhead while maintaining high output quality . |
PRISP: Privacy-Safe Few-Shot Personalization via Lightweight Adaptation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for large language model personalization are limited by data-rich settings and privacy risks. |
| Approach: | They propose a lightweight and privacy-safe personalization framework tailored to constraints in large language models. |
| Outcome: | Experiments on a few-shot variant of the LaMP benchmark show that PRISP achieves strong overall performance compared to prior approaches. |
SpecHub: Provable Acceleration to Multi-Draft Speculative Decoding (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have limited inference speed due to sequential token generation . Spechub is a novel, efficient sampling-verification method for MDSD that improves acceptance rates with only linear computational overhead. |
| Approach: | They propose a method that uses a smaller draft model to generate multiple token sequences . Spechub generates 0.05-0.27 and 0.02-0.16 more tokens per step than RRS and RRS without replacement . |
| Outcome: | The proposed method improves acceptance rates with only linear computational overhead. |
A Joint Optimization Framework for Enhancing Efficiency of Tool Utilization in LLM Agents (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing efforts for tool utilization involve an LLM agent that contains instructions on using the description of the available tools to determine and call the tools required to solve the problem. |
| Approach: | They propose to optimize the context of LLM agents by combining the instructions provided in agent prompts and tool descriptions to enhance their interaction. |
| Outcome: | The proposed framework improves both the instructions provided in agent prompt and tool description, enhancing their interaction. |
Pruning before Fine-tuning: A Retraining-free Compression Framework for Pre-trained Language Models (2024.lrec-main)
Copied to clipboard
| Challenge: | Structured pruning is an effective technique for compressing pre-trained language models (PLMs), but it requires retraining, leading to additional computational overhead. |
| Approach: | They propose a task-specific pruning framework that prunes redundant modules of pre-trained language models before fine-tuning them. |
| Outcome: | The proposed pruning framework achieves higher performance on GLUE, SQUAD, WikiText-2, Wik-103, and PTB datasets while reducing the time required for fine-tuning. |
Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and Adaptation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing LLMs require users to submit raw text regardless of its sensitivity, resulting in substantial computational overhead and degrade model performance. |
| Approach: | They propose a new training pipeline that allows a client-side encoder to condition on k-pooled prompt embeddings instead of raw text and a server-side projection module to fine-tune the projection module and LLM on private, domain-specific data using noise-injected embeddables. |
| Outcome: | The proposed approach eliminates the need for transmitting raw prompt text while maintaining a favorable balance between privacy preservation and model utility for both clients and service providers. |
Don’t Reinvent the Wheel: Efficient Instruction-Following Text Embedding based on Guided Space Transformation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for text embedding require re-encoding the entire corpus for each instruction. |
| Approach: | They propose a framework that generates dynamic text embeddings that adapt to user instructions, highlighting specific attributes of text. |
| Outcome: | The proposed framework improves instruction-following text embedding quality over state-of-the-art methods while speeding up processing on large datasets. |
Attention Basin: Why Contextual Position Matters in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are sensitive to the contextual position of information in input. |
| Approach: | They introduce Attention-Driven Reranking (AttnRank) which estimates a model’s intrinsic positional attention preferences using a small calibration set and reorders retrieved documents or few-shot examples to align the most salient content with these high-attention positions. |
| Outcome: | Experiments on multi-hop QA and few-shot in-context learning tasks show that AttnRank achieves substantial improvements across 10 large language models of varying architectures and scales, without modifying model parameters or training procedures. |
