Papers with compact
Smarter, not Bigger: Fine-Tuned RAG-Enhanced LLMs for Automotive Hardware-in-the-Loop Testing (2026.acl-industry)
Copied to clipboard
| Challenge: | Hardware-in-the-Loop (HIL) testing is essential for automotive validation but suffers from fragmented and underutilized test artifacts. |
| Approach: | They propose to integrate semantic retrieval with domain-adapted large language models to support test engineers in real-world HIL workflows. |
| Outcome: | The proposed system improves perceived helpfulness, truthfulness, and satisfaction over general-purpose LLMs. |
Language Directions in Multilingual LLMs: A Layer-wise Diagnostic Study of Token Alignment and Pretraining Imprint (2026.acl-srw)
Copied to clipboard
| Challenge: | Using a unified probing framework, we analyze six multilingual LLMs across five languages. |
| Approach: | They analyze multilingual representations across five languages and analyze their behavior . they find that accuracy rises by +73.5 to +80.7 points from L0 to L1 on average . |
| Outcome: | The proposed framework enables a consistent and substantial early jump in accuracy across models . the token–language alignment measures where vocabulary sharing peaks . |
ELEVANT: A Fully Automatic Fine-Grained Entity Linking Evaluation and Analysis Tool (2022.emnlp-demos)
Copied to clipboard
| Challenge: | a tool for fine-grained evaluation of entity linkers on benchmarks is presented . a typical evaluation shows that overall precision and recall are poor . particular benchmarks often require very particular skills from an entity linker . |
| Approach: | They propose a tool for fine-grained evaluation of entity linkers on benchmarks . they use a graph-based tool to analyze performance of a set of entity links . |
| Outcome: | The proposed tool provides an automatic breakdown of the performance by error categories and by entity type. |
Improving Knowledge Graph Embedding Using Simple Constraints (P18-1)
Copied to clipboard
| Challenge: | Recent efforts focused on designing more complicated models or incorporating extra information beyond triples. |
| Approach: | They propose to use non-negativity constraints on entity representations and approximate entailment constraints on relation representations to improve KG embedding. |
| Outcome: | The proposed model outperforms baseline models on WordNet, Freebase, and DBpedia. |
MULSUM: A Multimodal Summarization System with Vis-Aligner and Diversity-Aware Image Selection (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing systems that condense text and images into concise, faithful digests are inefficient and require large fusion transformers. |
| Approach: | They propose a framework that uses image embeddings to generate a visually informed text summary and a Diversity-Aware Image Selector to maximize images-relevance to the summary. |
| Outcome: | The proposed framework outperforms baselines on automatic metrics such as ROUGE and human evaluation shows that selected images act as explanatory evidence rather than ornamental add-ons. |
Co-DETECT: Collaborative Discovery of Edge Cases in Text Classification (2025.emnlp-demos)
Copied to clipboard
Chenfei Xiong, Jingwei Ni, Yu Fan, Vilém Zouhar, Donya Rooein, Lorena Calvo-Bartolomé, Alexander Miserlis Hoyle, Zhijing Jin, Mrinmaya Sachan, Markus Leippold, Dirk Hovy, Mennatallah El-Assady, Elliott Ash
| Challenge: | Social scientists often need to develop codebooks that can be reliable but require significant human effort. |
| Approach: | They propose a mixed-initiative annotation framework that integrates human expertise with automatic annotation guided by large language models. |
| Outcome: | The proposed framework integrates human expertise with automatic annotation guided by large language models. |
Compressing Sentence Representation for Semantic Retrieval via Homomorphic Projective Distillation (2022.findings-acl)
Copied to clipboard
| Challenge: | Existing pre-trained language models produce large sentence embeddings, resulting in performance gap between large and small models. |
| Approach: | They propose a method that augments a small Transformer encoder model with learnable projection layers to produce compact sentences while mimicking a large pre-trained language model to retain the sentence representation quality. |
| Outcome: | The proposed method achieves 2.7-4.5 points performance gain on STS and SR tasks while maintaining the quality of the pre-trained language models. |
Contextualized Query Embeddings for Conversational Search (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to conversational search use multiple inference pipelines that require long inference times . despite their effectiveness, such a pipeline often includes multiple neural models that require longer inference time. |
| Approach: | They propose to integrate conversational query reformulation directly into a dense retrieval model . they use a dataset with pseudo-relevance labels to overcome the lack of training data . |
