CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextGraphChain-of-Thought
🎯 What it does: This paper proposes a graph learning framework called GRAPH-R1 that completely does not rely on graph neural networks, transforming graph tasks such as node classification, link prediction, and graph classification into text reasoning problems, and generating interpretable chain-of-thought reasoning through large reasoning models (LRM).
Yuhao Yang (University of Hong Kong), Chao Huang (University of Hong Kong)
CodeRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelAgentic AITextGraph
🎯 What it does: Propose a multi-agent framework called GraphAgent that automatically generates semantic knowledge graphs (SKG), understands user queries, and performs prediction and generation tasks;
Grounded Semantic Role Labelling from Synthetic Multimodal Data for Situated Robot Commands
Claudiu Daniel Hromei (University of Rome Tor Vergata), Roberto Basili (University of Rome Tor Vergata)
CodeGenerationData SynthesisRobotic IntelligenceLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose a multimodal semantic role labeling (G-SRL) framework that aligns robot commands with perceptual context and designs an automated synthetic image generation and validation pipeline, producing approximately 11,000 image-command pairs with structured annotations.
Group-SAE: Efficient Training of Sparse Autoencoders for Large Language Models via Layer Groups
Davide Ghilardi (University of Milan-Bicocca), Matteo Palmonari (London School of Economics)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Significantly reduces the computational overhead of SAE training by clustering adjacent layers into groups and training a single sparse autoencoder (SAE) per group to reconstruct multi-layer activations.
Grouping Entities with Shared Properties using Multi-Facet Prompting and Property Embeddings
Amit Gajbhiye, Steven Schockaert (Cardiff University)
CodeClassificationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Leverage large language models (LLM) to generate multi-faceted (facet) attribute pairs (facet-property) for each entity, then use pre-trained text embeddings (LLM2Vec) to map these attribute pairs into vectors and cluster them. Finally, group entities under the same category if they appear in the same attribute cluster, thereby achieving general-purpose grouping of large-scale entity sets.
GuessingGame: Measuring the Informativeness of Open-Ended Questions in Large Language Models
Dylan Hutson (University of Cincinnati), Tianyu Jiang (University of Cincinnati)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Designed and implemented GuessingGame, an open-source protocol for evaluating the strategic performance of large language models in open-domain, open-ended questioning scenarios.
HealthCards: Exploring Text-to-Image Generation as Visual Aids for Healthcare Knowledge Democratizing and Education
Qian Wu (Chinese University of Hong Kong), Qi Dou (Chinese University of Hong Kong)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelFlow-based ModelImageTextBiomedical Data
🎯 What it does: Explored how to use text-to-image models to generate health education flashcards, constructed a high-quality medical knowledge flashcard dataset, and fine-tuned and validated the model.
HMoE: Heterogeneous Mixture of Experts for Language Modeling
An Wang (Tencent Hunyuan), Cheng-zhong Xu (University of Macau)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose a heterogeneous hybrid expert (HMoE) framework that uses experts of different scales to enhance the performance and efficiency of language models, and introduces a parameter penalty loss during training to balance expert activation.
HookMoE: A learnable performance compensation strategy of Mixture-of-Experts for LLM inference acceleration
Cheng Longkai (Nankai University), Tao Li (Nankai University)
CodeComputational EfficiencyTransformerSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Proposed a lightweight Hook compensation module combined with the TLNLE scheme to reduce the number of activated experts in MoE models for faster inference, while restoring performance through fine-tuning.
How Far Can LLMs Improve from Experience? Measuring Test-Time Learning Ability in LLMs with Human Comparison
Jiayin Wang (Tsinghua University), Min Zhang (Tsinghua University)
CodeMeta LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigated the test-time learning capability of large language models, designed an evaluation framework based on semantic games, and compared it with human performance.
Tu Nguyen (ETH Zürich), Ryan Cotterell (ETH Zürich)
CodeExplainability and InterpretabilityText
🎯 What it does: This paper proposes a Target Persuasion Score (TPS) based on Wasserstein distance to quantify the impact of context on the answer distribution of language models.
How Sememic Components Can Benefit Link Prediction for Lexico-Semantic Knowledge Graphs?
