CodeOptimizationComputational EfficiencyRobotic IntelligenceVision-Language-Action Model
π― What it does: Proposes the QVLA quantization framework for vision-language-action models, conducting quantization analysis on the action space for robot control and achieving unified channel-level bit-width allocation and pruning;
QWHA: Quantization-Aware Walsh-Hadamard Adaptation for Parameter-Efficient Fine-Tuning on Large Language Models
Hyesung Jeon (Seoul National University), Jae-Joon Kim (Seoul National University)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes QWHA, a sparse adapter combined with Walsh-Hadamard Transform (WHT), for efficient fine-tuning on large language models with low-bit quantization;
π― What it does: Construct multi-step, interdependent reasoning tasks through query composition (R-HORIZON), and design evaluation benchmarks and reinforcement learning training data based on this to explore the long-term reasoning capabilities of large reasoning models (LRM).
R-Zero: Self-Evolving Reasoning LLM from Zero Data
Chengsong Huang (Tencent AI Seattle Lab), Dong Yu (Washington University in St. Louis)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: Proposed the R-Zero framework, leveraging a challenger and solver from a homologous LLM to self-generate, filter, and learn incrementally difficult training sets through relative policy optimization without any manually annotated data, achieving self-evolving reasoning capability improvements.
R1-Code-Interpreter: LLMs Reason with Code via Supervised and Multi-stage Reinforcement Learning
Yongchao Chen (MIT), Chuchu Fan (MIT)
CodeOptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
π― What it does: Trained and evaluated a Code Interpreter LLM for general tasks using a multi-round text/code generation framework combined with supervised learning and multi-stage Group Relative Policy Optimization (GRPO) reinforcement learning.
π― What it does: Propose R2-Dreamer, a fully decoder-free and data-augmentation-free image reinforcement learning framework that employs internal redundancy reduction (Barlow Twins) regularization to learn high-quality representations.
π― What it does: Proposed a worst-case robust real-time pursuit strategy (R2PS) under partial observability and dynamically changing graph structures, and implemented its cross-graph reinforcement learning training framework;
π― What it does: Proposes RACE Attention, a linear-time, memory-efficient alternative to Softmax Attention that supports training with ultra-long contexts;
π― What it does: Proposes the RADAR framework to address the modeling and reasoning of asymmetric distance matrices in neural vehicle routing planning
π― What it does: Propose a retrieval-based early exit framework RAEE, which dynamically determines the exit layer during inference by leveraging an offline-built exit database, using exit information from similar samples to achieve inference acceleration and improve model accuracy.
CodeTransformerLarge Language ModelTextBenchmarkChain-of-Thought
π― What it does: This paper proposes a gradient-free training method called RAIN-Merging, which can merge instruction tuning models (ITM) with large reasoning models (LRM) while preserving the LRM's 'thinking' format and reasoning quality.
RainPro-8: An Efficient Deep Learning Model to Estimate Rainfall Probabilities Over 8 Hours
Rafael Pablos Sarabia (Aarhus University), Ira Assent (Aarhus University)
CodeComputational EfficiencyConvolutional Neural NetworkTransformerMultimodalityTime Series
π― What it does: Propose an efficient deep learning model, RainPro-8, for predicting the probability distribution of precipitation over the next 8 hours within the European region, integrating multi-source data fusion and probabilistic output;
π― What it does: Propose an extremely simple RA-LDL framework that achieves class-incremental learning for medical image classification by leveraging pre-trained ViT features through random anchor projection, low-rank residual calibration, and closed-form Ridge regression.
π― What it does: Propose a time series learning framework based on stochastic control differential equations, constructing two stochastic dynamical models RF-CDE and R-RDE, which correspond to RBF-lifted signature kernels and rough signature kernels in the infinite width limit;
Random Policy Valuation is Enough for LLM Reasoning with Verifiable Rewards
Haoran He (Hong Kong University Of Science And Technology), Ling Pan (Hong Kong University Of Science And Technology)
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: Propose a RL method called ROVER based on uniform random strategy value assessment, which directly uses the Q-values of the random strategy for greedy or Softmax-based action selection, thereby enhancing mathematical reasoning and diversity in LLMs.