| Outcome: | The proposed model rewrites conversational queries as dense representations in conversational search and open-domain question answering datasets. |
Teacher-Student Training for Debiasing: General Permutation Debiasing for Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated impressive zero-shot capabilities and versatility in NLP tasks, however they often fail to maintain crucial invariances for specific tasks. |
| Approach: | They propose to distill the capabilities of a computationally intensive, debiased, teacher model into a more compact student model. |
| Outcome: | The proposed model outperforms larger, biased teacher models by reducing the invariances of the student model by incorporating encoders. |
Provenance: A Light-weight Fact-checker for Retrieval Augmented LLM Generation Output (2024.emnlp-industry)
Copied to clipboard
| Challenge: | Existing methods for fact checking RAG outputs rely on large language models. |
| Approach: | They propose a method that computes a factuality score that can be thresholded to yield a binary decision to check RAG outputs. |
| Outcome: | The proposed method is low latency and low cost at run-time and no need for LLM fine-tuning. |
Text2Mem: A Unified Memory Operation Language for Memory Operating System (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing memory frameworks lack a formal, executable specification for memory control. |
| Approach: | They propose a unified memory operation language that standardizes translation of natural-language instructions into reliable execution. |
| Outcome: | The proposed language standardizes translation of natural-language instructions into reliable execution. |
A Compact and Language-Sensitive Multilingual Translation Method (P19-1)
Copied to clipboard
| Challenge: | Existing paradigms for multilingual neural machine translation do not make full use of language commonality and parameter sharing. |
| Approach: | They propose a multilingual neural machine translation paradigm with one encoder-decoder model that makes full use of language commonality and parameter sharing. |
| Outcome: | The proposed method outperforms strong standard multilingual translation systems on WMT and IWSLT datasets. |
AWARE: Agentic Knowledge Warehousing for Contextual Intelligence (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models excel in information seeking tasks, but their knowledge is limited in coverage and timeliness. |
| Approach: | They propose an agentic knowledge warehousing framework that transforms unstructured data into minimal, task-conditioned knowledge representations consumable by LLMs. |
| Outcome: | Experiments on GAIA, WebWalker, and BrowseComp-Plus show improvements over baselines. |
MemoBrain: Executive Memory as an Agentic Brain for Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are inherently long-horizon, causing reasoning traces and tool artifacts to accumulate and strain the working context of large language models. |
| Approach: | They propose a model that constructs a dependency-aware memory over reasoning steps and captures salient intermediate states and their logical relations. |
| Outcome: | The proposed model prunes invalid steps, folds completed sub-trajectories, and preserves a compact, high-salience reasoning backbone under a fixed context budget. |
Learning from Diverse Reasoning Paths with Routing and Collaboration (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies suggest that the reasoning abilities of large language models (LLMs) grows with model size and pre-training data. |
| Approach: | They propose to combine quality filtering, conditional routing, and cooperative peer teaching to transfer knowledge from powerful teacher models to compact and transparent students. |
| Outcome: | Experiments show that QR-Distill is superior to traditional methods. |
TriSum: Learning Summarization Ability from Large Language Models with Structured Rationale (2024.naacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have advanced tasks like text summarization, but their size and computational demands limit their use in resource-constrained and privacy-centric settings. |
| Approach: | They propose a framework for distilling LLMs’ text summarization abilities into a compact, local model using a curriculum learning strategy that evolves from simple to complex tasks. |
| Outcome: | The proposed framework outperforms baseline models on CNN/DailyMail, XSum, and ClinicalTrial, and improves interpretability by providing insights into the summarization rationale. |
Compare, Compress and Propagate: Enhancing Neural Architectures with Alignment Factorization for Natural Language Inference (D18-1)
Copied to clipboard
| Challenge: | Using a new architecture, alignment pairs are compared, compressed and then propagated to upper layers for enhanced representation learning. |
| Approach: | They propose a new architecture where alignment pairs are compared, compressed and then propagated to upper layers for enhanced representation learning. |