Hansi Wang (Peking University), Yang Liu (Peking University)
CodeTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextGraphBenchmark
🎯 What it does: This paper proposes a method to enhance link prediction in lexico-semantic knowledge graphs (Lexico‑Semantic KG) by leveraging semantic units (sememes), and constructs the corresponding sememe prediction dataset SememeDef as well as Chinese link prediction benchmarks HN7 and CWN5.
HVGuard: Utilizing Multimodal Large Language Models for Hateful Video Detection
Yiheng Jing (Wuhan University), Hongxin Hu (University at Buffalo)
CodeClassificationTransformerLarge Language ModelMixture of ExpertsVision Language ModelVideoTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Designed an anti-abuse video detection framework HVGuard based on multi-modal large language models (MLLM), generating reasoning processes via chain-of-thought (CoT) and fusing multi-modal features using Mixture-of-Experts (MoE).
HYDRA: A Multi-Head Encoder-only Architecture for Hierarchical Text Classification
Fabian Karl (University of Ulm), Ansgar Scherp (University of Ulm)
CodeClassificationTransformerText
🎯 What it does: Propose HYDRA, a multi-head encoder-only architecture for hierarchical text classification, directly treating each level as an independent classification task while sharing the underlying encoder;
ICG: Improving Cover Image Generation via MLLM-based Prompting and Personalized Preference Alignment
Zhipeng Bian (Huazhong University of Science and Technology), Zhenhua Dong (Huawei Noah's Ark Lab)
CodeGenerationRecommendation SystemTransformerLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelImageTextMultimodality
🎯 What it does: Propose an end-to-end framework called ICG, which leverages multimodal large language models to extract semantic contexts from titles and reference images, generates personalized cover images by combining user embeddings, and achieves unsupervised optimization through multi-reward learning.
Identification of Multiple Logical Interpretations in Counter-Arguments
Wenzhi Wang (Tohoku University), Kentaro Inui (MBZUAI)
CodeClassificationExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a new multi-interpretation counting method, which performs fine-grained splitting and multi-annotation of logical structures in refutation arguments (CA), constructs the CALSA+ dataset, and trains models to identify these multiple logical explanations.
IG-Pruning: Input-Guided Block Pruning for Large Language Models
Kangyu Qiao (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences), Yang Feng (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences)
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes an input-feature-based block-level pruning method called IG-Pruning, which can dynamically select execution paths for Transformer layers during inference, significantly reducing computational costs for large language models.
🎯 What it does: Propose the iKnow-audio framework, combining knowledge graphs with the CLAP audio-text model to achieve zero-shot audio classification;
IL-PCSR: Legal Corpus for Prior Case and Statute Retrieval
Shounak Paul (IIT Kharagpur), Ashutosh Modi (IIT Kanpur)
CodeRetrievalGraph Neural NetworkLarge Language ModelTextBenchmark
🎯 What it does: This paper constructs the first parallel corpus in the Indian legal domain, IL-PCSR, simultaneously used for retrieving applicable legal articles and previous case law, and conducts systematic research on retrieval tasks based on this corpus.
Image Embedding Sampling Method for Diverse Captioning
Sania Waheed (University of Southampton), Na Min An (KAIST)
CodeSegmentationGenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Proposed the HBoP framework, which samples structured segmentation image embeddings from the last layer of the pre-trained visual model ViT, and generates multi-level (global, regional, fine-grained) diverse descriptions using BLIP;
Implicit Behavioral Alignment of Language Agents in High-Stakes Crowd Simulations
Yunzhe Wang (University of Southern California), Volkan Ustun (USC Institute for Creative Technologies)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AITabular
🎯 What it does: Developed the PEBA-PEvo framework, utilizing LLM-driven generative agents in high-risk scenarios (active shooter events) to iteratively optimize agent personas, enabling group behavior distributions to approach expert-defined real-world distributions.
Improving Clustering with Positive Pairs Generated from LLM-Driven Labels
Xiaotong Zhang, Ying Li (Zhejiang University)
CodeRepresentation LearningLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the PPLL framework: construct positive sample pairs using micro-cluster labels generated by LLMs, train the embedder with BYOL, perform K-means in conjunction with micro-cluster labels, and ultimately generate interpretable cluster labels.
Improving Informally Romanized Language Identification
Adrian Benton (Google Research), Brian Roark (Google Research)
CodeClassificationTransformerLarge Language ModelText
🎯 What it does: The paper enhances language recognition performance for Indian languages by generating natural spelling variants in Romanized text during training.