RRAtention: Dynamic Block Sparse Attention via Per-Head Round-Robin Shifts for Long-Context Inference (2026.acl-long)
Copied to clipboard
Siran Liu, Guoxia Wang, Sa Wang, Jinle Zeng, Haoyang Xie, Siyu Lou, Jiabin Yang, Dianhai Yu, Haifeng Wang, Chao Yang
| Challenge: | Existing approaches to dynamic sparse attention require preprocessing, lack global evaluation, violate query independence, or incur high computational overhead. |
| Approach: | They propose a dynamic sparse attention method that achieves all desirable properties through a head **r**ound-**r**obin (RR) sampling strategy. |
| Outcome: | Experiments on natural language understanding and multimodal video comprehension show that the proposed method achieves 2.4 speedup at 128K context length outperforming existing methods. |
TinyJudge: Unverifiable Constraint Alignment via Lightweight Specialist Ensembles (2026.acl-long)
Copied to clipboard
Yirong Zeng, Yufei Liu, Xiao Ding, Yutai Hou, Yuxian Wang, Wu Ning, Haonan Song, Dandan Tu, Qixun Zhang, Yuxiang He, Bibo Cai, Ting Liu
| Challenge: | Instruction Following (IF) is a core capability of LLMs, requiring strict adherence to diverse constraints. |
| Approach: | They propose a framework that uses tiny language models to evaluate instruction following . they propose to use a set of specialized tiny language model to provide rewards for soft constraints. |
| Outcome: | The proposed framework outperforms baseline models by 12% and speeds up training time by 3. |
T-FREE: Subword Tokenizer-Free Generative LLMs via Sparse Representations for Memory-Efficient Embeddings (2024.emnlp-main)
Copied to clipboard
| Challenge: | Tokenizers are crucial for encoding information in Large Language Models, but their development has stagnated. |
| Approach: | They propose a tokenizer that embeds words through sparse activation patterns over character triplets . they show competitive downstream performance with a parameter reduction of more than 85% . |
| Outcome: | The proposed approach achieves competitive downstream performance with a parameter reduction of more than 85% on embedding layers. |
ReflectDiffu: Reflect between Emotion-intent Contagion and Mimicry for Empathetic Response Generation via a RL-Diffusion Framework (2025.acl-long)
Copied to clipboard
| Challenge: | Existing models for empathetic dialogue generation neglect the intricate interplay between emotion and intent, leading to suboptimal controllability of empathy. |
| Approach: | They propose a framework that integrates emotion contagion and intent mimicry to enhance empathetic response generation. |
| Outcome: | The proposed framework outperforms existing models in relevance, controllability, and informativeness. |
Sketch-of-Thought: Efficient LLM Reasoning with Adaptive Cognitive-Inspired Sketching (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have enabled strong reasoning capabilities through Chain-of-Thought (CoT) prompting. |
| Approach: | They propose a framework that integrates cognitively inspired reasoning paradigms with linguistic constraints to reduce token usage while preserving reasoning accuracy. |
| Outcome: | The proposed framework reduces token usage while preserving reasoning accuracy across 18 reasoning datasets across multiple domains, languages, and modalities. |
Demystifying Data Organization for Enhanced LLM Training (2026.acl-long)
Copied to clipboard
Yalun Dai, Yangyu Huang, Tongshen Yang, Yonghan Wang, Xin Zhang, Wenshan Wu, Qihao Zhao, Hao Li, Yuanyuan Gao, Kim-Hui Yap, Scarlett Li
| Challenge: | Large Language Models (LLMs) have revolutionized various fields, yet their training efficiency is heavily reliant on effective data curation. |
| Approach: | They propose to reuse pre-computed sample-level scores originally generated for data efficiency and introduce two new data ordering methods to improve LLM training. |
| Outcome: | The proposed methods improve the stability and performance of LLM training. |
HomoGraphAdapter: A Homogeneous Graph Neural Network as an Effective Adapter for Vision-Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing adaptation methods overlook structural knowledge between text and image modalities or create overly complex graphs containing redundant information for alignment. |
| Approach: | They propose a method to adapt visual models to downstream tasks using text and image modalities. |
| Outcome: | The proposed method improves classification accuracy by 1.51% for 1-shot and 0.74% for 16-shot on 11 datasets. |