| Outcome: | The proposed architecture achieves competitive performance on three popular benchmarks, SNLI, MultiNLI and SciTail, while maintaining lightweight parameter size. |
Lion: Adversarial Distillation of Proprietary Large Language Models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for transferring knowledge from a proprietary large language model to an open-source LLM have been criticized for not incorporating any feedback to boost student model's proficiency iteratively. |
| Approach: | They propose a novel adversarial distillation framework that integrates feedback to identify challenging instructions where the student model’s performance falls short. |
| Outcome: | The proposed framework surpasses existing state-of-the-art LLMs by 55.4% in challenging zero-shot reasoning benchmarks and 16.7% on AGIEval. |
Defending Pre-trained Language Models from Adversarial Word Substitution Without Performance Sacrifice (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing defense approaches suffer from notable performance loss and complexities. |
| Approach: | They propose a framework for detecting anomalies with frequency-aware randomization to defend adversarial word substitution. |
| Outcome: | The proposed framework outperforms existing defense methods over various tasks with much higher inference speed. |
CacheNotes: Task-Aware Key-Value Cache Compression for Reasoning-Intensive Knowledge Tasks (2026.eacl-long)
Copied to clipboard
| Challenge: | Current methods for integrating external knowledge into Large Language Models (LLMs) face limitations with broad, multi-source queries, while long-context models are computationally prohibitive. |
| Approach: | They propose a task-aware key-value cache compression method that generates a sequence of CPTs from a corpus and guides a one-time compression of the corpus into a compact, reusable KV cache. |
| Outcome: | The proposed method outperforms Retrieval-Augmented Generation (RAG) on Question-Answering tasks and reduces latency by over 4. |
Evian: Towards Explainable Visual Instruction-tuning Data Auditing (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing data filtering methods rely on coarse-grained scores that lack granularity to identify nuanced semantic flaws. |
| Approach: | They propose a "Decomposition-then-Evaluation" paradigm that breaks model responses into constituent cognitive components. |
| Outcome: | The proposed model outperforms models trained on larger datasets in three key areas . the authors show that Logical Coherence is the most critical factor in data quality evaluation . |
VizoMem: A Visual-Textual Memory Framework for Efficient Long-Horizon Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing systems that use long-context modeling incur computational and memory overhead. |
| Approach: | They propose a visual memory framework that pre-rendered text into structured images and stored as visual notes for agentic systems. |
| Outcome: | The proposed system reduces token consumption while preserving effective long-term memory recall. |
Boosting the Capabilities of Compact Models in Low-Data Contexts with Large Language Models and Retrieval-Augmented Generation (2025.coling-main)
Copied to clipboard
| Challenge: | Existing language models lack data and computation power, but they are extremely parameter-heavy and difficult to train. |
| Approach: | They propose a retrieval augmented generation framework backed by a large language model to correct the output of a smaller model for morphological glossing. |
| Outcome: | The proposed model is highly effective in data-scarce settings and offers a state-of-the-art for morphological glossing. |
Generative Gamer: Learning Equilibrium Strategy by LLM-driven Dynamic Deduction (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) falter in domains requiring deep strategic reasoning. |
| Approach: | They propose a framework that trains LLMs to reason like an expert player . they propose action pruning based on policy confidence, state pruning via value estimation and branch pruning inspired by alpha-beta principles to train the model effectively. |
| Outcome: | Experiments on Tic-Tac-Toe and Leduc Poker show that GenGamer significantly improves the strategic capabilities of large language models. |
Phonotomizer: A Compact, Unsupervised, Online Training Approach to Real-Time, Multilingual Phonetic Segmentation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to phonetic segmentation are hierarchical and end-to-end . many mistakes in final output stem from subtle segmenter perturbations . |
| Approach: | They propose a phonetic segmentation system that trains on raw sound files alone . it can modulate computational exactness and reduce acoustic model size, they argue . |
| Outcome: | The proposed method reduces the size of the acoustic model and training epochs. |
Revisiting Pruning vs Quantization for Small Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Compressing Small Language Models (SLMs) is particularly suited for resource-constrained devices, but their compression dynamics remain underexplored compared to Large Language Model (LLMs). |
