Papers with convergence
BitDistiller: Unleashing the Potential of Sub-4-Bit LLMs via Self-Distillation (2024.acl-long)
Copied to clipboard
| Challenge: | Weight quantization has emerged as a popular solution to reduce memory and computational demands. |
| Approach: | They propose a framework that synergizes Quantization-Aware Training (QAT) with Knowledge Distillation (KD) to boost the performance of LLMs at sub-4-bit. |
| Outcome: | The proposed framework outperforms existing QAT methods on language understanding and complex reasoning benchmarks on sub-4-bit models. |
The Importance of Being Parameters: An Intra-Distillation Method for Serious Gains (2022.emnlp-main)
Copied to clipboard
| Challenge: | Recent pruning methods remove redundant parameters according to parameter sensitivity, a gradient-based measure reflecting the contribution of the parameters. |
| Approach: | They propose a general task-agnostic method to balance parameter sensitivity and a novel adaptive learning method to control strength of intra-distillation loss for faster convergence. |
| Outcome: | The proposed method can reduce redundant parameters by over 80% without obvious performance degradation. |
Your Pretrained Model Tells the Difficulty Itself: A Self-Adaptive Curriculum Learning Paradigm for Natural Language Understanding (2025.acl-srw)
Copied to clipboard
| Challenge: | Existing curriculum learning approaches rely on manually defined difficulty metrics which may not accurately reflect the model’s own perspective. |
| Approach: | They propose a self-adaptive curriculum learning paradigm that prioritizes fine-tuning examples based on difficulty scores predicted by pre-trained language models (PLMs) they evaluate four datasets covering binary and multi-class classification tasks. |
| Outcome: | The proposed model leads to faster convergence and improved performance compared to standard random sampling. |
How does Multi-Task Training Affect Transformer In-Context Capabilities? Investigations with Function Classes (2024.naacl-short)
Copied to clipboard
| Challenge: | Multi-task learning (MTL) for generalist models is a promising direction that offers transfer learning potential. |
| Approach: | They propose to combine multi-task learning (MTL) with in-context learning (ICL) to build models that can generalize to multiple tasks while being robust to out-of-distribution examples. |
| Outcome: | The proposed training strategies enable models to learn difficult tasks while mixing in prior tasks, denoted as mixed curriculum. |
Arabic Dialect Identification with a Few Labeled Examples Using Generative Adversarial Networks (2022.aacl-main)
Copied to clipboard
| Challenge: | Experimental results show that transformer-based models can handle Dialect Arabic (DA) classification tasks with a large corpus of labeled examples. |
| Approach: | They extend transformer-based models with unlabeled data in a generative adversarial setting using semi-supervised Generative Adversarial Networks (SS-GAN) they show that the model can produce high-quality embeddings for the Dialect Arabic examples and generalize for the downstream classification task given few labeled examples. |
| Outcome: | The proposed model outperforms models with unlabeled data in a generative adversarial setting with unlabelled examples and faster convergence when only a few labeled examples are available. |
Data Pruning for Efficient Model Pruning in Neural Machine Translation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large-scale pre-trained language models have demonstrated encouraging performance in various NLP tasks at the cost of over-parametrized networks and high memory requirements. |
| Approach: | They combine data pruning with movement pruning for Neural Machine Translation to enable efficient fine-pruning by leveraging cross-entropy scores of individual training instances. |
| Outcome: | The proposed pruning strategy outperforms other pruning methods on a translation task and shows that training cross-entropy scores can reduce the steps required for convergence and training time. |
CEAN: Contrastive Event Aggregation Network with LLM-based Augmentation for Event Extraction (2024.eacl-long)
Copied to clipboard
| Challenge: | Event Extraction is a crucial yet arduous task in natural language processing (NLP), as its performance is hindered by laborious data annotation. |
| Approach: | They propose a Contrastive Event Aggregation Network with LLM-based Augmentation to promote low-resource learning and reduce data noise for event extraction. |
| Outcome: | The proposed approach achieves new state-of-the-art results on the ACE2005 and ERE-EN datasets. |
Towards Federated Low-Rank Adaptation of Language Models with Rank Heterogeneity (2025.naacl-short)
Copied to clipboard
| Challenge: | Low-rank adaptation (LoRA) is an efficient alternative to full-weight adaptation in federated fine-tuning of language models, significantly reducing computational costs. |
| Approach: | They propose a low-rank adaptation method that freezes original weights and trains only the update parametrized as a product of two low-ranked matrices. |
| Outcome: | The proposed method accelerates convergence and enhances the global model’s predictive performance. |
CASPER: Bridging Discrete and Continuous Prompt Optimization through Feedback-Guided Gradient Descent (2026.eacl-industry)
Copied to clipboard
| Challenge: | Existing pipelines for generative tasks require extensive manual effort and domain expertise to achieve task-optimal performance. |
| Approach: | They propose a framework bridging discrete and continuous prompt optimization through feedback-guided gradient descent in embedding space. |
| Outcome: | The proposed framework bridges discrete and continuous prompt optimization through feedback-guided gradient descent in embedding space. |
LoRA-MGPO: Mitigating Double Descent in Low-Rank Adaptation via Momentum-Guided Perturbation Optimization (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Low-Rank Adaptation (LoRA) adapts large language models by training only a small fraction of parameters, but as the rank of the low-rank matrices increases, LoRA exhibits an unstable “double descent” phenomenon, which delays convergence and impairs generalization by causing instability due to the attraction to sharp local minima. |
| Approach: | They propose a framework that incorporates Momentum-Guided Perturbation Optimization (MGPO) MGPO stabilizes training dynamics by mitigating double descent phenomenon and guiding weight perturbations using momentum vectors from the optimizer’s state. |
| Outcome: | The proposed framework improves performance on natural language understanding benchmarks and shows that it improves convergence and generalization. |
Deep Dirichlet Multinomial Regression (N18-1)
Copied to clipboard
| Challenge: | supervised topic models can incorporate arbitrary document-level features to inform topic priors, but their ability to model corpora is limited by the representation and selection of these features. |
| Approach: | They propose a generative topic model that simultaneously learns document feature representations and topics. |
| Outcome: | The proposed model outperforms DMR and LDA on three datasets and human subjects judge it more representative of associated document features. |
Do language models accommodate their users? A study of linguistic convergence (2026.eacl-long)
Copied to clipboard
| Challenge: | In this paper, we examine how large language models adapt their language use to the linguistic patterns of their user. |
| Approach: | They examine whether large language models exhibit linguistic convergence, a pragmatic element of human language communication, and compare their results to original human responses. |
| Outcome: | The proposed model language use is significantly different from that of humans. |
Anti-Overestimation Dialogue Policy Learning for Task-Completion Dialogue System (2022.findings-naacl)
Copied to clipboard
| Challenge: | Recent research has focused on reinforcement learning (RL)-based dialogue policy. |
| Approach: | They propose a dynamic partial average estimator (DPAV) of the ground truth maximum action value to solve the overestimation problem. |
| Outcome: | The proposed method achieves better results on three dialogue datasets with a lower computational load compared to baselines on three different domains with lower bias. |
AdaZeta: Adaptive Zeroth-Order Tensor-Train Adaption for Memory-Efficient Large Language Models Fine-Tuning (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in memory-efficient zeroth-order methods have limited their widespread adoption due to performance drops and a high risk of divergence. |
| Approach: | They propose a memory-efficient zeroth-order framework to improve performance and convergence of the MeZO methods by using only forward passes. |
| Outcome: | The proposed framework improves performance and convergence of the proposed methods on Roberta-Large and Llama-2-7B models. |
GTA: Supervised-Guided Reinforcement Learning for Text Classification with Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Reinforcement learning fine-tuning methods suffer from inefficient exploration and slow convergence . supervised fine- tuning methods have limited performance ceiling and less solid theoretical foundation . |
| Approach: | They propose a Guess-Think-Answer framework that combines supervised and supervised learning in a unified training paradigm. |
| Outcome: | The proposed framework outperforms both standalone SFT and RL training models on three text classification benchmarks. |
Why is unsupervised alignment of English embeddings from different algorithms so hard? (D18-1)
Copied to clipboard
| Challenge: | a new paper challenges word embedding algorithms to align independent English word embeds with 100% precision . authors show that when two different embeddables are used, they fail to do so . |
| Approach: | They propose to use unsupervised bilingual dictionary induction to study English-English alignments. |
| Outcome: | The proposed approach is more of a challenge than a technical contribution . it shows that the results challenge unsupervised bilingual dictionary induction algorithms . |
Negating Negatives: Alignment with Human Negative Samples via Distributional Dispreference Optimization (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to steer LLMs towards human preference suffer from noisy positive-negative training pairs. |
| Approach: | They propose a distributional preference optimization method which maximizes discrepancy between dispreferred responses and generated non-negative ones. |