Improving Large Language Models Function Calling and Interpretability via Guided-Structured Templates
Hy Dang (University Of Notre Dame), Meng Jiang (University Of Notre Dame)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a reasoning framework and dataset (ToolGT) based on structured templates to enhance the accuracy and interpretability of large language models in function calls.
Improving Low-Resource Sequence Labeling with Knowledge Fusion and Contextual Label Explanations
Peichao Lai (Peking University), Bin Cui (Peking University)
CodeData SynthesisExplainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes a framework that combines the knowledge enhancement process of large language models (LLMs) with the span-based KnowFREE model to improve sequence labeling tasks under low-resource conditions. The framework generates extended entity labels, part-of-speech labels, and tokenization labels using LLMs, and enhances the model's understanding of entity boundaries and semantic distributions by generating contextualized entity explanations through interpretive prompts.
Improving Online Job Advertisement Analysis via Compositional Entity Extraction
Kai Krüger (Bundesinstitut für Berufsbildung), Alan Akbik (Humboldt-Universität zu Berlin)
CodeTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a compositional entity modeling framework for extracting job requirements from online job advertisements and construct the GOJA dataset
🎯 What it does: Investigate the bias of different multilingual models (LASER3, XLM-R, LaBSE) in parallel corpus filtering during web mining, and propose and systematically evaluate multiple rule-based heuristic filtering combinations to reduce cross-model discrepancies and improve NMT performance.
Improving Zero-shot Sentence Decontextualisation with Content Selection and Planning
Zhenyun Deng (University of Cambridge), Andreas Vlachos (University of Cambridge)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a zero-shot sentence decontextualization framework ECSP, which first identifies ambiguous EDUs through EDU segmentation and selects relevant discourse content, followed by content planning to rewrite sentences
In-Context Learning Boosts Speech Recognition via Human-like Adaptation to Speakers and Language Varieties
Nathan Roll (Stanford University), Dan Jurafsky (Stanford University)
CodeRecognitionDomain AdaptationTransformerLarge Language ModelPrompt EngineeringTextMultimodalityAudio
🎯 What it does: This paper proposes and implements a scalable context learning framework that leverages the Phi-4-MM model to rapidly adapt to different speakers and language variants in speech recognition through alternating audio-text examples during inference;
IndiGEC: Multilingual Grammar Error Correction for Low-Resource Indian Languages
Ujjwal Sharma (Indian Institute of Technology Bombay), Pushpak Bhattacharyya (Indian Institute of Technology Bombay)
CodeGenerationData SynthesisTransformerText
🎯 What it does: Propose the Mask-Translate&Fill (MTF) method, which generates high-quality synthetic grammar error correction data by leveraging monolingual data, machine translation, and masked language models; simultaneously release the multilingual IndiGEC corpus
Infini-gram mini: Exact n-gram Search at the Internet Scale with FM-Index
Hao Xu (University of Washington), Hannaneh Hajishirzi (University of Washington)
CodeRetrievalComputational EfficiencyText
🎯 What it does: Built a system named INFINI-GRAM MINI, which achieves efficient exact string search and counting on internet-scale (multi-petabyte) text corpora using FM-Index.
Information Integration in Large Language Models is Gated by Linguistic Structural Markers
Wei Liu (Zhejiang University), Nai Ding (Zhejiang University)
CodeTransformerLarge Language ModelText
🎯 What it does: This paper proposes a window-based prediction method to quantify the information integration window of large language models (LLMs) and humans at sentence and clause boundaries, measuring the difference between local window predictions and full context predictions using Jensen-Shannon divergence.
Interpretability Analysis of Arithmetic In-Context Learning in Large Language Models
Gregory Polyakov (University of Tübingen), Seyed Ali Bahrainian (University of Tübingen)
CodeExplainability and InterpretabilityTransformerText
🎯 What it does: This paper systematically studies the in-context learning (ICE) mechanisms of large language models (LLMs) in three arithmetic tasks through multiple mechanism explanation techniques.
Investigating How Pre-training Data Leakage Affects Models’ Reproduction and Detection Capabilities
Masahiro Kaneko (Mohamed bin Zayed University of Artificial Intelligence), Timothy Baldwin (Mohamed bin Zayed University of Artificial Intelligence)
CodeSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Systematically evaluate the pre-training data leakage rate of large language models (LLMs), explore how the proportion of leaked instances affects model reproduction rate and leakage detection rate, and propose a supervised detection method based on a few examples to mitigate the impact of low leakage rates.