Source-primed Multi-turn Conversation Helps Large Language Models Translate Documents (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Document-level machine translations have paved the way for truly simple document-level translation, but challenges such as omission errors remain. |
| Approach: | They propose a method for document-level machine translation that leverages previous contexts in a multi-turn conversational manner by decomposing documents into segments and iteratively translating them while maintaining previous turns. |
| Outcome: | The proposed method outperforms translations of entire documents in a single turn and translations independently according to multiple automatic metrics in representative LLMs. |
Runaway is Ashamed, But Helpful: On the Early-Exit Behavior of Large Language Model-based Agents in Embodied Environments (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Experiments with 4 different LLMs across 5 embodied environments show significant efficiency improvements, with only minor drops in agent performance. |
| Approach: | They propose an intrinsic method that injects exit instructions during generation and an extransic system that verifies task completion to determine when to halt an agent’s trial. |
| Outcome: | The proposed method injects exit instructions during generation and an exit method verifies task completion to determine when to halt an agent’s trial. |
RG-VQA: Leveraging Retriever-Generator Pipelines for Knowledge Intensive Visual Question Answering (2025.findings-emnlp)
Copied to clipboard
Settaluri Lakshmi Sravanthi, Pulkit Agarwal, Debjyoti Mondal, Rituraj Singh, Subhadarshi Panda, Ankit Mishra, Kiran Pradeep, Srihari K B, Godawari Sudhakar Rao, Pushpak Bhattacharyya
| Challenge: | Existing methods to improve the reasoning capabilities of VQA systems are limited due to complexity of graph neural networks and end-to-end training. |
| Approach: | They propose a method to integrate Dense Passage Retrievers with Vision Language Models to boost the reasoning capabilities of VQA systems. |
| Outcome: | The proposed method outperforms human accuracy and GPT-4 in the ScienceQA dataset. |
Efficient Latent Semantic Clustering for Scaling Test-Time Computation of LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for scaling test-time computation rely on external models that introduce substantial computational overhead and fail to capture context-aware semantics. |
| Approach: | They propose a method that leverages the generator LLM’s internal hidden states for clustering, eliminating the need for external models. |
| Outcome: | The proposed method improves the computational efficiency of test-time scaling while maintaining or exceeding the performance of existing methods. |
Odysseus Navigates the Sirens’ Song: Dynamic Focus Decoding for Factual and Diverse Open-Ended Text Generation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing decoding methods struggle to balance factuality and diversity . Deterministic decoding approaches suffer from degeneration and lack of diversity - a problem that is not addressed by the current literature. |
| Approach: | They propose a plug-and-play stochastic approach that adjusts decoding focus based on distributional differences across layers, leveraging the modular nature of factual knowledge within LLMs. |
| Outcome: | Extensive experiments on seven datasets show that DFD significantly improves performance. |
Latent Attention Denoising: A Training-Free Energy-Based Framework for Mitigating Hallucinations in Vision-Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing models for large vision-Language Models lack reliable visual hallucination . despite advances in multimodal perception, they are still limited in real-world applications . |
| Approach: | They propose a framework that recasts attention calibration as a one-step score-based denoising process. |
| Outcome: | The proposed framework achieves superior performance on generative and discriminative tasks while maintaining efficiency comparable to standard decoding. |
Learn Like Humans: Use Meta-cognitive Reflection for Efficient Self-Improvement (2026.acl-long)
Copied to clipboard
| Challenge: | Existing self-improving frameworks rely on inefficient, multi-turn recursive loops that incur high computational costs. |
| Approach: | They propose a framework that achieves efficient self-evolution within a single recurrence cycle. |
| Outcome: | The proposed framework outperforms state-of-the-art self-evolving systems while significantly reducing computational overhead. |