| Outcome: | The proposed method achieves comparable generation quality and surpasses the latest strong baselines in producing less harmful and more informative responses with better training stability and faster convergence. |
Memory-Efficient Structured Backpropagation for On-Device LLM Fine-Tuning (2026.acl-industry)
Copied to clipboard
| Challenge: | Existing approaches for fine-tuning large language models require a trade-off between exact gradients with high memory and low memory with noisy estimates (MeZO). |
| Approach: | They propose a method which derivates gradients from LoRA's low-rank structure and manually deriving backward passes to exploit the low-level structure. |
| Outcome: | The proposed method reduces peak memory from 361MB to 136MB for Qwen2.5-0.5B, enabling fine-tuning scenarios previously infeasible on memory-constrained devices. |
Teacher Intervention: Improving Convergence of Quantization Aware Training for Ultra-Low Precision Transformers (2023.eacl-main)
Copied to clipboard
| Challenge: | Quantization-aware training (QAT) is a promising method to lower the implementation cost and energy consumption. |
| Approach: | They propose a method for fast converging QAT of pre-trained Transformers using a layer-wise signal propagation method with the intact signal from the teacher. |
| Outcome: | The proposed method achieves superior accuracy with significantly lower fine-tuning iterations on well-known Transformers of natural language processing as well as computer vision compared to the state-of-the-art methods. |
OFA: A Framework of Initializing Unseen Subword Embeddings for Efficient Large-scale Multilingual Continued Pretraining (2024.findings-naacl)
Copied to clipboard
| Challenge: | Existing methods to pretrain multilingual models are limited by the number of embedding parameters and the complexity of the model. |
| Approach: | They propose a framework that initializes the embeddings of unseen subwords and can adapt a model to multiple languages. |
| Outcome: | The proposed framework can adapt a pre-trained model to multiple languages efficiently and effectively. |
NLoRA: Nyström-Initiated Low-Rank Adaptation for Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Parameter-efficient fine-tuning is essential for adapting large language models (LLMs). However, LoRA suffers from slow convergence and some recent LoRA variants, such as PiSSA, rely on Singular Value Decomposition (SVD) for initialization. |
| Approach: | They propose to introduce a small intermediate matrix between the low-rank matrices (A) and (B) and propose NyströmLoRA (NLoRA) which leverages Nyström-based initialization for SLoRA to improve its effectiveness and efficiency. |
| Outcome: | The proposed approach improves on 5 natural language generation tasks and 8 natural language understanding tasks with minimal parameter overhead. |
Client-Customized Adaptation for Parameter-Efficient Federated Learning (2023.findings-acl)
Copied to clipboard
| Challenge: | Pre-trained language models have a large memory footprint and are difficult to use in federated learning (FL) |
| Approach: | They propose a hypernetwork-based FL framework that generates client-specific adapters by conditioning the client information. |
| Outcome: | The proposed framework maximizes the utility of shared model parameters while minimizing divergence caused by client heterogeneity. |
Multilingual Machine Translation with Hyper-Adapters (2022.emnlp-main)
Copied to clipboard
| Challenge: | Multilingual machine translation suffers from negative interference across languages. |
| Approach: | They propose a rescaling fix that reduces the number of parameters and enables training larger hyper-networks. |
| Outcome: | The proposed approach outperforms regular adapters and achieves the same performance with 12 times less parameters. |
Exploiting Curriculum Learning in Unsupervised Neural Machine Translation (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Experimental results show that the proposed method achieves consistent improvements with faster convergence speed. |
| Approach: | They propose a curriculum learning method to gradually utilize pseudo bi-texts based on their quality from multiple granularities. |
| Outcome: | The proposed method achieves consistent improvements with faster convergence speed on WMT 14 En-Fr, WMT14 En-De, and LDC En-Zh translation tasks. |
RoZO: Geometry-Aware Zeroth-Order Fine-Tuning on Low-Rank Adapters for Black-Box Large Language Models (2026.eacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated exceptional performance across a wide range of tasks, yet fine-tuning them efficiently under black-box or memory-constrained settings remains challenging. |
| Approach: | They propose a Riemannian zeroth-order optimization framework that constrains updates to the tangent space of the LoRA manifold. |
| Outcome: | The proposed framework achieves more stable convergence, tighter variance bounds, and superior performance compared to existing ZO methods. |
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. |
Improving Reinforcement Learning Based Image Captioning with Natural Language Prior (D18-1)
Copied to clipboard
| Challenge: | Recent research shows that Reinforcement Learning (RL) approaches suffer from the exposure bias problem. |
| Approach: | They propose a Reinforcement Learning (RL) based training framework that constrains the action space using an n-gram language prior. |
| Outcome: | The proposed model is more human readable and graceful. |
Improving Visual-Semantic Embedding with Adaptive Pooling and Optimization Objective (2023.eacl-main)
Copied to clipboard
Zijian Zhang, Chang Shu, Ya Xiao, Yuan Shen, Di Zhu, Youxin Chen, Jing Xiao, Jey Han Lau, Qian Zhang, Zheng Lu
| Challenge: | Recent VSE models combine simple pooling methods with hard triplet loss to improve performance. |
| Approach: | They propose an adaptive pooling strategy that allows the model to learn how to aggregate features through a combination of simple pooling methods. |
| Outcome: | The proposed strategy outperforms current state-of-the-art systems on image-to-text and text-toimage retrieval. |
Adversarial Contrastive Estimation (P18-1)
Copied to clipboard
| Challenge: | Noise contrastive estimation (NCE) is a general strategy used in word embeddings and translations for knowledge graphs. |
| Approach: | They propose to augment negative sampler into mixture distribution with adversarially learned sampler and to combine it with noise contrastive estimation (NCE) they observe faster convergence and improved results on multiple metrics. |
| Outcome: | The proposed model performs better on word embeddings, order embedds and knowledge graph embeddments and faster convergence and improved results on multiple metrics. |
From Data-Centric to Sample-Centric: Enhancing LLM Reasoning via Progressive Optimization (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) has recently advanced the reasoning capabilities of large language models (LLMs). |
| Approach: | They propose a method that incorporates partial solution prefixes from expert demonstrations to guide the policy. |
| Outcome: | The proposed methods outperform strong baselines, yielding faster convergence and a higher performance ceiling. |
SQUIRE: A Sequence-to-sequence Framework for Multi-hop Knowledge Graph Reasoning (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for multi-hop knowledge graph reasoning suffer from slow and poor convergence . a transformer model can be used to learn and predict in an end-to-end fashion, giving faster convergence compared to previous methods . |
| Approach: | They propose a Sequence-to-sequence based multi-hop reasoning framework . it uses an encoder-decoder transformer structure to translate the query to a path . |
| Outcome: | The proposed framework can learn and predict in an end-to-end fashion, which gives better and faster convergence. |
Layer-wise Importance Matters: Less Memory for Better Performance in Parameter-efficient Fine-tuning of Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Parameter-Efficient Fine-Tuning (PEFT) methods have gained popularity for adapting pre-trained Large Language Models (LLMs) to downstream tasks. |
| Approach: | They propose a method to optimize the importance of full layers with layer-wise importance scoring by leveraging the estimated importance scores. |
| Outcome: | The proposed method is compatible with PEFT methods that operate on a per-layer basis and achieves better performance. |
Data Selection Curriculum for Neural Machine Translation (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Neural Machine Translation models are typically trained on heterogeneous data that are concatenated and randomly shuffled. |
| Approach: | They propose a two-stage curriculum training framework where a NMT model is fine-tuned on subsets of data, selected by deterministic scoring and online scoring. |
| Outcome: | The proposed framework improves on six language pairs comprising low- and high-resource languages and shows up to +2.2 BLEU improvement and faster convergence. |
Natural Language to Structured Query Generation via Meta-Learning (N18-2)
Copied to clipboard
| Challenge: | Conventional supervised training is a pervasive paradigm for NLP problems . however, examples of the same problem may vary widely . a few-shot meta-learning scenario is used to learn multiple models . |
| Approach: | They propose a learning protocol that treats each example as a unique pseudo-task . they use a few-shot meta-learning scenario to reduce the original learning problem to a single example . |
| Outcome: | The proposed learning protocol achieves 1.1%–5.4% accuracy gains over non-meta-learning counterparts on a WikiSQL dataset. |
Found in the Middle: Permutation Self-Consistency Improves Listwise Ranking in Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Large language models exhibit positional bias in how they use context, which affects listwise ranking. |
| Approach: | They propose a method to marginalize out different list orders in the prompt to produce an order-independent ranking with less positional bias. |
| Outcome: | The proposed method improves on five datasets in sorting and passage reranking by 34-52% . it marginalizes out different list orders in the prompt to produce an order-independent ranking . |
LTRS: Improving Word Sense Disambiguation via Learning to Rank Senses (2025.coling-main)
Copied to clipboard
| Challenge: | Conventional training strategies only consider predefined senses for target words and learn each of them from relatively limited instances, neglecting the influence of similar ones. |
| Approach: | They propose a method to rank senses to improve the task of word Sense Disambiguation (WSD) by ranking an expanded list of sense definitions. |