Iterative Prompt Refinement for Safer Text-to-Image Generation
Jinwoo Jeon (Korea University), Byung-Jun Lee (Korea University)
CodeGenerationSafty and PrivacySupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Propose an iterative prompt refinement algorithm (IPR) that leverages vision-language models to repeatedly adjust prompts based on generated images, thereby enhancing the safety of text-to-image models.
🎯 What it does: Proposes iTool, which combines easy-to-difficult SFT, MCTS path exploration, and iterative ReFT to enhance LLM tool usage capabilities, especially in complex scenarios.
iVISPAR — An Interactive Visual-Spatial Reasoning Benchmark for VLMs
Julius Mayer (Osnabrück University), Elia Bruni (Osnabrück University)
CodePrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the iVISPAR benchmark, utilizing interactive sliding geometric puzzles (SGP) to evaluate the spatial reasoning and planning capabilities of vision-language models (VLMs) under three input modalities: 3D vision, 2D vision, and text.
Jailbreak LLMs through Internal Stance Manipulation
Shuangjie Fu (Chinese Academy of Sciences), Xueqi Cheng (Chinese Academy of Sciences)
CodeAdversarial AttackLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes a Stance Manipulation (SM) method that achieves automated jailbreaking by suppressing the internal rejection posture of LLMs, further elucidating the internal formation process of LLM rejection mechanisms;
🎯 What it does: Investigated and demonstrated that attacking closed-source language models through 'jailbreak-tuning' can completely undermine their safety defenses and generate high-quality harmful responses.
Job Unfair: An Investigation of Gender and Occupational Bias in Free-Form Text Completions by LLMs
Camilla Casula (Fondazione Bruno Kessler), Sara Tonelli (Fondazione Bruno Kessler)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Examined gender and occupational bias in 9 English-Italian bilingual large language models (LLMs) during free-text completion tasks, combining manual annotation, vector embedding clustering, and lexical relevance analysis to systematically evaluate gender distribution, thematic direction, subject misunderstanding, and agent/affinity tendencies in model outputs.
JOLT-SQL: Joint Loss Tuning of Text-to-SQL with Confusion-aware Noisy Schema Sampling
Jinwang Song (Zhengzhou University), Min Peng (Wuhan University)
CodeOptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose a single-stage joint loss optimization framework called JOLT-SQL to simultaneously optimize schema linking and SQL generation, thereby improving Text-to-SQL performance;
JUREX-4E: Juridical Expert-Annotated Four-Element Knowledge Base for Legal Reasoning
Huanghai Liu (Tsinghua University), Yansong Feng (Peking University)
CodeClassificationRetrievalTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the JUREX-4E expert-annotated four-element knowledge base, covering 155 common criminal charges in Chinese criminal cases. A hierarchical legal interpretation framework was used to provide precise and complete annotations for each element, and its value was validated in tasks such as distinguishing similar charges and case retrieval.
KCS: Diversify Multi-hop Question Generation with Knowledge Composition Sampling
Yangfan Wang (Harbin Institute of Technology), Jingchi Jiang (MemTensor Technology Co., Ltd.)
CodeGenerationData SynthesisTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose the Knowledge Composition Sampling (KCS) framework, which, given an answer and a long-text context, first selects knowledge compositions using a sentence-level sequence prediction model, then samples diverse knowledge compositions through random decoding, and finally generates multi-hop questions using a pre-trained multi-hop question generation model.
Keep Security! Benchmarking Security Policy Preservation in Large Language Model Contexts Against Indirect Attacks in Question Answering
Hwan Chang (Chung-Ang University), Hwanhee Lee (Chung-Ang University)
CodeSafty and PrivacyTransformerPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a new large-scale benchmark dataset called CoPriva to evaluate whether large language models can adhere to safety policies defined in the context when facing direct and indirect attacks, primarily tested in question-answering tasks.
KG-CQR: Leveraging Structured Relation Representations in Knowledge Graphs for Contextual Query Retrieval
Chi Minh Bui (Viettel AI, Viettel Group), Khac-Hoai Nam Bui (Viettel AI, Viettel Group)
CodeRetrievalRepresentation LearningLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the KG-CQR framework, which improves the retrieval phase of retrieval-augmented generation systems by leveraging semantically rich knowledge graph subgraphs to enhance the input query with context.