RTTC: Reward-Guided Collaborative Test-Time Compute (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Reward-Guided Test-Time Compute (RTTC) is a powerful paradigm for large language models . indiscriminate application of TTC strategy incurs substantial computational overhead . |
| Approach: | They propose a framework that adaptively selects the most effective TTC strategy for each query via a pretrained reward model. |
| Outcome: | The proposed framework maximizes accuracy across diverse domains and tasks. |
Search-in-Context: Efficient Multi-Hop QA over Long Contexts via Monte Carlo Tree Search with Dynamic KV Retrieval (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to multihop question answering (MHQA) over long contexts are often neglecting explicit reasoning or incurring expensive computational costs due to full-attention mechanisms over long contextuals. |
| Approach: | They propose a framework that integrates Monte Carlo Tree Search (MCTS) with dynamic key-value retrieval to enable iterative, context-aware reasoning. |
| Outcome: | The proposed framework integrates Monte Carlo Tree Search (MCTS) with dynamic key-value (KV) retrieval to enable iterative, context-aware reasoning. |
EGSS: Entropy-guided Stepwise Scaling for Reliable Software Engineering (2026.acl-long)
Copied to clipboard
Chenhui Mao, Yuanting Lei, Zhixiang Wei, Ming Liang, Zhixiang Wang, Jingxuan Xu, Dajun Chen, Wei Jiang, Yong Li
| Challenge: | Entropy-Guided Stepwise Scaling (EGSS) is a novel TTS framework for software engineering tasks. |
| Approach: | They propose an entropy-guided stepwise scaling framework that balances efficiency and effectiveness through entropic-guide encoding and robust test-suite augmentation. |
| Outcome: | EGSS boosts performance by 5–10% across all evaluated models, and reduces inference-time token usage by over 28% . compared to existing methods, EGS reduces token usage and reduce inference time by over 20% . |
Chronos: Learning Temporal Dynamics of Reasoning Chains for Test-Time Scaling (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for testing time scales treat reasoning traces or tokens equally, ignoring substantial variations in trajectory quality and localized logical failures. |
| Approach: | They propose a chronological reasoning scorer that models each trajectory as a time series. |
| Outcome: | The proposed method achieves relative improvements of 34.21% over Pass@128 and 22.70% over Maj@135 on HMMT25, highlighting its effectiveness. |
HASH-RAG: Bridging Deep Hashing with Retriever for Efficient, Fine Retrieval and Augmented Generation (2025.findings-acl)
Copied to clipboard
| Challenge: | Experimental evaluations on NQ, TriviaQA, and HotpotQA datasets demonstrate that our approach achieves a 90% reduction in retrieval time compared to conventional methods while maintaining considerate recall performance. |
| Approach: | They propose a framework that integrates deep hashing techniques with systematic optimizations to address these limitations. |
| Outcome: | The proposed framework outperforms retrieval/non-retrieval baselines by 1.4-4.3% in EM scores on NQ, TriviaQA, and HotpotQA datasets. |
CODEPROMPTZIP: Code-specific Prompt Compression for Retrieval-Augmented Generation in Coding Tasks with LMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing prompt compression techniques for natural language lack fine-grained control over compression ratios. |
| Approach: | They propose a code-aware prompt compression framework for RAG that enables precise length control while preserving critical information. |
| Outcome: | The proposed framework outperforms baselines on three code-related tasks while maintaining the most informative tokens. |
Automated Fine-Grained Mixture-of-Experts Quantization (2025.findings-acl)
Copied to clipboard
| Challenge: | specialized quantization framework for Mixture of Experts architectures is inadequate for model compression. |
| Approach: | They propose a specialized quantization framework for Mixture of Experts architectures . they find that expert networks exhibit distinctive channel-wise outlier distributions ." |
| Outcome: | The proposed framework improves on the Mixtral-8x7b-v0.1 architecture while maintaining minimal computational overhead. |
CollabCoder: Plan-Code Co-Evolution via Collaborative Decision-Making for Efficient Code Generation (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing multi-agent code generation frameworks are constrained by static planning, isolated execution, high computational overhead, and limited adaptability to complex tasks. |