| Outcome: | The proposed method achieves a SOTA F1 score of 79.6% in Chinese WSD and shows faster convergence than previous methods. |
TEN: Table Explicitization, Neurosymbolically (2026.acl-industry)
Copied to clipboard
| Challenge: | Existing methods for extracting tabular data from semistructured text are error-prone and costly. |
| Approach: | They propose a neurosymbolic approach to extract tabular data from semistructured text . TEN is a triadic feedback loop that iteratively refines table hypotheses . |
| Outcome: | The proposed approach outperforms neural baselines in exact match accuracy and lower hallucination rates. |
Self-supervised Rewiring of Pre-trained Speech Encoders:
Towards Faster Fine-tuning with Less Labels in Speech Processing (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Pre-trained speech encoders have facilitated great success across various speech processing tasks, but fine-tuning them for downstream tasks requires large training data to converge or to achieve state-of-the-art. |
| Approach: | They propose to rewire pre-trained speech encoders to improve their representation space without task-specific labels by neutrally synthesising audio inputs and frame masking. |
| Outcome: | The proposed model shows consistent improvement in isotropy in the representation space on 6 speech processing tasks. |
Taming LLMs with Gradient Grouping (2025.acl-long)
Copied to clipboard
| Challenge: | a new study presents scaling with gradient grouping (SGG) the adaptive learning rate scaling approach is based on per-parameter statistics, which incurs memory overhead. |
| Approach: | They propose an optimizer wrapper that improves adaptive learning rate estimation by dynamic grouping and group-specific scaling. |
| Outcome: | The proposed algorithm improves learning rate estimation on diverse models with different model sizes and batch sizes. |
Training Dynamics for Curriculum Learning: A Study on Monolingual and Cross-lingual NLU (2022.emnlp-main)
Copied to clipboard
| Challenge: | Current approaches for NLU use CL to improve in-distribution data performance via heuristic-oriented or task-agnostic difficulties. |
| Approach: | They propose to use CL to improve in-distribution data performance by taking advantage of training dynamics as difficulty metrics instead of heuristic-oriented or task-agnostic difficulties. |
| Outcome: | The proposed model schedulers improve on in-distribution, out-of-distortion and zero-shot cross-lingual transfer datasets while being 20% faster on average. |
An Efficient Task-Oriented Dialogue Policy: Evolutionary Reinforcement Learning Injected by Elite Individuals (2025.acl-long)
Copied to clipboard
| Challenge: | Evolutionary Algorithms (EAs) have been proven to effectively explore the solution space of neural networks by maintaining population diversity. |
| Approach: | They propose an elite individual injection mechanism to enhance EA’s search efficiency by adaptively introducing best-performing individuals into the population. |
| Outcome: | Experiments on four datasets show that the proposed approach significantly improves the balance between exploration and exploitation, boosting performance. |
Upsample or Upweight? Balanced Training on Heavily Imbalanced Datasets (2025.naacl-long)
Copied to clipboard
| Challenge: | a lack of data across domains creates significant imbalances in training data sizes . a recent study shows that temperature sampling and scaling are equivalent but differ under stochastic gradient descent due to differences in gradient variance. |
| Approach: | They propose a method that upsamples low-resource languages and upweights their loss functions to address this disparity. |
| Outcome: | The proposed method competes effectively with existing data re-weighting techniques while offering computational efficiency. |
Shallow Focus, Deep Fixes: Enhancing Shallow Layers Vision Attention Sinks to Alleviate Hallucination in LVLMs (2025.emnlp-main)
Copied to clipboard
Xiaofeng Zhang, Yihao Quan, Chen Shen, Chaochen Gu, Xiaosong Yuan, Shaotian Yan, Jiawei Cao, Hao Cheng, Kaijie Wu, Jieping Ye
| Challenge: | Multimodal large language models (MLLMs) demonstrate excellent abilities for understanding visual information, but the hallucination remains a challenging problem. |
| Approach: | They propose a training-free approach to enhance vision attention sinks to facilitate convergence of the image token attention sink within shallow layers. |
| Outcome: | The proposed approach improves the convergence of the image token attention sink within shallow layers and strengthens the layer’s focus on the image itself. |
Can We Predict Before Executing Machine Learning Agents? (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to scientific discovery rely on expensive physical execution . a Generate-Execute-Feedback paradigm is costly and slow . |
| Approach: | They propose to internalize execution priors to substitute costly runtime checks with instantaneous predictive reasoning, drawing inspiration from World Models. |
| Outcome: | The proposed framework achieves 61.5% accuracy and robust confidence calibration when primed with a Verified Data Analysis Report. |
Exploring the Relationship between In-Context Learning and Instruction Tuning (2024.findings-emnlp)
Copied to clipboard
| Challenge: | In-Context Learning (ICL) and Instruction Tuning (IT) are two primary paradigms of adopting Large Language Models (LLMs) to downstream applications, but they are significantly different. |
| Approach: | They examine how the hidden states of Large Language Models change in these two paradigms by examining how they differ in implementation. |
| Outcome: | The proposed model changes the hidden states of LLMs as if its accompanying demonstrations were used to instructionally tune the model. |
Re-entry Prediction for Online Conversations via Self-Supervised Learning (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing work on re-entry prediction ignores conversation thread patterns and repeated engagement of target users. |
| Approach: | They propose to use conversation thread patterns to predict whether a user will come back to a conversation they once participated in to train a model on labels that are automatically derived from the data. |
| Outcome: | The proposed task outperforms the state-of-the-art models on two social media datasets with fewer parameters and faster convergence. |
Tending Towards Stability: Convergence Challenges in Small Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Increasing the number of parameters in language models is a common strategy to enhance performance, but smaller models often underperform compared to their larger counterparts due to their reduced representational capacity. |
| Approach: | They use the Pythia model suite to analyse the training dynamics that underlie this phenomenon. |
| Outcome: | The proposed model suite enables us to examine the training dynamics of small models. |
Entropy Scheduling in Reinforcement Learning for Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | entropy in reinforcement learning functions analogously to the learning rate in LLMs. |
| Approach: | They propose an entropy scheduling system that optimizes different pre-set goals by controlling and scheduling entropicy at each step of the RL process. |
| Outcome: | The proposed method improves AIME2024 from 50.9 to 54.9 within 40 training steps. |
Multimodal and Multiresolution Speech Recognition with Transformers (2020.acl-main)
Copied to clipboard
| Challenge: | Existing audio visual automatic speech recognition systems rely on audio input to produce transcriptions. |
| Approach: | They propose an audio visual automatic speech recognition system using a transformer-based architecture and incorporate a multitask training criterion for multiresolution ASR. |
| Outcome: | The proposed system can generate character and subword transcriptions with visual information. |
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. |
Imperfect also Deserves Reward: Multi-Level and Sequential Reward Modeling for Better Dialog Management (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing research on taskoriented dialog systems mainly includes pipeline and end-to-end methods due to its non-differentiable nature. |
| Approach: | They propose a multi-level reward modeling approach that factorizes a reward into a three-level hierarchy: domain, act, and slot. |
| Outcome: | The proposed approach significantly improves performance and speed of training in a wide range of dialog systems. |
Half-S: Halving the Scale for Near-Lossless 4-Bit LLM Training (2026.findings-acl)
Copied to clipboard
Jinyang Du, Ruihao Gong, Linghan Ai, Zining Wang, Yunke Peng, Yao Wang, Lei Yan, null Wxuefei, Yaoyuan Wang, Jinyang Guo, Dahua Lin, Xianglong Liu
| Challenge: | Existing 4-bit training pipelines rely on max-scaling, which causes representation collapse . despite this, there are limitations in the accuracy of 4-bit LLM training . |
| Approach: | They propose a scaling strategy that uses half-scaling as a hardware-friendly default . they propose fp4 support that allows for a faster scaling of large language models . |
| Outcome: | The proposed scaling strategy narrows the gap between theoretical optimum and BF16 while maintaining the efficiency benefits of 4-bit training. |
CAME: Confidence-guided Adaptive Memory Efficient Optimization (2023.acl-long)
Copied to clipboard
| Challenge: | Existing memory-efficient methods require second-moment estimates of the per-parameter gradients to maintain their performance. |
| Approach: | They propose to use memory-efficient optimizers to reduce memory usage by preserving second-moment estimates of gradients. |
| Outcome: | The proposed method achieves fast convergence and lower memory usage across training tasks. |
DiffusionBERT: Improving Generative Masked Language Models with Diffusion Models (2023.acl-long)
Copied to clipboard
| Challenge: | Existing generative masked language models have a shared training objective, i.e., denoising. |
| Approach: | They propose a noise schedule for the forward diffusion process that controls the degree of noise added at each step based on the information of each token. |
| Outcome: | The proposed model improves on existing models in terms of perplexity and BLEU score. |
Knowledge Representation Learning with Contrastive Completion Coding (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing knowledge representation learning methods suffer from immaturity on tackling potentially-imperfect knowledge graphs and highly-imbalanced positive-negative instances during training. |
| Approach: | They propose a framework for knowledge representation learning that incorporates two functional components to achieve robust embedding for each entity/relation. |