Wenxing Zhou (University of Edinburgh), Timothy Ivor Cannings (University of Edinburgh)
CodeRepresentation Learning
π― What it does: Propose a dimensionality reduction method based on random projection ensemble (RPEDR), which selects the best-performing projection within multiple projection groups, aggregates their outer products, and obtains a low-dimensional subspace through singular value decomposition;
Fangzhou Wu (University of Wisconsin-Madison), Qiuyi Zhang (Google DeepMind)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper proposes a unified mathematical model for KV cache and query routing, and designs two algorithms: randomized leaf token eviction (RLT) and learning-based greedy routing (LBGR), significantly improving inference efficiency in a multi-LLM server environment.
Rank-GRPO: Training LLM-based Conversational Recommender Systems with Reinforcement Learning
Yaochen Zhu (University of Virginia), Nathan Kallus (University of Virginia)
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper proposes the ConvRec-R1 framework, which employs a two-phase training approach: first, generating high-quality, catalog-based demonstration data through the Remap-Reflect-Adjust pipeline for behavior cloning, and then aligning the conversational recommendation model using Rank-GRPO for reinforcement learning.
RAS: Retrieval-And-Structuring for Knowledge-Intensive LLM Generation
Pengcheng Jiang (University of Illinois Urbana Champaign), Jiawei Han (University of Illinois Urbana Champaign)
CodeRetrievalExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelTextGraphBenchmarkRetrieval-Augmented Generation
π― What it does: Construct and utilize a problem-specific knowledge graph for multi-round retrieval and reasoning to enhance the inference accuracy of large language models in knowledge-intensive tasks.
π― What it does: Propose a pragmatic rate-distortion theory based on information theory and implement the RDcomm framework, significantly enhancing task performance and communication efficiency in multi-agent collaborative perception.
Rating Quality of Diverse Time Series Data by Meta-learning from LLM Judgment
Shunyu Wu (Sun Yat-sen University), See-Kiong Ng (National University of Singapore)
CodeData-Centric LearningMeta LearningLarge Language ModelPrompt EngineeringTime Series
π― What it does: This paper proposes the TSRating framework, which uses LLM to comparatively evaluate the quality of multi-domain time series data and trains a Meta learning TSRater model for efficient quality scoring.
REA-RL: Reflection-Aware Online Reinforcement Learning for Efficient Reasoning
Hexuan Deng (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes REA-RL, a reflection-aware online reinforcement learning framework that utilizes small reflection models for online sequential revision and incorporates reflection rewards to reduce overthinking in large-scale reasoning models (LRM), while maintaining or even improving reasoning accuracy.
Read the Room: Video Social Reasoning with Mental-Physical Causal Chains
Lixing Niu (Peking University), Lifeng Fan (BIGAI)
CodeLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: This paper constructs a high-quality video social reasoning benchmark, R3-Bench, and generates a large-scale training set, R3-FDT, through an automated pipeline to evaluate and enhance the psychophysical causal reasoning capabilities of multimodal models.
π― What it does: Analyze the processing of VLM visual encoders layer by layer through logit lens, study the two-stage attribute recognition and semantic disambiguation in the 'what' path, and the spatial geometric structure of 2D RoPE in the 'where' path, and propose instruction-agnostic Run-Length Encoding token compression and RoPE Scaling position enhancement methods based on this.
REAL: Reading Out Transformer Activations for Precise Localization in Language Model Steering
Li-Ming Zhan (Hong Kong Polytechnic University), Xiao-Ming Wu (Hong Kong Polytechnic University)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningText
π― What it does: This paper proposes a framework called REAL for behavior regulation of large language models during inference. It first performs nonlinear decomposition of activations for each attention head or layer using a vector quantized autoencoder (VQ-AE), and learns behavior-related latent subspaces through supervised contrastive loss. Subsequently, it models discrete code sequences with a self-regressive prior, calculates behavior relevance scores for each module, and selects and activates them with importance-weighted injection.