| Approach: | They evaluated post-training pruning and quantization methods across six SLMs from 0.5 to 3.8B, seven languages, and seven downstream tasks. |
| Outcome: | The proposed methods outperform pruning and quantization on six SLMs from 0.5 to 3.8B, seven languages, and seven downstream tasks. |
Model Internal Sleuthing: Finding Lexical Identity and Inflectional Features in Modern Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Prior work suggests hierarchical organization where different layers specialize in capturing distinct levels of linguistic structure. |
| Approach: | They probe 25 models from BERT Base to Qwen2.5-7B focusing on linguistic properties: lexical identity and inflectional features. |
| Outcome: | The proposed model maintains inflectional features across layers while trading off lexical identity for compact, predictive representations. |
Logits-Based Finetuning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for developing compact and efficient large language models lack token-level dependencies and linguistic diversity. |
| Approach: | They propose a logits-based fine-tuning framework that integrates supervised learning and knowledge distillation to build enriched training targets using teacher logits and ground truth labels. |
| Outcome: | The proposed method outperforms existing methods on a large-scale logits dataset and a series of science-focused models. |
BcQLM: Efficient Vision-Language Understanding with Distilled Q-Gated Cross-Modal Fusion (2025.findings-emnlp)
Copied to clipboard
| Challenge: | proposed lightweight MLLM framework for end-to-end visual question answering . proposed framework uses BreezeCLIP, a vision-language encoder optimised for efficient multimodal understanding . |
| Approach: | proposed lightweight MLLM framework is based on BreezeCLIP, a vision-language encoder . it offers a promising path toward deployable ML models under practical hardware constraints. |
| Outcome: | The proposed model significantly reduces computational cost while achieving performance comparable to standard-size MLLMs. |
Learning to Extract Rational Evidence via Reinforcement Learning for Retrieval-Augmented Generation (2026.findings-acl)
Copied to clipboard
| Challenge: | Retrieval-Augmented Generation (RAG) is effective in Large Language Models (LLMs). However, retrieval noises undermine the quality of LLMs’ generation, necessitating the development of denoising mechanisms. |
| Approach: | They propose a model which integrates reasoning and extracting into one unified trajectory, followed by knowledge token masking to avoid information leakage. |
| Outcome: | Extensive experiments on five benchmark datasets show the superiority of EviOmni, which provides compact and high-quality evidence, enhances the accuracy of downstream tasks, and supports both traditional and agentic RAG systems. |
Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing frameworks treat memory as a static append-only archive . Existing systems focus on passive accumulation, resulting in a 'passive accumulation' of memory. |
| Approach: | They propose a framework for experience-driven agent evolution that integrates procedural memory with contextual information to create a high-quality experience pool. |
| Outcome: | Experiments on BFCL-V3 and AppWorld show that ReMe outperforms memoryless Qwen3-8B. |
U-Fold: Dynamic Intent-Aware Context Folding for User-Centric Agents (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing context-folding methods are designed for single-query or single-intent scenarios. |
| Approach: | They propose a dynamic context-folding framework tailored to user-centric tasks that preserves fine-grained information through dynamic context folding. |
| Outcome: | The proposed framework outperforms ReAct and previous folding frameworks on long, noisy tasks. |
Targeted Distillation for Sentiment Analysis (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies demonstrate that large language models exhibit remarkable capabilities and achieve state-of-the-art performance in diverse sentiment analysis tasks. |
| Approach: | They propose a distillation framework that decouples knowledge from alignment and introduces a sentiment analysis benchmark that covers a diverse set of tasks. |
| Outcome: | The proposed framework improves models' generalization to unseen tasks and their generalization is strong against existing small-scale models. |
Multilingual and Cross-Lingual Citation Needed Detection on Wikipedia for Lower-Resource Languages (2026.acl-long)
Copied to clipboard
| Challenge: | Existing research has largely overlooked lower-resource languages for automated fact-checking. |
| Approach: | They propose a multilingual CND corpus spanning 18 languages across three resource levels and a small decoder-based language model for CND. |
| Outcome: | The proposed model outperforms prompted LLMs in cross-lingual CND across languages. |
Less is More: The Effectiveness of Compact Typological Language Representations (2025.emnlp-main)