| Outcome: | The proposed framework achieves better convergence against state-of-the-art methods on several benchmarks. |
TextVerifier: Robustness Verification for Textual Classifiers with Certifiable Guarantees (2023.findings-acl)
Copied to clipboard
| Challenge: | a textual classifier must withstand word-level alteration attacks due to inherent vulnerability. |
| Approach: | They propose a formal verification framework with certifiable guarantees on deep neural networks in natural language processing against word-level alteration attacks. |
| Outcome: | The proposed framework provides an approximation of the maximal safe radius with tight bounds . it yields an efficient speed edge and reliable anytime estimation . |
Neuro-Symbolic Reinforcement Learning with First-Order Logic (2021.emnlp-main)
Copied to clipboard
Daiki Kimura, Masaki Ono, Subhajit Chaudhury, Ryosuke Kohita, Akifumi Wachi, Don Joven Agravante, Michiaki Tatsubori, Asim Munawar, Alexander Gray
| Challenge: | Existing deep reinforcement learning methods require many trials before convergence and no direct interpretability of trained policies is provided. |
| Approach: | They propose a novel RL method which can learn symbolic and interpretable rules in their differentiable network. |
| Outcome: | The proposed method can learn symbolic and interpretable rules in their differentiable network. |
Efficient One-shot Compression via Low-Rank Local Feature Distillation (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing structured pruning approaches for large language models require calibration data and costly continued pretraining on billions of tokens to recover lost performance. |
| Approach: | They propose a method that locally distills activations with low-rank weights . they compress Mixtral-8x7B on a single GPU and Phi-2 3B by 40% . |
| Outcome: | The proposed method compresses Mixtral-8x7B on a single A100 GPU, removing 10 billion parameters while retaining over 95% of its original performance. |
F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching (2025.acl-long)
Copied to clipboard
| Challenge: | Recent research in Text-to-Speech (TTS) has experienced great advancement . current models can synthesize speech for any given text and mimic the speaker of audio prompt. |
| Approach: | They propose a fully non-autoregressive text-to-speech system based on flow matching with Diffusion Transformer (DiT) without complex designs such as duration model, text encoder, and phoneme alignment, the text input is simply padded with filler tokens to the same length as input speech, and then denoising is performed for speech generation. |
| Outcome: | The proposed system achieves an inference RTF of 0.15, which is greatly improved compared to state-of-the-art diffusion-based models. |
Rethinking Skip Connection with Layer Normalization (2020.coling-main)
Copied to clipboard
| Challenge: | Existing methods to solve the optimization problem of deep neural networks are not linear, but can be used as a modulating mechanism between the input and output. |
| Approach: | They propose to use skip connection to adjust the scale of the input and output to improve the performance. |
| Outcome: | The proposed approach improves performance and convergence of deep neural networks and can be applied to machine translation and image classification datasets. |
Dynamically Adjusting Transformer Batch Size by Monitoring Gradient Direction Change (2020.acl-main)
Copied to clipboard
| Challenge: | Compared to previous studies, the performance of neural models is likely to be affected by the choice of hyper-parameters. |
| Approach: | They propose to automatically and dynamically determine batch sizes by accumulating gradients of mini-batches and performing an optimization step at just the time when the direction of gradients starts to fluctuate. |
| Outcome: | The proposed approach improves the Transformer model with a fixed 25k batch size by +0.73 and +0.82 BLEU respectively. |
Reservoir Transformers (2021.acl-long)
Copied to clipboard
| Challenge: | Using random initialization, we show that some transformers obtain impressive performance even when some of the layers are frozen. |
| Approach: | They propose to freeze transformer layers and use them to improve performance . they find that the transformers obtain impressive performance even when some of the layers are randomly initialized and never updated. |
| Outcome: | The proposed model improves on translation and language modelling tasks even when some layers are frozen. |
AGGC: Adaptive Group Gradient Clipping for Stabilizing Large Language Model Training (2026.findings-acl)
Copied to clipboard
| Challenge: | Adaptive group-wise gradient clipping (AGGC) is a new approach to stabilize training of Large Language Models. |
| Approach: | They propose a method to stabilize gradient clipping by partitioning parameters into groups based on functional types and a time-dependent scheduling mechanism to balance exploration and convergence. |
| Outcome: | The proposed algorithm outperforms standard LoRA and achieves 72.93% accuracy . it can be integrated into existing pipelines with negligible overhead. |
KinyaBERT: a Morphology-aware Kinyarwanda Language Model (2022.acl-long)
Copied to clipboard
| Challenge: | Pre-trained language models such as BERT are sub-optimal at handling morphologically rich languages. |
| Approach: | They propose a two-tier BERT architecture that leverages a morphological analyzer and explicitly represents morphology in a low-resource Kinyarwanda language. |
| Outcome: | The proposed model outperforms baseline models on the low-resource morphologically rich Kinyarwanda language by 2% in F1 score and 4.3% in average score of GLUE benchmark. |
Pretraining Methods for Dialog Context Representation Learning (P19-1)
Copied to clipboard
| Challenge: | Existing methods for pretraining dialog context encoders are still in their infancy. |
| Approach: | They propose to use unsupervised pretraining objectives for dialog context representations to fine-tune and evaluate them on a set of downstream dialog tasks. |
| Outcome: | The proposed methods improve performance on a set of dialog tasks and are less data hungry. |
Active Learning for Rumor Identification on Social Media (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for rumor tracking depend on a significant amount of labeled data. |
| Approach: | They propose an Active-Transfer Learning strategy to identify rumors with limited amount of annotated data. |
| Outcome: | The proposed approach achieves faster convergence in terms of the F-score while requiring fewer annotated samples (42% of the whole dataset for the best model). |
Debate, Deliberate, Decide (D3): A Cost-Aware Adversarial Framework for Reliable and Interpretable LLM Evaluation (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing evaluation tools for Large Language Models (LLMs) are inconsistency, bias, and lack of transparent decision criteria. |
| Approach: | They propose a cost-aware, adversarial multi-agent framework that orchestrates structured debate among role-specialized agents to produce reliable and interpretable evaluations. |
| Outcome: | The proposed framework orchestrates structured debate among role-specialized agents to produce reliable and interpretable evaluations. |
Asymmetric Relational-Geometry Driven Universal Adversarial Perturbations for Vision-Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing universal adversarial perturbation (UAP) methods suffer from limited cross-model transferability in black-box scenarios. |
| Approach: | They propose an optimization-based framework that learns universal perturbations under an asymmetric relational-geometry driven objective. |
| Outcome: | The proposed framework outperforms state-of-the-art models in black-box transfer settings. |
Enhancing Language Model Hypernetworks with Restart: A Study on Optimization (2025.naacl-long)
Copied to clipboard
| Challenge: | a comprehensive investigation into optimization strategies for hypernetworks remains lacking. |
| Approach: | They propose restart optimization strategies to improve hypernetworks' performance for language models. |
| Outcome: | The proposed restart strategy improves hypernetworks' performance for language models, compared to conventional deep neural networks. |
Denoising based Sequence-to-Sequence Pre-training for Text Generation (D19-1)
Copied to clipboard
| Challenge: | PoDA pre-trains encoders and decoders by denoising noise-corrupted text . Unlike encoder-only or decode-only methods, it can be used for text generation tasks without using any task-specific techniques. |
| Approach: | They propose a sequence-to-sequence (seq2sequ) pre-training method PoDA which denoises autoencoders by denoising noise-corrupted text. |
| Outcome: | The proposed method improves model performance over strong baselines without using any task-specific techniques and significantly speed up convergence. |
Guiding Attention for Self-Supervised Learning with Transformers (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies show that self-attention patterns in trained models contain a majority of non-linguistic regularities. |
| Approach: | They propose a technique to allow efficient self-supervised learning with bi-directional Transformers by using an auxiliary loss function to guide attention heads to conform to such patterns. |
| Outcome: | The proposed method achieves state-of-the-art in low-resource settings and is agnostic to pre-training objectives. |
MUZO: Leveraging Multiple Queries and Momentum for Zeroth-Order Fine-Tuning of Large Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning large language models incur memory overhead due to the need for activation storage for back-propagation (BP). |
| Approach: | They propose a method that estimates gradients through finite differences without activation storage for back-propagation. |
| Outcome: | The proposed method demonstrates superior performance in fine-tuning various LLMs. |
Knowledge Distillation with Reptile Meta-Learning for Pretrained Language Model Compression (2022.coling-1)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) can transfer knowledge from the original model into a compact model to achieve model compression. |
| Approach: | They propose a knowledge distillation method with reptile meta-learning to facilitate the transfer of knowledge from the teacher to the student. |
| Outcome: | Extensive experiments on the GLUE benchmark show the proposed method performs better than previous methods. |
Acquiring Clean Language Models from Backdoor Poisoned Datasets by Downscaling Frequency Space (2024.acl-long)
Copied to clipboard