| Approach: | They propose a plan-code co-evolution framework that allows dynamic multi-agent collaboration to improve code quality and robustness across tasks. |
| Outcome: | The proposed framework improves code quality and robustness across tasks while reducing the number of API calls by an average of 4-10 per execution. |
Divide-Then-Aggregate: An Efficient Tool Learning Method via Parallel Tool Invocation (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) demonstrate remarkable capabilities but their ability to autonomously execute complex real-world tasks remains limited. |
| Approach: | They propose a parallel tool invocation framework that decomposes tasks into parallel tool-using subtasks while aggregating results for subsequent decisions. |
| Outcome: | The proposed method significantly improves task performance while reducing token consumption and inference time. |
ORPP: Self-Optimizing Role-playing Prompts to Enhance Language Model Capabilities (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing research has explored model-driven strategies for prompt optimization, but these methods suffer from high computational overhead or require strong optimization capabilities from the model itself, which limits their broad applicability. |
| Approach: | They propose a framework that optimizes and generates role-playing prompts by limiting the prompt search space to role-player scenarios. |
| Outcome: | The proposed framework matches and surpasses existing prompt optimization methods in terms of performance. |
Multi-view-guided Passage Reranking with Large Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing models rely on autoregressive generation and sliding window strategies to rank passages, which incur heavy computational overhead as the number of passages increases. |
| Approach: | They propose a non-generative LLM-based reranking method that encodes query-passage information into diverse view embeddings without being influenced by external biases. |
| Outcome: | The proposed model matches the performance of much larger 7B-scale fine-tuned models while achieving a 100x reduction in inference latency. |
Cross-MoE: An Efficient Temporal Prediction Framework Integrating Textual Modality (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing models ignore dynamic and different relations between time series patterns and textual features, which leads to poor performance in temporal-textual feature fusion. |
| Approach: | They propose a temporal-textual fusion framework that replaces Cross Attention with Cross-Ranker to reduce computational complexity and enhances modality-aware correlation memorization with Mixture-of-Experts (MoE) networks to tolerate the distributional shifts in time series. |
| Outcome: | The proposed framework reduces MSE by 8.78% compared to the current SOTA model and requires only 75% of computational overhead and 12.5% of activated parameters. |
HyperAdaLoRA: Accelerating LoRA Rank Allocation During Training via Hypernetworks without Sacrificing Performance (2026.findings-acl)
Copied to clipboard
Hao Zhang, Zhenjia Li, Yifan Gao, Xi Xiao, Heng Zhang, Shuyang Zhang, null Xiaoxincc, Bo Huang, Yuhang Wu, Tianyang Wang, Hao Xu
| Challenge: | Low-Rank Adaptation (LoRA) assumes a uniform rank r for each incremental matrix, not accounting for the varying significance of weight matrices across modules and layers. |
| Approach: | They propose a framework that allows for faster convergence of low-rank adaptive models . they use a hypernetwork to prune the outputs of the hypernetworks to generate parameters . |
| Outcome: | The proposed framework accelerates convergence of AdaLoRA by leveraging a hypernetwork. |
LDEDE: LRP-Driven Efficient Detection and Editing Framework for LLM Privacy Neurons (2026.findings-acl)
Copied to clipboard
Zhao Zhengyuan, Cao Lifeng, null Sunhaodong, Shi Haotian, Du Xuehui, Liu Aodi, Niu Lanjie, Yang Xiaocheng
| Challenge: | Existing privacy protection methods fail to cover context-dependent sensitive information and are prone to performance degradation. |
| Approach: | They propose a Layer-wise Relevance Propagation-driven framework for efficient privacy neuron detection and editing. |
| Outcome: | The proposed framework achieves 80% higher efficiency than gradient attribution methods while reducing leakage risks of Phone, Email, and medical privacy by 42.7%–73.5% on average and cutting computational time by 60%–90%. |