🎯 What it does: Propose a probabilistic calibration framework KGEC for knowledge graph embedding models, addressing the shortcomings of traditional calibration methods in large class spaces and ranking preservation.
Koel-TTS: Enhancing LLM based Speech Generation with Preference Alignment and Classifier Free Guidance
Shehzeen Samarah Hussain (NVIDIA Corporation), Jason Li (NVIDIA Corporation)
CodeGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextAudio
🎯 What it does: Propose a self-regressive TTS model Koel-TTS based on large language models, achieving fast and natural multilingual speech synthesis through a low-frame-rate audio codec.
KRETA: A Benchmark for Korean Reading and Reasoning in Text-Rich VQA Attuned to Diverse Visual Contexts
Taebaek Hwang (Waddle), Hyunjun Eun (SK Telecom)
CodeLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed and publicly released KRETA—a VQA benchmark for Korean text-rich images, covering 15 industry domains and 26 image types, evaluated using a two-tier reasoning framework (System 1 for basic recognition, System 2 for high-level reasoning).
LaMP-QA: A Benchmark for Personalized Long-form Question Answering
Alireza Salemi (University of Massachusetts Amherst), Hamed Zamani (University of Massachusetts Amherst)
CodeTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed and made public the LaMP-QA benchmark for evaluating long-form personalized question-answering systems, and proposed an evaluation framework based on user history and question narratives.
🎯 What it does: Propose a text-to-audio generation framework called Siren based on language models, which splits the prediction of multi-layer RVQ tokens into multiple collaborative Transformers and uses reinforcement learning for reverse causal alignment to address gradient conflicts and exposure bias caused by orthogonality and semantic decay in RVQ layers.
Lucius E.j. Bynum, Kyunghyun Cho (New York University)
CodeGenerationData SynthesisTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the Sequence-Driven Structural Causal Models (SD-SCM) framework, which automatically generates observational, interventional, and counterfactual sequence data by utilizing language models and user-specified directed acyclic graphs (DAGs), and uses this to construct a new causal inference benchmark;
Language-Guided Temporal Token Pruning for Efficient VideoLLM Processing
Yogesh Kumar (Indian Institute of Technology Jodhpur)
CodeComputational EfficiencyTransformerVision Language ModelVideoText
🎯 What it does: Proposes a language-guided temporal token pruning (LGTTP) method for efficiently processing long videos in vision-language models, which adaptively retains important frames and removes irrelevant frames based on temporal prompts in queries.
Large Language Models for Automated Literature Review: An Evaluation of Reference Generation, Abstract Writing, and Review Composition
Xuemei Tang (Hong Kong Polytechnic University), Zhenguang Cai (Chinese University of Hong Kong)
CodeTransformerLarge Language ModelTextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed a framework for automatically evaluating the ability of large language models (LLMs) to write literature reviews, designed three independent tasks (reference generation, abstract writing, and review writing), and quantified LLM performance across three tasks using multi-dimensional evaluation metrics (false positive rate, accuracy, coverage, factual consistency, semantic coverage).
Large Language Models Meet Knowledge Graphs for Question Answering: Synthesis and Opportunities
Chuangtao Ma (Aalborg University), Haofen Wang (Tongji University)
CodeGraph Neural NetworkTransformerLarge Language ModelGraphReview/Survey Paper
🎯 What it does: Reviews the integration methods of large language models (LLMs) and knowledge graphs (KGs) in question answering, proposes a structured multi-dimensional classification framework, systematically organizes and aligns solutions for various complex question answering tasks, analyzes the advantages and disadvantages of existing technologies, and outlines future research directions.
LASER: An LLM-based ASR Scoring and Evaluation Rubric
Amruta Parulekar (Indian Institute of Technology Bombay), Preethi Jyothi (Indian Institute of Technology Bombay)
CodeRecognitionLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkAudio
🎯 What it does: Proposed a LASER metric for ASR evaluation based on large language models (LLMs), which learns error types through carefully designed prompts and examples, and provides fine-grained scores.