| Challenge: | Prior work attempts to mitigate backdoor learning during training LMs on poisoned datasets . backdoor attack poisons a small portion of training data by implanting specific text patterns . |
| Approach: | They propose a multi-scale low-rank adaptive model that prioritizes learning of clean mapping . they propose radial scalings to reduce the success rate of diverse backdoor attacks . |
| Outcome: | The proposed model outperforms baselines significantly in the frequency space . it reduces the success rate of diverse backdoor attacks to below 15% across datasets . |
E2CL: Exploration-based Error Correction Learning for Embodied Agents (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Language models are exhibiting increasing capability in knowledge utilization and reasoning, but they often suffer from misalignment between their intrinsic knowledge and environmental knowledge, leading to infeasible actions. |
| Approach: | They propose a framework that leverages exploration-induced errors and environmental feedback to enhance environment alignment for embodied agents. |
| Outcome: | The proposed framework outperforms baseline methods and exhibits superior self-correction capabilities. |
Improving Non-Autoregressive Neural Machine Translation via Modeling Localness (2022.coling-1)
Copied to clipboard
| Challenge: | Existing non-autoregressive neural machine translation models suffer from poor localization quality due to sequential dependencies within the target sentence. |
| Approach: | They propose to introduce local information into NAT models by explicitly introducing local information about surrounding words into the encoder and decoder sides to achieve localness-aware representations. |
| Outcome: | The proposed method can achieve significant improvements over strong NAT baselines. |
Curriculum Debiasing: Toward Robust Parameter-Efficient Fine-Tuning Against Dataset Biases (2025.acl-long)
Copied to clipboard
| Challenge: | Parameter-efficient fine-tuning (PEFT) addresses the memory footprint issue of full fine- tuning by modifying only a subset of model parameters. |
| Approach: | They propose a framework that debiases models in a biased-to-unbiased order and uses only a subset of parameters to modify model parameters. |
| Outcome: | The proposed framework accelerates convergence on unbiased examples by approximately twofold and improves ID and OOD performance by 1.2% and 8.0%, respectively. |
The Devil in Linear Transformer (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing linear transformers suffer from performance degradations on various tasks and corpus. |
| Approach: | They propose a new linear attention that replaces scaling with a normalization to stabilize gradients and confine attention to neighbouring tokens in early layers. |
| Outcome: | The proposed model outperforms vanilla transformers on the long-range arena benchmark while being significantly more space-time efficient. |
Symmetric Dot-Product Attention for Efficient Training of BERT Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Transformer-based models are stretched to enormous sizes, requiring increasingly larger training datasets and unsustainable amount of compute resources. |
| Approach: | They propose an alternative compatibility function for the Transformer-based attention mechanism that exploits an overlap in the learned representation of the traditional scaled dot-product attention mechanism. |
| Outcome: | The proposed model achieves 79.36 on the GLUE benchmark against 78.74 for the traditional implementation and reduces the number of trainable parameters by 6%. |
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. |
FPT: Improving Prompt Tuning Efficiency via Progressive Training (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Recent prompt tuning (PT) has gained increasing attention as a parameter-efficient way of tuning pre-trained language models (PLMs). |
| Approach: | They propose a prompt tuning algorithm that uses a small-scale partial PLM and progressively expands its depth and width until the full-model size. |
| Outcome: | The proposed method could save over 30% of training computations while achieving comparable performance. |
Multi-Attribute Controlled Text Generation with Contrastive-Generator and External-Discriminator (2022.coling-1)
Copied to clipboard
| Challenge: | Existing studies on controlled text generation focus on single-attribute control, but in practical applications, they lack controllability. |
| Approach: | They propose a framework for multi-attribute controlled text generation that can effectively generate texts with more attributes. |
| Outcome: | The proposed framework achieves remarkable controllability while keeping the text fluent and diverse. |
CorefUD 1.0: Coreference Meets Universal Dependencies (2022.lrec-1)
Copied to clipboard
| Challenge: | Recent advances in standardization for annotated language resources have led to successful large scale efforts, such as the Universal Dependencies (UD) project for multilingual syntactically annotized data. |
| Approach: | They propose a multilingual collection of corpora and a standardized format for coreference resolution compatible with morphosyntactic annotations in the UD framework. |
| Outcome: | The proposed framework is compatible with morphosyntactic annotations and includes facilities for related tasks such as named entity recognition. |
Large Language Models Can Help Mitigate Barren Plateaus in Quantum Neural Networks (2026.findings-acl)
Copied to clipboard
| Challenge: | Quantum Neural Networks (QNNs) are often hindered by barren plateaus (BPs) barren peaks are where gradient variance vanishes exponentially as qubit size increases . |
| Approach: | They propose a framework that leverages large language models with the submartingale property to iteratively synthesize initial parameters for QNNs that yield non-negligible gradient variance. |
| Outcome: | The proposed framework outperforms existing initialization methods in maintaining higher gradient variance across various QNN scales. |
Exploiting Tree Structure for Credit Assignment in Reinforcement Learning with Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement learning has shown strong promise for strengthening reasoning ability of large language models, but sparse, delayed rewards make token-level credit assignment a central challenge. |
| Approach: | They propose a critic-free algorithm that rewards tokens that change the solution. |
| Outcome: | The proposed algorithm improves on in-distribution benchmarks and out-of-disttribution settings. |
Provably Safe Offline-to-Online RL: Decoupling Learning from Data-Driven Safety Enforcement (2026.acl-long)
Copied to clipboard
| Challenge: | Hybrid offline–online reinforcement learning (O2O RL) promises both sample efficiency and robust exploration, but suffers from instability due to distribution shift between offline and online data. |
| Approach: | They propose a framework that decouples policy optimization from safety enforcement . they propose dynamic curricula that gradually extend temporal horizons and anneal offline–online data mixing . |
| Outcome: | The proposed framework preserves the exploratory value of online interactions without collapsing to conservative policies. |
Data Selection Curriculum for Abstractive Text Summarization (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Abstractive Text Summarization (ATS) models are commonly trained using large-scale data that is randomly shuffled. |
| Approach: | They propose a data selection curriculum scoring system that measures the learning difficulty of an ATS model and expected performance on an instance. |
| Outcome: | The proposed system surpasses baselines on CNN/DailyMail dataset, utilizing 20% of available instances. |
Scaling Laws Under the Microscope: Predicting Transformer Performance from Small Scale Experiments (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Neural scaling laws define a predictable relationship between a model’s parameter count and its performance after training in the form of a power law. |
| Approach: | They perform an empirical investigation of language understanding tasks and evaluate their results to determine whether scaling laws can be used to accelerate model development. |
| Outcome: | The proposed scaling laws can be exploited for debugging convergence when training large models, and can predict the performance of larger models. |
CriticSearch: Fine-Grained Credit Assignment for Search Agents via a Retrospective Critic (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing search agent pipelines rely on sparse outcome rewards, leading to inefficient exploration and unstable training. |
| Approach: | They propose a tool-integrated reasoning framework that provides turn-level feedback via a retrospective critic mechanism. |
| Outcome: | The proposed framework outperforms baselines in multi-hop reasoning benchmarks and achieves faster convergence and training stability. |
Learning Temporally-Aware Sample Weights for Preference Optimization (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for preference optimization rely on static functions of instantaneous model states and ignore temporal learning dynamics. |
| Approach: | They propose a framework that meta-learns adaptive weights using three temporal features: reward margin evolution, learning volatility, and reference deviation. |
| Outcome: | The proposed framework achieves statistically significant improvements over baselines on models ranging from 7B to 70B parameters. |
ORPO: Monolithic Preference Optimization without Reference Model (2024.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained language models with vast training corpora have shown remarkable abilities in diverse natural language processing tasks. |
| Approach: | They propose a model-free monolithic odds ratio preference optimization algorithm, ORPO, to improve preference alignment. |
| Outcome: | The proposed algorithm outperforms state-of-the-art language models with more than 7B and 13B parameters on the ultrafeedback alone. |
MeanAudio: Fast and Faithful Text-to-Audio Generation with Mean Flows (2026.acl-long)
Copied to clipboard
| Challenge: | Recent advances in Text-to-Audio Generation (TTA) systems suffer from slow inference speed, authors report . authors demonstrate that MeanAudia achieves state-of-the-art performance in single-step audio generation . |
| Approach: | They propose a text-to-audio generator capable of rendering realistic sound with only one function evaluation. |
| Outcome: | The proposed system achieves state-of-the-art performance in single-step audio generation. |
Less Noise, More Voice: Reinforcement Learning for Reasoning via Instruction Purification (2026.findings-acl)
Copied to clipboard