Copied to clipboard
| Challenge: | Linguistic feature datasets such as URIEL+ have high dimensionality and sparsity, especially for low-resource languages. |
| Approach: | They propose a pipeline to optimize the URIEL+ typological feature space by feature selection and imputation. |
| Outcome: | The proposed pipeline produces compact yet interpretable typological representations on linguistic distance alignment and downstream tasks. |
Breaking Token Into Concepts: Exploring Extreme Compression in Token Representation Via Compositional Shared Semantics (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Standard language models employ unique, monolithic embeddings for each token, limiting their ability to capture multifaceted meanings. |
| Approach: | They propose a compositional structure that accumulates diverse semantic facets for tokens . they apply this representational scheme to standard transformer architectures and a biomedical domain benchmark . |
| Outcome: | The proposed representational scheme achieves extreme compression in embedding parameters while maintaining >95% task performance relative to the base model. |
EviReport: From Reasoned Outlines to Evidence Tracked Long-Form Reports (2026.findings-acl)
Copied to clipboard
Zihan Liu, Jianhui li, Zexin Wang, Fei Sun, Jingjing LI, Zheyuan Li, Ke Xiang, Hang Cui, Houhua Gong, Changhua Pei, Gaogang Xie
| Challenge: | Evidence-intensive reports often produce fluent but under-supported drafts . eviReport is an evidence-grounded workflow for automated long-form report generation . |
| Approach: | They propose an evidence-tracked workflow that organizes corpus evidence into compact, traceable units and retrieves query-relevant subgraphs into retrieval-ready packages. |
| Outcome: | The proposed workflow outperforms baselines in factual coverage, factual accuracy and visual evidence integration. |
LLM-induced Rationales for More Compact Explainable Style Classification Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for extracting explanations from complex models are based on discovering a large number of features, and this affects interpretability. |
| Approach: | They propose a model that leverages Large Language Models and clustering algorithms to discover a compact set of interpretable features. |
| Outcome: | The proposed model reduces the number of features on 3 Style Classification tasks by 85–99% while reducing the number by 85. |
Piece of Table: A Divide-and-Conquer Approach for Selecting Subtables in Table Question Answering (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to QA tables rely on independent row or column selection, fail to capture cross-row and cross-column dependencies, or attempt global reasoning. |
| Approach: | They propose a divide-and-conquer subtable selection framework that aggregates local evidence without requiring explicit global reasoning. |
| Outcome: | The proposed framework outperforms previous approaches to table QA in the noisy context. |
VoiceBBQ: Investigating Effect of Content and Acoustics in Social Bias of Spoken Language Model (2025.emnlp-main)
Copied to clipboard
| Challenge: | Due to the nature of speech modality, social bias in Spoken Language Models (SLMs) can emerge from two distinct sources: 1) content aspect and 2) acoustic aspect. |
| Approach: | They propose a dataset that measures social bias by presenting ambiguous or disambiguated contexts followed by questions that may elicit stereotypical responses. |
| Outcome: | The proposed dataset converts every BBQ context into controlled voice conditions, enabling per-axis accuracy, bias, and consistency scores comparable to the original text benchmark. |
IterCOMP: Reasoning-aware Adaptive Prompt Compression for Multi-hop Question Answering (2026.acl-long)
Copied to clipboard
| Challenge: | Existing prompt compression methods are designed for single-turn queries and fail to capture interdependent reasoning steps. |
| Approach: | They propose a unified, training-free prompt compression framework that integrates multi-hop reasoning within an iterative compression loop. |
| Outcome: | Experiments on MusiQue, 2WikiMultiHopQA, and HotpotQA show that iterCOMP achieves significant improvements in Exact Match and F1 scores while reducing the token budget. |
Lizard: An Efficient Linearization Framework for Large Language Models (2026.acl-long)
Copied to clipboard
Chien Van Nguyen, Huy Huu Nguyen, Ruiyi Zhang, Hanieh Deilamsalehy, Puneet Mathur, Viet Dac Lai, Haoliang Wang, Jayakumar Subramanian, Ryan A. Rossi, Trung Bui, Nikos Vlassis, Franck Dernoncourt, Thien Huu Nguyen
| Challenge: | Existing linearization frameworks that rely on softmax attention with quadratic time and memory complexity pose significant computational and memory bottlenecks for long-context applications. |
| Approach: | They propose a linearization framework that transforms pretrained Transformer-based Large Language Models into subquadratic architectures that closely approximate softmax attention while preserving model quality. |