Latent Inter-User Difference Modeling for LLM Personalization
Yilun Qiu, Fuli Feng (University Of Science And Technology Of China)
CodeRepresentation LearningTransformerLarge Language ModelPrompt EngineeringAuto EncoderContrastive LearningText
🎯 What it does: This paper proposes a personalized LLM framework called DEP, which models differences between users in the latent space by generating difference-aware embeddings using user history and contrast information from similar users;
Layer-Aware Representation Filtering: Purifying Finetuning Data to Preserve LLM Safety Alignment
Hao Li (Shanghai Artificial Intelligence Laboratory), Lei Sha (Beihang University)
CodeSafty and PrivacyRepresentation LearningData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Propose Layer-Aware Representation Filtering (LARF), a method that filters training samples causing safety alignment degradation by identifying safety-sensitive layers in LLMs and leveraging their representations.
Learn and Unlearn: Addressing Misinformation in Multilingual LLMs
TaiMing Lu (Johns Hopkins University), Philipp Koehn (Johns Hopkins University)
CodeSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper investigates the spread of misinformation across different languages by injecting training data containing false information into a multilingual LLM (LLaMA3-8B), and systematically evaluates the effectiveness of three 'unlearning' approaches (only English, same-source language, cross-lingual) for eliminating misinformation. Finally, the paper proposes and verifies a method that combines English and the original language of the false information for unlearning, which can almost completely eliminate misinformation in all languages.
Learning from Diverse Reasoning Paths with Routing and Collaboration
Zhenyu Lei (University of Virginia), Jundong Li (University of Virginia)
CodeKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the QR-Distill framework, which efficiently trains small models using multi-path reasoning through three modules: quality filtering, conditional routing, and peer distillation.
🎯 What it does: Proposes an evaluation framework and benchmark (NoisyToolBench) for addressing ambiguous instruction issues in large language model tool usage, along with a prompting technique (Ask-when-Needed, AwN) that enables models to proactively ask for clarification, and implements an automated evaluation tool called ToolEvaluator.
CodeGenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper studies the impact of post-training pruning on the fairness of generating opinion summaries in large language models and proposes a new pruning strategy.
Less Is MuRE: Revisiting Shallow Knowledge Graph Embeddings
Victor Charpenay (Mines Saint-Etienne), Steven Schockaert (Cardiff University)
CodeRepresentation LearningGraphBenchmark
🎯 What it does: This paper systematically studies shallow knowledge graph embedding models, particularly focusing on MuRE as a core framework. It conducts theoretical and experimental evaluations of MuRE's expressiveness, training strategies, and design choices (linear vs. bilinear, trainable bias, scaling and translation, cross-coordinate comparison, region-based width). MuRE is compared with its extension ExpressivE and mainstream shallow models, proposing that MuRE and ExpressivE can serve as new baselines.
🎯 What it does: Designed the LeTS framework, integrating process-level rewards with result-level rewards to enhance reasoning-retrieval behavior in retrieval-augmented generation (RAG);
Leveraging Cognitive Complexity of Texts for Contextualization in Dense Retrieval
Effrosyni Sokli (University of Milano-Bicocca), Gabriella Pasi (University of Milano-Bicocca)
CodeClassificationRetrievalTransformerMixture of ExpertsContrastive LearningText
🎯 What it does: This paper proposes a dense retrieval model called DenseC3, which achieves semantic contextualization of vectors by embedding the cognitive complexity (based on Bloom's cognitive hierarchy) of text into the vector representations of queries and documents.
Leveraging Knowledge Graph-Enhanced LLMs for Context-Aware Medical Consultation
Su-Hyeong Park (Catholic University of Korea), Kang-Min Kim (Catholic University of Korea)
CodeTransformerLarge Language ModelTextGraphElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Propose the ILlama framework, integrating retrieval-augmented generation (RAG) with the UMLS structured knowledge graph to achieve hallucination-free, context-aware medical question answering.
Leveraging Large Models to Evaluate Novel Content: A Case Study on Advertisement Creativity
Zhaoyi Joey Hou (University of Pittsburgh), Xiang Lorraine Li (University of Pittsburgh)
CodeTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper proposes a new benchmark for visual ad creativity evaluation, decomposing creativity into two dimensions: originality and atypicality, and constructing distributed annotations using 25 diverse human ratings.
Leveraging Semantic Triples for Private Document Generation with Local Differential Privacy Guarantees
Stephen Meisenbacher (Technical University of Munich), Florian Matthes (Technical University of Munich)
CodeGenerationSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposes a local differential privacy text generation method DP-ST based on semantic triplets. First, documents are decomposed into SVO triplets, then privacy is applied within the semantic neighborhood using the exponential mechanism, followed by reconstructing coherent text with an LLM.