| Challenge: | Experimental results show that LENS outperforms GRPO in delivering higher performance and faster convergence. |
| Approach: | They propose a framework that purifies prompts by identifying and removing interference tokens and then transfers successful rollouts to supervise policy optimization on original noisy prompts. |
| Outcome: | The proposed framework outperforms GRPO in the real-world, with a 3.88% gain and speedup. |
A Probabilistic Inference Scaling Theory for LLM Self-Correction (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated the capability to refine their generated answers through self-correction, enabling continuous performance improvement over multiple rounds. |
| Approach: | They propose a probabilistic theory to model the dynamics of accuracy change and explain performance improvements observed in multi-round self-correction. |
| Outcome: | The proposed model can predict accuracy curves and improve accuracy over multiple rounds. |
BranchNorm: Robustly Scaling Extremely Deep Transformers (2024.findings-acl)
Copied to clipboard
| Challenge: | Recent work on DeepNorm scales Transformers into extremely deep (1000 layers) due to the training instability of Transformers, the depths of these SOTA models are still relatively shallow. |
| Approach: | They propose a branch-rescaled model which dynamically rescales the non-residual branch of Transformer in accordance with the training period. |
| Outcome: | The proposed approach significantly outperforms existing shallow models on multiple translation tasks and achieves better training stability and convergent performance. |
Recyclable Tuning for Continual Pre-training (2023.findings-acl)
Copied to clipboard
Yujia Qin, Cheng Qian, Xu Han, Yankai Lin, Huadong Wang, Ruobing Xie, Zhiyuan Liu, Maosong Sun, Jie Zhou
| Challenge: | Continual pre-training is the paradigm where pre-trained language models acquire fresh knowledge and gradually get upgraded. |
| Approach: | They propose to use adapted weights to recycle old PLMs for continual pre-training . they propose to combine initialization and distillation methods to achieve better performance . |
| Outcome: | The proposed method improves the convergence and performance of the upgraded PLM. |
Probing the Emergence of Cross-lingual Alignment during LLM Training (2024.findings-acl)
Copied to clipboard
| Challenge: | Multilingual Large Language Models (LLMs) achieve remarkable levels of zero-shot cross-lingual transfer performance. |
| Approach: | They propose that LLMs can align languages without explicit supervision from parallel sentences without a single linguistic feature. |
| Outcome: | The proposed model can perform zero-shot cross-lingual transfer even when the vocabularies of two languages have a null intersection, i.e., no tokens are shared. |
KBAlign: Efficient Self Adaptation on Specific Textual Knowledge Bases (2025.findings-emnlp)
Copied to clipboard
Zheni Zeng, Yuxuan Chen, Shi Yu, Ruobing Wang, Yukun Yan, Zhenghao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, Maosong Sun
| Challenge: | Existing methods for retrieval-augmented generation (RAG) are limited and fine-tuning incurs prohibitive costs of external signals. |
| Approach: | They propose a self-supervised framework that enhances RAG systems through efficient model adaptation. |
| Outcome: | The proposed framework achieves 90% of the performance gain obtained through GPT-4-supervised adaptation while relying entirely on self-annotation of much smaller models. |
AdaLomo: Low-memory Optimization with Adaptive Learning Rate (2024.findings-acl)
Copied to clipboard
| Challenge: | Large language models require substantial memory for training, thereby setting a high hardware threshold. |
| Approach: | They propose a low-memory optimization technique that reduces memory footprint . they propose an adaptive learning rate for each parameter and a grouped update normalization to stabilize convergence . |
| Outcome: | The proposed low-memory optimization performs better than the prevailing algorithm for large language models, AdamW. |
Wasserstein Selective Transfer Learning for Cross-domain Text Mining (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to improve the learning of data-scarce target domains have negative transfer due to the data distributions between source and target domain. |
| Approach: | They propose a method that uses a reinforced selector to select helpful data for transfer learning and a Wasserstein-based discriminator to maximize the distance between the selected data and target data. |
| Outcome: | The proposed method performs better on three real-world text mining tasks. |
Training with Fewer Bits: Unlocking Edge LLMs Training with Stochastic Rounding (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Quantized training improves computational and memory efficiency but introduces quantization noise. |
| Approach: | They propose to use stochastic rounding to improve LLM training but introduce quantization noise. |
| Outcome: | The proposed method can compensate for reduced accuracy during backpropagation. |
MolTRES: Improving Chemical Language Representation Learning for Molecular Property Prediction (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for chemical representation learning often lead to overfitting and limited scalability due to early convergence. |
| Approach: | They propose a framework to train Transformers on SMILES sequences to learn from structural examples and integrate external materials embedding to enrich molecular representations. |
| Outcome: | The proposed model outperforms state-of-the-art models on molecular property prediction tasks. |
Efficient Layer-wise LLM Fine-tuning for Revision Intention Prediction (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models have shown extraordinary success across text generation tasks . however, their potential for simple yet essential text classification remains underexplored . |
| Approach: | a plug-and-play layer-wise parameter-efficient fine-tuning framework is proposed . it fine- tunes a subset of important LLM layers while freezing redundant ones . |
| Outcome: | a plug-and-play framework fine-tunes a subset of important LLM layers while freezing redundant layers. |
Semantic Aware Linear Transfer by Recycling Pre-trained Language Models for Cross-lingual Transfer (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly incorporating multilingual capabilities, fueling the demand to transfer them into target language-specific models. |
| Approach: | They propose a novel cross-lingual transfer technique that recycles embeddings from target language Pre-trained Language Models to transmit deep representational strengths to LLMs. |
| Outcome: | The proposed technique outperforms existing methods in cross-lingual understanding setups and achieves faster convergence and lower loss during language adaptation. |
Seeing No Evil: Blinding Large Vision-Language Models to Safety Instructions via Adversarial Attention Hijacking (2026.acl-long)
Copied to clipboard
| Challenge: | Existing attacks optimize image perturbations to maximize harmful output likelihood, but suffer from slow convergence due to gradient conflict between adversarial objectives and the model’s safety-retrieval mechanism. |
| Approach: | They propose a push-pull approach which suppresses attention to system-prompt tokens and anchors generation on adversarial image features to avoid collisions. |
| Outcome: | The proposed approach reduces gradient conflict by 45% and achieves 94.4% attack success rate on Qwen-VL (vs. 68.8% baseline) with 40% fewer iterations. |
A Comprehensive Taxonomy of Negation for NLP and Neural Retrievers (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a new taxonomy of negation is proposed to improve neural information retrieval models . negation types are covered in existing datasets, allowing for faster convergence . |
| Approach: | They propose a taxonomy of negation that derives from philosophical, linguistic, and logical definitions . they also propose analyzing the performance of retrieval models on existing datasets using a logic-based classification mechanism. |
| Outcome: | The proposed taxonomy produces a balanced data distribution over negation types . it also provides a better training setup that leads to faster convergence on the NevIR dataset . |
Beyond End-to-End: Dynamic Chain Optimization for Private LLM Adaptation on the Edge (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are revolutionizing mobile intelligence, but their implementation on mobile devices is severely bottlenecked by the prohibitive resource demands of LLMs. |
| Approach: | They propose a paradigm that forgoes end-to-end updates in favor of a sequential, layer-by-layer manner. |
| Outcome: | Extensive experiments on multiple benchmarks demonstrate the superiority of ChainFed over existing methods, boosting average accuracy by up to 46.46%. |
INK: Injecting kNN Knowledge in Nearest Neighbor Machine Translation (2023.acl-long)
Copied to clipboard
| Challenge: | Neural machine translation models induce a non-smooth representation space, which harms its generalization results. |
| Approach: | They propose a framework to smooth the representation space by adjusting neighbor representations with a small number of new parameters. |
| Outcome: | The proposed framework outperforms the state-of-the-art kNN-MT system with average gains of 1.99 COMET and 1.0 BLEU on four benchmark datasets. |
Visualising Policy-Reward Interplay to Inform Zeroth-Order Preference Optimisation of Large Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | ZOPrO is a novel algorithm designed for *Preference Optimisation* in large language models. |
| Approach: | They propose a ZO algorithm designed for *Preference Optimisation* in LLMs that uses function evaluations instead of gradients to reduce memory usage. |
| Outcome: | The proposed method improves reward signals while achieving convergence times comparable to first-order methods. |
FSUIE: A Novel Fuzzy Span Mechanism for Universal Information Extraction (2023.acl-long)
Copied to clipboard
| Challenge: | Existing Universal Information Extraction models rely heavily on span boundaries in data during training, which does not reflect the reality of span annotation challenges. |
| Approach: | They propose a framework that uses fuzzy spans to model various IE tasks . they propose generative Universal Information Extraction (UIE) to unify various ie tasks based on fuzzy span boundaries . |
| Outcome: | The proposed framework improves on a series of main IE tasks with small amounts of data and training epochs. |
SAGE: Sign-Adaptive Gradient for Memory-Efficient LLM Optimization (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to train LLMs consume memory equivalent to twice the model size, resulting in a hybrid design that reverts to AdamW and negates the memory gains. |