REAP the Experts: Why Pruning Prevails for One-Shot MoE compression
Mike Lasby (Cerebras Systems Inc.), Vithursan Thangarasa (University of Calgary)
CodeCompressionMixture of ExpertsText
π― What it does: Proposed a novel expert pruning method (REAP) based on routing weights and expert activation norms, achieving compression and evaluation on various scales of SMoE language models.
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTabularBenchmarkChain-of-Thought
π― What it does: A reinforcement learning-driven inference framework that directly learns individual political views from survey data and generates structured reasoning and answers.
Reasoning on Time-Series for Financial Technical Analysis
Kelvin J.L. Koa (National University of Singapore), Tat-Seng Chua (National University of Singapore)
CodeExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningTime SeriesFinance RelatedChain-of-Thought
π― What it does: Propose the Verbal Technical Analysis (VTA) framework, integrating natural language reasoning with time series models to achieve interpretable multi-step stock price prediction.
Reasoning or Retrieval? A Study of Answer Attribution on Large Reasoning Models
Yuhui Wang (Stony Brook University), Ting Wang (Stony Brook University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Investigate the answer generation mechanism of large-scale reasoning models (LRM), demonstrating that answers are generated both through chain-of-thought (CoT) reasoning and internal retrieval memory, and verifying the coexistence of these two mechanisms via joint perturbation experiments.
Reasoning Scaffolding: Distilling the Flow of Thought from LLMs
Xiangyu Wen (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
CodeKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes the Reasoning Scaffolding framework, which uses discrete interpretable semantic signals rather than directly copying text for knowledge distillation in LLMs.
Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning
Yunhao Gou (Southern University of Science and Technology), Yu Zhang (Southern University of Science and Technology)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelMultimodalityBenchmark
π― What it does: Proposed the RAPID framework, which separates perception and reasoning in multi-modal large language models (MLLMs). The framework uses MLLM to generate visual descriptions and temporary answers, which are then passed as text context to a powerful text LLM for reasoning. Visual Perception Optimization (VPO) aligns the MLLM's visual descriptions with the final reasoning results, improving reasoning quality.
ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory
Siru Ouyang (UIUC), Tomas Pfister (Google Cloud AI Research)
CodeReinforcement Learning from Human FeedbackLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes the REASONINGBANK memory framework, which can distill transferable reasoning strategies from both successful and failed experiences, and combines with MATTS to achieve self-evolution during testing
Reassessing Layer Pruning in LLMs: New Insights and Methods
Yao Lu (Zhejiang University of Technology), Zhaowei Zhu (D5 Data)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper systematically evaluates hierarchical pruning methods for large language models (LLMs), proposes and verifies an efficient pruning strategy combining simple reverse-order pruning (Reverse-Order) and partial fine-tuning of residual layers (Partial-Layer Fine-Tuning), provides theoretical gradient flow analysis, and ultimately achieves significant pruning effects on Llama-3.1-8B-Instruct, Llama-3-8B, and Llama-3-70B.
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Designed the RECAST framework and constructed the RECAST-30K dataset to train LLMs to accurately follow instructions under multiple constraints, and proposed the RLVC reinforcement learning method to further improve constraint satisfaction rates.
π― What it does: Propose the RECON method, which utilizes unsupervised class pose decomposition and explicit normalization to discover instance-specific symmetric distributions and achieve data alignment canonicalization.
CodeRestorationSuper ResolutionConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingComputed TomographyPhysics Related
π― What it does: Propose a lightweight general-purpose model RAM that can solve multiple computational imaging inverse problems (e.g., deblurring, CT/MRI reconstruction, super-resolution) in a single framework.
π― What it does: Propose a cross-layer KV cache reconstruction method called FusedKV and its lightweight version FusedKV-Lite, which learns to fuse KV caches from lower and middle layers to reduce memory consumption in higher-layer caches.