Resonating with RoPE: Spectral Quantization for High-Fidelity Key Cache Compression (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to reduce memory footprint of long-context LLMs rely on RoPE-induced oscillations. |
| Approach: | They propose a frequency-domain framework that converts RoPE-induced oscillations into sparse spectral representations. |
| Outcome: | The proposed framework achieves efficient compression with performance comparable to FP16 benchmarks. |
Slim-SC: Thought Pruning for Efficient Scaling with Self-Consistency (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies show that Test-Time Scaling (TTS) can improve reasoning performance without retraining the model. |
| Approach: | They propose a step-wise pruning strategy that identifies and removes redundant chains using inter-chain similarity at the thought level. |
| Outcome: | The proposed method reduces inference latency and KVC usage by up to 45% and 26% with R1-Distill while maintaining or improving accuracy. |
Listen, Pause, and Reason: Toward Perception-Grounded Hybrid Reasoning for Audio Understanding (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent Large Audio Language Models (LALMs) have shown strong capabilities in audio understanding, yet their reasoning remains vulnerable to perceptual errors. |
| Approach: | They propose a large-scale dataset for **Perception-Aware Question Answering** that uses a hierarchical decoupling strategy to separate speech from environmental sounds and distinguishes among multiple speakers. |
| Outcome: | The proposed model improves on MMAU-mini, MMAR, and PAQA while maintaining comparable performance on multiple benchmarks. |
Beyond Single Representations: Multi-Model Embedding Fusion for Stable Text Classification (2026.acl-long)
Copied to clipboard
| Challenge: | Existing studies on embedding fusion have not evaluated the effectiveness of individual layers or the impact of combining embeddables from multiple models. |
| Approach: | They propose to combine embeddings from multiple models to improve performance across NLP tasks. |
| Outcome: | The proposed method improves performance on low-resource datasets and reduces the impact of any single model as the number of integrated models increases. |
ConvX: A Lightweight Converter to Bridge Indexed Dense Representations and Large Language Models for Retrieval-Augmented Generation (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing RAG pipelines suffer from critical efficiency limitations due to their complexity and complexity. |
| Approach: | They propose a compression-based RAG framework that directly leverages indexed dense representations produced by a retriever, substituting to long text contexts. |
| Outcome: | Empirical results show that the proposed model achieves competitive performances compared to the state-of-the-art model that uses a large ad-hoc context compressor while offering substantially improved inference efficiency. |
Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models (2026.acl-long)
Copied to clipboard
Woody Haosheng Gan, Deqing Fu, Julian Asilis, Ollie Liu, Vatsal Sharan, Robin Jia, Willie Neiswanger
| Challenge: | Steering methods have emerged as effective tools for guiding large language models’ behavior, yet multimodal large language model lacks comparable techniques due to architectural diversity and limited availability of multimodal steering vectors. |
| Approach: | They validate steering vectors derived solely from text-only LLM backbones and use a cross-modal transfer technique to reuse existing interpretability tools. |
| Outcome: | The proposed steering vectors can guide and enhance multimodal models using SPAR, Mean Shift, and Linear Probing. |
Structure Guided Retrieval-Augmented Generation for Factual Queries (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for RAG produce factually incorrect outputs, resulting in incorrect answers. |
| Approach: | They propose a novel problem that explicitly incorporates structural information into RAG for factual questions to satisfy all query conditions. |
| Outcome: | The proposed method significantly outperforms baselines on ERQA while maintaining reasonable computational overhead. |
Long-Context Reasoning Through Proxy-Based Chain-of-Thought Tuning (2026.acl-long)
Copied to clipboard
| Challenge: | Recent large language models support inputs of up to 10 million tokens, yet they perform poorly on long-context tasks. |
| Approach: | They propose a training framework that transfers reasoning capabilities from proxy contexts to full long contexts. |