| Approach: | They propose a new, memory-efficient O(d) adaptive scale that replaces AdamW in a hybrid structure that combines a Lion-style update direction with a memory-saving adaptive scale. |
| Outcome: | The proposed model outperforms existing methods on LLMs up to 1.3B parameters while significantly reducing optimizer state memory. |
Dynamic Expert Specialization: Towards Catastrophic Forgetting-Free Multi-Domain MoE Adaptation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to adapt Mixture-of-Experts models to multiple domains are prohibitive computation, cross-domain interference or require separate runs per domain. |
| Approach: | They propose a dynamic expert specialization framework for multi-domain adaptation of Mixture-of-Experts models. |
| Outcome: | The proposed framework reduces forgetting by 89% compared to full fine-tuning as domains scale from 2 to 6 and achieves faster convergence than conventional methods. |
Breaking Block Boundaries: Anchor-based History-stable Decoding for Diffusion Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Semi-autoregressive (Semi-AR) decoding suffers from inherent block constraints . naive lookahead decoding is unreliable, token stability closely correlates with convergence trend, and historical information is isolated. |
| Approach: | They propose a training-free, plug-and-play dynamic decoding strategy that monitors the stability of tokens in real time through dynamic anchors. |
| Outcome: | The proposed approach reduces decoding steps by 80% while improving performance by 3.67% on the BBH benchmark. |
Hop, skip, jump to Convergence: Dynamics of Learning Rate Transitions for Improved Training of Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Modern deep neural networks have achieved stateof-the-art performance across a wide range of machine learning tasks. |
| Approach: | They propose to switch the learning rate at a predetermined time during training to improve the performance of large language models. |
| Outcome: | The proposed model shows that switching the learning rate causes the loss curves to contract towards each other. |
Make Prompt-based Black-Box Tuning Colorful: Boosting Model Generalization from Three Orthogonal Perspectives (2024.lrec-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have shown increasing power on NLP tasks. however, tuning these models for downstream tasks usually requires exorbitant costs. |
| Approach: | They propose a black-box tuning technique that optimizes task-specific prompts without accessing gradients and hidden representations. |
| Outcome: | The proposed method improves performance under few-shot learning scenarios. |
Learning from Cognition: Enhancing RL Efficiency for LLM Reasoning via Hierarchical Metacognitive Decomposition and Refinement (2026.acl-long)
Copied to clipboard
Zexu Sun, Yongcheng Zeng, Erxue Min, Heyang Gao, Bokai Ji, Dugang Liu, Xing Tang, Xiuqiang He, Xu Chen
| Challenge: | Recent advances in Large Language Models have demonstrated notable inferential capacities via reinforcement learning (RL) however, “zero-RL” approaches relying on fixed prompt templates introduce substantial sampling inefficiencies for weak LLMs. |
| Approach: | They propose a hierarchical metacognitive RL framework that decomposes zero-accuracy problems into subproblems and prompts the policy to refine answers by referencing previous wrong solutions. |
| Outcome: | The proposed framework improves sample utilization and sample efficiency and accelerates convergence compared to baselines. |
Comparing human and language models sentence processing difficulties on complex structures (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) that converse with humans are a reality, but do LLMs experience human-like processing difficulties? |
| Approach: | They systematically compare human and LLM sentence comprehension across seven challenging linguistic structures. |
| Outcome: | The proposed model achieves near perfect accuracy on non-GP structures, but struggles on GP structures. |
Anti-Length Shift: Dynamic Outlier Truncation for Training Efficient Reasoning Models (2026.acl-long)
Copied to clipboard
Wei Wu, Liyi Chen, Congxi Xiao, Tianfu Wang, Qimeng Wang, Chengqiang Lu, Yan Gao, null Yiwu, Yao Hu, Hui Xiong
| Challenge: | Existing efficient reasoning methods rely on explicit length penalties for excessive verbosity on simple queries. |
| Approach: | They propose a training-time intervention that selectively suppresses redundant tokens . they find length shift occurs when models generate unnecessary reasoning on trivial inputs - a phenomenon that is often unexplored . |
| Outcome: | The proposed method reduces inference token usage by 78% while increasing accuracy compared to the initial policy and surpasses state-of-the-art efficient reasoning methods. |
LESA: Learnable LLM Layer Scaling-Up (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for depth scaling-up rely on empirical heuristic rules for layer duplication, resulting in poor initialization and slower convergence during continual pre-training. |
| Approach: | They propose a method for learning latent parameters between layers by concatenating parameters from each layer and applying Singular Value Decomposition. |
| Outcome: | Experiments show that LESA outperforms baseline models with less than half the cost of existing methods. |
Text-Guided Multi-Scale Frequency Representation Adaptation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning visual signals are limited by their size and complexity. |
| Approach: | They propose a multi-scale frequency-based fine-tuning method that integrates textual information and performs multi-level fine- tuning of visual signals in the frequency domain. |
| Outcome: | Extensive experiments on multimodal models, including CLIP and LLaVA, demonstrate that the proposed method significantly improves performance and efficiency with minimal cost and fast convergence within one epoch. |
BadScientist: Can a Research Agent Write Convincing but Unsound Papers that Fool LLM Reviewers? (2026.acl-long)
Copied to clipboard
| Challenge: | Existing evidence suggests that LLMs are not able to detect scientifically unsound work from malicious or poorly designed research agents. |
| Approach: | They develop a framework that evaluates whether fabrication-oriented paper generation agents can deceive multi-model LLM review systems. |
| Outcome: | The proposed framework shows that fabricated papers achieve acceptance rates up to 18% . the framework shows only marginal improvements, with detection accuracy barely exceeding random chance. |
GMFL: Efficient Global Masking for Federated LLM Fine-tuning (2026.acl-long)
Copied to clipboard
| Challenge: | Low-Rank Adaptation (LoRA) has emerged as a prominent solution to mitigate the communication and computation costs in federated fine-tuning of Large Language Models (LLMs). |
| Approach: | They propose a plug-and-play layer freezing mechanism to integrate with existing federated fine-tuning frameworks. |
| Outcome: | The proposed solution reduces communication overhead and lowers computational costs while preserving the performance of the underlying federated fine-tuning methods. |
ZERA: Zero-init Instruction Evolving Refinement Agent – From Zero Instructions to Structured Prompts via Principle-based Optimization (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to improve large language model performance focus on user prompts and require large sample sizes and long iteration cycles. |
| Approach: | They propose a framework that jointly optimizes both system and user prompts . they evaluate ZERA across five LLMs and nine diverse datasets spanning reasoning, summarization, and code generation tasks. |
| Outcome: | The proposed framework improves prompt construction over baselines and is available on github . it scores prompts using eight generalizable criteria and revises prompts based on structured critiques. |
Cat-MoD: Accelerating Multimodal Alignment via Caption Token Guided Asymmetric Mixture-of-Depths (2026.acl-long)
Copied to clipboard
| Challenge: | Existing query-based alignment modules enforce uniform cross-attention across all layers, leading to computational redundancy. |
| Approach: | They propose a framework that allows for asynchronous query-based alignment with large-scale visual features. |
| Outcome: | The proposed framework matches or surpasses baseline performance while reducing alignment FLOPs by approximately 37% during training and inference. |
Rewiring the Transformer with Depth-Wise LSTMs (2024.lrec-main)
Copied to clipboard
| Challenge: | Stacking non-linear layers allows deep neural networks to model complicated functions . but residual connections within each layer fail to fuse information from previous layers effectively . |
| Approach: | They propose a Transformer with depth-wise LSTMs connecting cascading Transformer layers and sub-layers. |
| Outcome: | The proposed model improves in English-German / French and multilingual tasks with BLEU. |
SOCIA-EVO: Automated Simulator Construction via Dual-Anchored Bi-Level Optimization (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) demonstrate strong capabilities in translating natural language into code, but applying them to this domain remains challenging. |
| Approach: | They propose a dual-anchored evolutionary framework that combines a static blueprint and a bi-level optimization to decouple structural refinement from parameter calibration. |
| Outcome: | The proposed framework identifies two failure modes in long-horizon LLM agents: contextual drift and optimization instability arising from conflating structural and parametric errors. |
HELENE: Hessian Layer-wise Clipping and Gradient Annealing for Accelerating Fine-tuning LLM with Zeroth-order Optimization (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) face memory challenges due to the high cost of backpropagation. |
| Approach: | They propose a zeroth-order (ZO) optimization that matches memory usage to inference . they propose scalable and memory-efficient zeroth order (ZE) optimizer that integrates annealed A-GNB gradients with diagonal Hessian estimation and layer-wise clipping as a second-order pre-conditioner. |
| Outcome: | The proposed algorithm outperforms state-of-the-art methods with an average speedup of 20 over MeZO on RoBERTa-large and OPT-1.3B. |
SDAR-VL: Stable and Efficient Block-wise Diffusion for Vision-Language Understanding (2026.acl-long)
Copied to clipboard
| Challenge: | Existing block-wise discrete diffusion models lack robust autoregressive (AR) decoders. |