π― What it does: Proposed a unified evaluation framework RDβ―3, systematically re-evaluating and correcting the post-evaluation inconsistencies of existing separated dataset distillation methods;
ReDDiT: Rehashing Noise for Discrete Visual Generation
Tianren Ma (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
CodeGenerationTransformerDiffusion modelImageText
π― What it does: Proposes ReDDiTβa discrete diffusion transformer that leverages re-hashed noise, significantly improving the quality of discrete visual generation;
π― What it does: Proposed a new two-dimensional (discovery rate and statistical regularity) classification task for corrupted data delearning, and introduced a general delearning method called REM that can balance performance in this two-dimensional space;
Reducing Belief Deviation in Reinforcement Learning for Active Reasoning of LLM Agents
Deyu Zou (Chinese University of Hong Kong), Yu Gong (ByteDance)
CodeTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmark
π― What it does: This paper proposes a reward method called T3 based on early truncation, which controls the belief deviation of LLM agents in multi-round active reasoning, thereby improving credit assignment and training stability in reinforcement learning.
π― What it does: Investigated the issue of inter-class accuracy differences in deep neural networks on balanced data, and proposed a novel regularization method called MR2 to reduce this gap.
Reducing Contextual Stochastic Bilevel Optimization via Structured Function Approximation
Maxime Bouscary (Massachusetts Institute of Technology), Saurabh Amin (Massachusetts Institute of Technology)
CodeOptimizationHyperparameter Search
π― What it does: Proposes a method to reduce the complexity of contextual stochastic bilevel optimization (CSBO) problems through structured function approximation.
Reducing information dependency does not cause training data privacy. Adversarially non-robust features do.
Rasmus Torp (Dartmouth College), Adam Breuer (Dartmouth College)
CodeClassificationSafty and PrivacyAdversarial AttackConvolutional Neural NetworkImage
π― What it does: This paper challenges the traditional view that the stronger the model's dependence on training data, the more susceptible it is to model inversion attacks (MIA) through a series of experiments. It points out that privacy leakage mainly originates from the 'non-robust features' (generalizable but imperceptible features) learned by the model. Subsequently, the Anti-Adversarial Training (AT-AT) training framework is proposed, which actively encourages the model to learn non-robust features, significantly reducing the reconstruction accuracy under MIA while maintaining or even improving model accuracy.
π― What it does: Achieved precise exploitability computation for large-scale incomplete information games, conducting over 7,000 training experiments on seven DRL algorithms;
Reference-guided Policy Optimization for Molecular Optimization via LLM Reasoning
Xuan Li (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
CodeOptimizationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextGraph
π― What it does: Studied instruction-driven molecular optimization tasks, proposing a reference-guided policy optimization (RePO) method that enables large language models to perform multi-step reasoning and generate optimized molecules satisfying similarity constraints without trajectory supervision.
References Improve LLM Alignment in Non-Verifiable Domains
Kejian Shi (Yale University), Arman Cohan (Yale University)
CodeData-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
π― What it does: Propose using high-quality reference answers to guide LLMs as evaluators (RefEval/RefMatch), and apply it to self-improvement of LLM alignment.
π― What it does: Propose the InVirtuoGen model, which utilizes a unified source discrete flow to perform molecule generation, property optimization, and lead optimization on fragmented SMILES.
Refine Now, Query Fast: A Decoupled Refinement Paradigm for Implicit Neural Fields
Tianyu Xiong (Ohio State University), Han Wei Shen
CodeSuper ResolutionComputational EfficiencyRepresentation LearningNeural Radiance FieldPhysics Related
π― What it does: Propose the Decoupled Representation Refinement (DRR) framework, which separates the deep reconstruction network from the low-cost embedding structure to achieve efficient, high-quality Implicit Neural Representation (INR);
RefineStat: Efficient Exploration for Probabilistic Program Synthesis
Madhav Kanda (University of Illinois Urbana-Champaign), Sasa Misailovic (University of Illinois Urbana-Champaign)
CodeComputational EfficiencyAI Code AssistantLarge Language ModelTabular
π― What it does: Propose the REFINESTAT framework, which automatically generates probabilistic programs that satisfy Bayesian workflow reliability metrics using small language models within a semantic constraint and diagnostic-driven iterative process.