| Outcome: | Experiments show that the proposed framework outperforms existing methods by 9.4 - 24.5 points on the 5-shot MMLU benchmark and demonstrating superior associative recall. |
Octopus: Gated Selective Attention for Memory-Bounded Long-Context Inference in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Subquadratic architectures rely on aggressive state compression that degrades performance on complex reasoning tasks. |
| Approach: | They propose a framework that confers fixed-memory inference onto pretrained Transformers . they use a learnable module that enforces an adaptive sparsity policy over the context history . |
| Outcome: | The proposed framework outperforms state-of-the-art linearized baselines on the GSM8K benchmark by over 36 points under identical memory constraints. |
Distilling Examples into Task Instructions: Enhanced In-Context Learning for Real-World B2B Conversations (2026.findings-acl)
Copied to clipboard
| Challenge: | In-context learning (ICL) is the standard method for low-resource classification, yet its efficacy in specialized domains remains largely unexplored. |
| Approach: | They propose a framework that distills verbose examples into compact, interpretable representations of structured classification criteria and precise task descriptions. |
| Outcome: | The proposed method achieves 99% reduction in token usage and improves macro-averaged AUC by up to 7% over traditional ICL. |
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. |
Why Does Reinforcement Learning Generalize? A Feature-Level Mechanistic Study of Post-Training in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning (RL)-based post-training often improves the reasoning performance of large language models beyond the training domain, while supervised fine-tuning (SFT) frequently leads to general capabilities forgetting. |
| Approach: | They propose a feature-level mechanistic analysis methodology to probe RL generalization using a controlled experimental setup. |
| Outcome: | The proposed method identifies a compact, task-agnostic set of features that directly mediate generalization across diverse tasks. |
Rhombus: Incentivizing Coordination in Parallel Thinking through Reinforcement Learning (2026.findings-acl)
Copied to clipboard
Ziyuan Nan, Qi Yi, Di Huang, Yutong Wu, Guanhua Huang, Xue Gong, Kejiao Li, Yuhao Jiang, Chenchen Zhang, Zenan Xu, Xing Hu, Bo Zhou
| Challenge: | Parallel thinking is a promising avenue for scaling test-time compute in Large Language Models . however, coordinating the exploration and aggregation stages remains challenging . |
| Approach: | They propose a parallel thinking framework that explicitly incentivizes coordination between components via end-to-end reinforcement learning. |
| Outcome: | The proposed framework improves accuracy by 6.0% over long chain-of-thought baselines while reducing wall-clock latency by 39.4% under matched token budgets. |
Emergence and Localisation of Semantic Role Circuits in LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Despite displaying semantic competence, large language models’ internal mechanisms that ground abstract semantic structure remain insufficiently characterised. |
| Approach: | They propose a causal-temporal methodology that uses contrastive minimal pairs, edge-attribution circuit discovery, and training-time tracking to characterise semantic-role circuits along three dimensions. |
| Outcome: | The proposed method is applicable to any behaviour isolatable through contrastive minimal pairs, including factual recall, syntactic agreement, and logical reasoning. |
POLARIS: A Gödel Agent Framework for Small Language Models through Experience-Abstracted Policy Repair (2026.findings-acl)
Copied to clipboard
| Challenge: | Gödel agent Polaris makes policy level changes with small, auditable patches that persist in the policy and are reused on unseen instances within each benchmark. |
| Approach: | They propose a Gödel agent that performs policy repair via experience abstraction . Polaris makes policy level changes with small, auditable patches that persist in the policy . |
| Outcome: | The proposed agent improves on MGSM, DROP, GPQA, and LitBench models over the base policy and competitive baselines. |
Multiplication in Multimodal LLMs: Computation with Text, Image, and Audio Inputs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks lack systematically paired instances across modalities, making it difficult to compare genuine arithmetic limits . a model that computes 4736 may fail on a nearby instance like 8967, despite a well-tuned internal router. |
| Approach: | They propose a controlled multimodal multiplication benchmark that factorially varies digit length, digit sparsity, representation, and modality with paired instances from a reproducible generator. |
| Outcome: | The proposed model can perceive numerical content across modalities but fails to perform exact multi-digit multiplication when presented as numerals, number words, images, or in audio form. |
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. |