| Approach: | They propose a block-wise discrete diffusion framework for large-scale vision-language understanding with a progressive beta noise curriculum. |
| Outcome: | The proposed framework improves training efficiency, convergence stability, and task performance over conventional block diffusion. |
MARS-RA: Rank Aggregation for Credit Assignment via Multimodal Comparisons in Embodied Multi-Agent Cooperation (2026.acl-long)
Copied to clipboard
Dawei Wang, Di Zhao, Xinyuan Liu, Marci Chi Ma, Xiaoyang Liu, Chengming Zhou, Gary Ushaw, Richard Davison
| Challenge: | Embodied AI systems are open, where agents may leave or enter mid-task due to hardware failures or task-related errors. |
| Approach: | They propose a framework that reformulates credit assignment as a rank aggregation problem using contribution-based pairwise comparisons among agents generated by large multimodal models. |
| Outcome: | The proposed framework can guide agents toward effective cooperation in complex tasks of different types. |
Why Supervised Fine-Tuning Fails to Learn: A Systematic Study of Incomplete Learning in Large Language Models (2026.acl-long)
Copied to clipboard
Chao Xue, Yao Wang, Mengqiao Liu, Di Liang, Xingsheng Han, Peiyang Liu, Xianjie Wu, Chenyao Lu, Lei Jiang, Yu Lu, Haibo Shi, Shuang Liang, Minlong Peng, Flora D. Salim
| Challenge: | Incomplete learning is widespread and heterogeneous in large language models . authors identify five recurrent sources of incomplete learning: missing prerequisite knowledge, conflicts between SFT supervision and pre-training knowledge, internal inconsistencies within SFT data, left-side forgetting during sequential fine-tuning, and insufficient optimization for rare or complex patterns. |
| Approach: | They propose a diagnostic-first framework that maps incomplete learning to causes . they identify five recurrent sources of incomplete learning: missing prerequisite knowledge, conflicts between supervision and pre-training knowledge, internal inconsistencies, left-side forgetting during sequential fine-tuning, and insufficient optimization for rare or complex patterns. |
| Outcome: | The proposed framework maps incomplete learning to causes using observable training and inference signals. |
SEE: Strategic Exploration and Exploitation for Cohesive In-Context Prompt Optimization (2025.acl-long)
Copied to clipboard
Wendi Cui, Jiaxin Zhang, Zhuohang Li, Hao Sun, Damien Lopez, Kamalika Das, Bradley A. Malin, Sricharan Kumar
| Challenge: | Existing approaches separate the optimization of prompt instructions and in-context learning examples, leading to incohesive, suboptimal results. |
| Approach: | They propose a framework that refines both prompt instructions and in-context learning examples. |
| Outcome: | The proposed framework outperforms state-of-the-art prompt optimization methods on 35 benchmark tasks. |
SAFO: Stable Adaptive Fairness Optimization for LLM-Based Social Survey Simulation (2026.acl-long)
Copied to clipboard
| Challenge: | Social survey simulations are increasingly used to improve minority performance and social-welfare metrics. |
| Approach: | They propose a dynamic utility–fairness optimization framework for LLM-based survey simulation that explicitly targets fairness and training stability. |
| Outcome: | The proposed framework improves minority performance and social-welfare metrics on three large-scale survey datasets from China, the U.S. and Europe. |
AmpleHate: Amplifying the Attention for Versatile Implicit Hate Detection (2025.emnlp-main)
Copied to clipboard
| Challenge: | Current approaches to detect hate speech rely on contrastive learning to distinguish hate from non-hate sentences. |
| Approach: | They propose a novel approach to detect implicit hate speech by identifying explicit targets . they use a pretrained Named Entity Recognition model to capture explicit target information . |
| Outcome: | The proposed approach outperforms current methods and achieves faster convergence. |
Can LLMs Really Judge? A Progressive Argumentation-Mining Framework for Distinguishing Understanding from Aggregation (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing evaluations of large language models rely on dataset-based generation accuracy . however, generative correctness does not guarantee discriminative capability to verify solutions . |
| Approach: | They propose a diagnostic framework that explicitly controls context and isolates discriminative behaviors. |
| Outcome: | The proposed framework explicitly controls context and isolates discriminative behaviors. |
Enhancing RLHF with Human Gaze Modeling (2025.emnlp-main)
Copied to clipboard
| Challenge: | Reinforcement Learning from Human Feedback (RLHF) is a powerful paradigm for aligning language models with human values and preferences. |
| Approach: | They propose to use gaze-aware reward models and gaze-based distribution of sparse rewards to enhance RLHF. |
| Outcome: | The proposed models achieve faster convergence while maintaining or slightly improving performance, reducing computational requirements during policy training. |
Astra: Activation-Space Tail-Eigenvector Low-Rank Adaptation of Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning pre-trained models are limited due to suboptimal activation subspaces. |
| Approach: | They propose a method that leverages tail eigenvectors of model output activations to construct low-rank adapters. |
| Outcome: | The proposed method outperforms existing methods across 16 benchmarks and surpasses full fine-tuning in certain scenarios. |
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. |
Convergence and Divergence of Language Models under Different Random Seeds (2025.emnlp-main)
Copied to clipboard
| Challenge: | a large body of work has examined the training dynamics of language models. |
| Approach: | They investigate the convergence of language models (LMs) trained under different random seeds . they find that larger models reconverge faster in later training stages, while smaller models never actually reconverge. |
| Outcome: | The proposed model size and training checkpoints influence convergence of language models under different seeds. |
Turning Failures into Value: Negative Experience Replay for RLVR via Confidence Gating and Boundary Failure Sampling (2026.acl-long)
Copied to clipboard
| Challenge: | Existing experience replay methods for RLVR ignore sample inefficiency . expensive reasoning trajectories are discarded immediately after a single gradient update . |
| Approach: | They propose a method to replay failure trajectories to improve model refinement . they propose 'nexGRPO' which employs mid-confidence gating to filter invalid noise and saturated errors. |
| Outcome: | The proposed model outperforms strong baaselines and achieves improved out-of-distribution generalization. |
Biomed-Enriched: Data-Efficient Biomedical Pretraining via Paragraph-Level Annotation (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models have demonstrated remarkable capabilities across a wide range of general tasks, from question answering to code generation. |
| Approach: | They use a paragraph-level pipeline to annotate PubMed Central paragraphs . they use XLM-RoBERTa to fine-tune the pipeline and propagate annotations to the full corpus . |
| Outcome: | The proposed approach improves performance on 11 tasks while using 2.5x fewer tokens and only public data. |
Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models (2026.findings-acl)
Copied to clipboard
Hoang Phan, Xianjun Yang, Yuanshun Yao, Jingyu Zhang, Shengjie Bi, Xiaocheng Tang, Madian Khabsa, Lijuan Liu, Deren Lei
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning . however, the recipe introduces a significant risk of capability regression, where models forget foundational skills after prolonged training without employing regularization strategies. |
| Approach: | They propose a replay strategy with dynamic objective reweighting for general knowledge preservation using short-horizon signals of convergence and instability. |
| Outcome: | The proposed method preserves general capabilities and improves reasoning . it can be applied to existing RLVR pipelines without training additional models or tuning . |
AG-GRPO: Answer-Guided GRPO for Masked Diffusion Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Recent work on large language models (LLMs) has emphasized not only final-answer accuracy but also reliability of reasoning on challenging tasks. |
| Approach: | They propose an answer-guided group-relative policy optimization for masked diffusion language models which generates text through iterative mangled token restoration. |
| Outcome: | The proposed approach improves over pretrained dLLMs and prior RL methods across mathematics, puzzle-solving, and code-generation benchmarks. |
AMQ: Enabling AutoML for Mixed-precision Weight-Only Quantization of Large Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Weight-only quantization is a powerful optimization technique for large language models . pushing below 4 bits often leads to substantial accuracy degradation due to increased quantization error. |
| Approach: | They propose a framework that assigns layer-wise quantization bit-widths to optimize model quality and memory usage. |
| Outcome: | The proposed framework can optimize for large language models under memory constraints. |
MDP-GRPO: Stabilized Group Relative Policy Optimization for Multi-Constraint Instruction Following (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) can follow many natural-language instructions, yet they remain brittle when a request bundles multiple explicit constraints, such as asking the LLM to respond in a particular structure with an exact ending phrase. |
| Approach: | They propose a method which stabilizes learning through multi-temperature sampling to increase reward dispersion, dual-anchor advantages to restore gradients in homogeneous groups, prospect-theoretic shaping to bound updates and penalize violations based on Kahneman Tversky’s theory and asymmetric KL regularization. |
| Outcome: | The proposed method outperforms standard GRPO on FollowBench, IFEval, and a curated multi-constraint dataset, improving strict constraint satisfaction by up to 5.0% on Llama-3.2-3B. |
Entropy-Aware Reshaping of Reinforcement Signals for Multi-Answer Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) is a standard post-training paradigm for large language models. |
| Approach: | They propose a framework that reshapes how learning signals are normalized and aggregated. |
| Outcome: | Experiments on MCTACO and MMLU-Multi show that the proposed framework improves accuracy, training stability and cross-dataset transfer performance. |
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. |