| Outcome: | The proposed framework outperforms baseline models with reduced computational overhead. |
CRISP: Compressing Redundancy in Chain-of-Thought via Intrinsic Saliency Pruning (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to compress CoT with external compressors fail to align with the model’s internal reasoning dynamics, resulting in the loss of critical logical steps. |
| Approach: | They propose a framework that exploits the model’s intrinsic saliency to compress CoT by exploiting its reasoning termination token . |
| Outcome: | The proposed framework reduces redundancy in reasoning chain by exploiting the model’s intrinsic saliency. |
Thinking Economically: A Hierarchical Framework for Adaptive-Complexity Reasoning in LLMs (2026.findings-acl)
Copied to clipboard
Yubo Gao, Haotian Wu, Hong Chen, Junquan Huang, Yibo Yan, Jungang Li, Zihao Dongfang, Sicheng Tao, PS Tan, Jie Zhang, Xuming Hu
| Challenge: | Existing efficiency methods for Chain-of-Thought (CoT) generate excessively long rationales without commensurate accuracy gains. |
| Approach: | They propose a training framework that operationalizes this principle through coarse-to-fine budgeting. |
| Outcome: | Experiments on GSM8K and MATH500 show that HAB surpasses standard CoT in accuracy and reduces token usage, achieving stronger performance-efficiency trade-off than baselines. |
ACBQ: Adaptive Cross-Block Quantization of Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for post-training quantization struggle to support weight–activation joint quantization and extreme low-bit weight quantization. |
| Approach: | They propose a framework that addresses weight–activation joint quantization and extreme weight quantization. |
| Outcome: | The proposed framework achieves superior performance under both W4A4 and highly aggressive W2 settings while incurring negligible additional computational overhead. |
DynamixSFT: Dynamic Mixture Optimization of Instruction Tuning Collections (2026.findings-acl)
Copied to clipboard
| Challenge: | Several studies rely on additional models to optimize mixtures. |
| Approach: | They propose a method that dynamically optimizes instruction-tuning dataset mixtures by prior-scaled Boltzmann Exploration and a multi-armed bandit setup. |
| Outcome: | The proposed method improves the TÜLU-2-mixture and TÜLO-3-mixtures across 10 benchmarks while introducing minimal computational overhead over naive sampling. |
Small Data, Big Noise: Adversarial Training for Robust ParameterEfficient Fine-Tuning (2026.findings-acl)
Copied to clipboard
| Challenge: | Parameter-Efficient Fine-Tuning (PEFT) is essential for adapting foundation models to downstream tasks, but current methods struggle with robustness to noise and performance degradation on limited training data. |
| Approach: | They propose a framework that brings adversarial training to PEFT to enhance model robustness and generalization, outperforming alternative approaches. |
| Outcome: | Experiments with two variants of the proposed framework show that it outperforms existing methods in low-resource settings and under word-level and character-level corruptions. |
Render-of-Thought: Rendering Textual Chain-of-Thought as Images for Visual Latent Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Recent work on Chain-of-Thought prompting imposes substantial computational overhead . lack of supervision obscures the analyzability of the latent reasoning chain. |
| Approach: | They propose a framework to render latent reasoning chain into images, making latent rationale explicit and traceable. |
| Outcome: | The proposed framework achieves 3-4 token compression and substantial inference acceleration compared to explicit CoT prompting. |
Tandem: Riding Together with Large and Small Language Models for Efficient Reasoning (2026.findings-acl)
Copied to clipboard
Zichuan Fu, Xian Wu, Guojing Li, Yejing Wang, Yijun Chen, Zhao Zihao, Luo Yixuan, Hanyu Yan, Yefeng Zheng, Xiangyu Zhao
| Challenge: | Recent advances in large language models (LLMs) have catalyzed the rise of reasoningintensive inference paradigms, where models perform explicit step-by-step reasoning before generating final answers. |
| Approach: | They propose a large-small LLM collaboration framework that synergizes large and small language models to achieve high-quality reasoning with significantly reduced computational cost. |
| Outcome: | The proposed framework outperforms the mentor LLM while preserving the benefits of the thinking paradigm of LLMs. |