Leveraging What’s Overfixed: Post-Correction via LLM Grammatical Error Overcorrection
Taehee Park (POSTECH), Gary Lee (POSTECH)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes a two-stage GEC method called PoCO, which first utilizes an LLM to intentionally trigger over-correction to improve recall, and then performs post-correction using a fine-tuned small model to enhance precision.
Liaozhai through the Looking-Glass: On Paratextual Explicitation of Culture-Bound Terms in Machine Translation
Sherrie Shen (University of Edinburgh), Alexandra Birch (University of Edinburgh)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Investigates the generation of paratextual explicitation for culturally bound terms in machine translation, proposing a task to enhance cross-cultural understanding by supplementing annotations from external textual sources;
CodeCompressionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Propose the LightThinker method, which trains LLMs to dynamically compress intermediate thinking steps during inference, condensing lengthy reasoning into a few gist tokens to reduce KV Cache size and inference cost.
🎯 What it does: Proposed a linear-time demonstration selection algorithm based on gradient estimation, which can quickly select the most suitable k examples from a large number of examples as context prompts;
LingGym: How Far Are LLMs from Thinking Like Field Linguists?
Changbing Yang (University Of British Columbia), Jian Zhu (University Of British Columbia)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed and released the LINGGYM benchmark, utilizing Interlinear Glossed Text (IGT) from 18 publicly available reference grammars and designing a Word-Gloss inference task to evaluate large language models' metalinguistic reasoning and structural generalization capabilities in low-resource languages.
Linguistic Neuron Overlap Patterns to Facilitate Cross-lingual Transfer on Low-resource Languages
Yuemei Xu (Beijing Foreign Studies University), Lin Gui (King's College London)
CodeDomain AdaptationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Propose the BridgeX-ICL method, which selects the optimal bridge language in zero-shot cross-lingual context learning for low-resource languages by leveraging inter-lingual neuron overlap patterns;
LinkAlign: Scalable Schema Linking for Real-World Large-Scale Multi-Database Text-to-SQL
Yihan Wang (China Academy of Information and Communications Technology), Xin Yang (China Academy of Information and Communications Technology)
CodeComputational EfficiencyData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the LinkAlign framework, which addresses Schema Linking in large-scale multi-database environments through a step-by-step process, including multi-round semantic retrieval and query rewriting, response filtering to eliminate noise from irrelevant databases, and schema parsing to identify key tables and columns; it also achieves a balance between efficiency and accuracy through pluggable Pipeline and Agent modes.
LiTEx: A Linguistic Taxonomy of Explanations for Understanding Within-Label Variation in Natural Language Inference
Pingjun Hong (LMU Munich), Barbara Plank (LMU Munich)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposed a linguistic explanation classification system named LITEX, specifically designed to analyze different reasoning paths under the same label in natural language inference (NLI), with annotation, verification, and analysis of free-text explanations on the e-SNLI dataset.
LLaMP: Large Language Model Made Powerful for High-fidelity Materials Knowledge Retrieval
Yuan Chiang (University of California Berkeley), Janosh Riebesell (Lawrence Berkeley National Laboratory)
CodeRetrievalTransformerLarge Language ModelAgentic AITabularPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes LLaMP—a hierarchical multi-agent framework for high-fidelity material knowledge retrieval and experimental process simulation in materials science research through large language models (LLMs);
LLM-Driven Completeness and Consistency Evaluation for Cultural Heritage Data Augmentation in Cross-Modal Retrieval
Jian Zhang (Xi'an Jiaotong-Liverpool University), Dongming Lu (Zhejiang University)
CodeRetrievalTransformerLarge Language ModelContrastive LearningMultimodalityChain-of-Thought
🎯 What it does: Propose the C3 framework, which enhances the completeness and consistency of cross-modal retrieval for cultural heritage through LLM-driven text augmentation.
CodeClassificationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Proposes a semantic relation reasoning framework (LGSRR) guided by large language models (LLMs), which automatically mines fine-grained semantics and constructs three logical relationships (importance, complementarity, and inconsistency), enabling efficient reasoning for multimodal intent recognition.
🎯 What it does: Propose a lightweight LLM-independent adaptive retrieval method based on external information and evaluate its effectiveness on multiple datasets.