Joint Optimization of Training Data and Policy in RLHF (2026.findings-acl)
Copied to clipboard
Zhuohao Yu, Jiali Zeng, Weizheng Gu, Mengyuan Sun, Yidong Wang, Fandong Meng, Jie Zhou, Shikun Zhang, Wei Ye
| Challenge: | JODP optimizes policies on fixed training inputs, limiting the diversity of learning signals. |
| Approach: | They propose a framework where policy generates improved variants of training problems to enhance its own learning. |
| Outcome: | The proposed framework improves on safety alignment tasks by allowing 4B models to reach 8B model performance with less than 1% additional computational overhead. |
EconProver: Towards More Economical Test-Time Scaling for Automated Theorem Proving (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have recently advanced the field of Automated Theorem Proving (ATP) Existing cost analyses regulate only the number of sampling passes, ignoring the substantial disparities in sampling costs. |
| Approach: | They propose to integrate two complementary methods into a unified EconRL pipeline to increase pass rates under constrained sampling passes. |
| Outcome: | The proposed method reduces token usage and sample passes while maintaining the original performance. |
Thinking-Based Non-Thinking: Solving the Reward Hacking Problem in Training Hybrid Reasoning Models via Reinforcement Learning (2026.acl-long)
Copied to clipboard
Siyuan Gan, Jiaheng Liu, Boyan Wang, Tianpei Yang, Runqing Miao, Yuyao Zhang, Fanyu Meng, Junlan Feng, Linjian Meng, Jing Huo, Yang Gao
| Challenge: | Existing work on large reasoning models (LRMs) focuses on using reinforcement learning (RL) to train hybrid reasoning models that automatically decide whether to engage in thinking or not based on the complexity of the query. |
| Approach: | They propose to use reinforcement learning to train hybrid reasoning models that automatically decide whether to engage in thinking or not based on the complexity of the query. |
| Outcome: | The proposed model reduces token usage by around 50%$ compared to DeepSeek-R1-Distill-Qwen-1.5B/7B and DeepScaleR-1.5b, while significantly improving accuracy. |
Defending LLMs against Jailbreak Attacks via Template-Based ICL with a Defensive Suffix (2026.findings-acl)
Copied to clipboard
| Challenge: | State-of-the-art large language models (LLMs) are vulnerable to jailbreak attacks, such as GCG and AutoDAN. |
| Approach: | They propose to take the advances of online In-Context Learning and an offline defensive suffix and optimize them using an iterative algorithm and an online stochastic random search to identify the most effective ICL demonstrations. |
| Outcome: | The proposed method reduces attack success rate to nearly *0% while maintaining the model’s utility on benign tasks and incurring only *negligible* computational overhead. |
Dictionary Guided Sparse Logit Editing for Reliable Jailbreak Attacks (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to optimize large language models suffer from high computational costs and produce uninterpretable, high-perplexity inputs. |
| Approach: | They propose a sparse index-based intervention that bypasses guardrails via sparser logit editing. |
| Outcome: | The proposed method bypasses guardrails by modifying pre-softmax logits without gradients or auxiliary models. |
Preference Heads in Large Language Models: A Mechanistic Framework for Interpretable Personalization (2026.acl-long)
Copied to clipboard
Weixu Zhang, Ye Yuan, Changjiang Han, Yuxing Tian, Zipeng Sun, Linfeng Du, Jikun Kang, Hong Kang, Xue Liu, Haolun Wu
| Challenge: | Large Language Models exhibit strong implicit personalization ability, but most approaches treat this behavior as a black box. |
| Approach: | They propose a mechanistic interpretation perspective and propose 'sparse' set of Preference Heads . they compute a Preference Contribution Score for each attention head and compare their predictions . |
| Outcome: | The proposed framework computes a Preference Contribution Score (PCS) for each attention head and measures its causal impact on user aligned outputs. |
Influence-based Online Experience Selection for Effective RLHF (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for RL fail to establish an interpretable connection between data and optimization objectives. |
| Approach: | They propose a data selection method that dynamically estimates the influence of individual training samples on policy optimization. |
| Outcome: | The proposed method significantly improves training effectiveness with fewer optimization steps. |