LMR-BENCH: Evaluating LLM Agent’s Ability on Reproducing Language Modeling Research
Shuo Yan (University of Texas at Dallas), Xinya Du (University of Texas at Dallas)
CodeAI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed the LMR-BENCH benchmark for systematically evaluating the ability of large language model (LLM) agents to reproduce code from natural language processing (NLP) research papers.
LoCt-Instruct: An Automatic Pipeline for Constructing Datasets of Logical Continuous Instructions
Hongyu Sun (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
CodeData SynthesisTransformerLarge Language ModelTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built a fully automated pipeline, LoCt-Pipeline, to generate the MSQA dataset LoCt-Instruct containing logically coherent, multi-turn instruction chains.
LogiCoL: Logically-Informed Contrastive Learning for Set-based Dense Retrieval
Yanzhen Shen (Stanford University), Dan Roth (University of Pennsylvania)
CodeRetrievalTransformerContrastive LearningText
🎯 What it does: Propose LOGICOL, a framework that integrates logical consistency constraints into contrastive learning to enhance the retrieval performance of dense retrieval models on queries containing logical connectives (e.g., AND, OR, NOT).
LogiDynamics: Unraveling the Dynamics of Inductive, Abductive and Deductive Logical Inferences in LLM Reasoning
Tianshi Zheng (Hong Kong University of Science and Technology), Simon See (NVIDIA)
CodeExplainability and InterpretabilityLarge Language ModelImageTextChain-of-Thought
🎯 What it does: Systematic experiments on the logical reasoning dynamics of System 1 (direct reasoning) and System 2 (inductive/deductive reasoning) in large language models;
Jingyao Li (Chinese University of Hong Kong), Jiaya Jia (Hong Kong University of Science and Technology)
CodeKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningBenchmark
🎯 What it does: Proposed a logits-based fine-tuning framework that combines the teacher model's logits with true labels to construct richer training objectives, thereby enhancing the reasoning performance of small LLMs.
🎯 What it does: This paper constructs the largest full-scale ISLR dataset Logos for Russian Sign Language and investigates the impact of cross-lingual transfer learning and visual similar gesture (VSSigns) annotations on model performance.
🎯 What it does: Proposes the Understanding-to-Reasoning Transition (URT) fine-tuning framework, which first enables the model to incorporate partial chain-of-thought (CoT) in the input and learn to generate the remaining reasoning, thereby achieving training for long-chain reasoning;
Long-Form Information Alignment Evaluation Beyond Atomic Facts
Danna Zheng (University of Edinburgh), Jeff Z. Pan (University of Edinburgh)
CodeLarge Language ModelTextSequentialBenchmark
🎯 What it does: Proposes a new attack method for information alignment evaluation called the MontageLie benchmark and designed the DOVESCORE framework that considers event order.
Look Again, Think Slowly: Enhancing Visual Reflection in Vision-Language Models
Pu Jian (Institute of Automation, Chinese Academy of Sciences), Jiajun Zhang (Institute of Automation, Chinese Academy of Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMultimodality
🎯 What it does: Proposed the concept of visual reflection and constructed the Reflection-V model through a two-stage training strategy (cold start data construction + RL based on visual attention rewards).
Lookahead Q-Cache: Achieving More Consistent KV Cache Eviction via Pseudo Query
Yixuan Wang (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
CodeComputational EfficiencyTransformerText
🎯 What it does: Propose the Lookahead Q-Cache (LAQ) framework, which uses pre-generated low-quality pseudo queries to approximate real inference queries during KV cache clearance, significantly improving the accuracy of cache retention;
LyapLock: Bounded Knowledge Preservation in Sequential Large Language Model Editing
Peng Wang (Chinese Academy of Sciences), Songlin Hu (Chinese Academy of Sciences)
CodeOptimizationTransformerLarge Language ModelText
🎯 What it does: Introduce the LyapLock framework in large model continuous editing, utilizing Lyapunov optimization to maintain long-term knowledge accuracy and model stability.
M-BRe: Discovering Training Samples for Relation Extraction from Unlabeled Texts with Large Language Models
Zexuan Li (Nanjing University of Aeronautics and Astronautics), Piji Li (Nanjing University of Aeronautics and Astronautics)
CodeClassificationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Built an M-BRe framework based on large language models for efficiently generating relation extraction training samples from unannotated text