Refining Hybrid Genetic Search for CVRP via Reinforcement Learning-Finetuned LLM
Rongjie Zhu (Nanjing University of Information Science and Technology), Zhiguang Cao (Singapore Management University)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringGraphBenchmark
π― What it does: Proposes RFTHGS, a reinforcement learning-based framework to fine-tune a small LLM (14B parameters) for generating high-performance crossover (and subpopulation) operators, thereby enhancing the solving capacity of HGS for the Capacitated Vehicle Routing Problem (CVRP).
π― What it does: To address the problem of small object removal, a two-stage framework is proposed: the first stage utilizes camera adaptive zoom + LoRA fine-tuning to enhance texture restoration of small targets; the second stage eliminates color discrepancies and shadows through mask-based stitching and a shadow repair module that only fine-tunes the decoder, maintaining background consistency.
Reforming the Mechanism: Editing Reasoning Patterns in LLMs with Circuit Reshaping
Zhenyu Lei (University of Virginia), Jundong Li (University of Virginia)
CodeExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningText
π― What it does: Proposes the 'Reasoning Editing' paradigm, which enables selective modification of specific reasoning patterns in large language models (LLMs) while preserving other reasoning capabilities.
Xintong Hao (ByteDance Seed), Chenggang Li (ByteDance Seed)
CodeData SynthesisTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
π― What it does: Propose a data augmentation framework called Massive Genre-Audience reformulation (MGA), which rephrases the original text by generating diverse genre-audience pairs, achieving large-scale, low repetition corpus expansion.
RefTool: Reference-Guided Tool Creation for Knowledge-Intensive Reasoning
Xiao Liu (Peking University), Yansong Feng (Peking University)
CodeComputational EfficiencyAI Code AssistantTransformerLarge Language ModelTextPhysics RelatedChain-of-Thought
π― What it does: Propose a framework called REFTOOL that automatically generates executable tools based on reference materials and uses them in reasoning.
ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding
Jia-Nan Li (Renmin University of China), Chongxuan Li (Renmin University of China)
CodeGenerationData SynthesisComputational EfficiencyTransformerLarge Language ModelDiffusion modelTextBenchmark
π― What it does: Propose a diffusion-based large language model named REFUSION, which combines sequence recombination with causal attention, achieving efficient generation through slot-level parallel selection and autoregressive filling.
ReIn: Conversational Error Recovery with Reasoning Inception
Takyoung Kim (University of Illinois Urbana Champaign), Dilek Hakkani-TΓΌr (University of Illinois Urbana Champaign)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Propose the REIN method, which helps LLM-based dialogue agents quickly diagnose and recover from interaction failures caused by user errors without modifying LLM parameters or system prompts, by injecting an initial reasoning block during the conversation.
Reinforced Latent Reasoning for LLM-based Recommendation
Yang Zhang (National University of Singapore), Tat-Seng Chua (National University of Singapore)
CodeRecommendation SystemTransformerLarge Language ModelReinforcement LearningTextSequentialChain-of-Thought
π― What it does: Propose the LatentR3 framework in recommendation systems, replacing explicit Chain-of-Thought (CoT) reasoning with implicit latent reasoning to enhance LLM recommendation performance.
π― What it does: Under the maximum likelihood estimation (MLE) environment, small language models (LMs) are updated via reinforcement learning gradients to enhance model performance on specific tasks, rather than relying solely on prompts from large models.
Yijun Tian (Xi'an Jiaotong University), Wei Wang (University of Utah)
CodeLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes a Reinforcement Mid-Training (RMT) framework, aiming to leverage unlabeled pre-training data to enhance the model's reasoning and mathematical capabilities during the intermediate phase between pre-training and post-training through reinforcement learning.
Reinforcing Diffusion Models by Direct Group Preference Optimization
Yihong Luo (Hong Kong University Of Science And Technology), Jing Tang (Hong Kong University Of Science And Technology)
CodeGenerationOptimizationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageOrdinary Differential Equation
π― What it does: In the post-training of diffusion models, the Direct Group Preference Optimization (DGPO) algorithm is proposed for model reinforcement learning;
π― What it does: Analyzed and improved the cross-entropy loss in knowledge distillation to better transmit ranking information in recommendation systems; proposed the RCE-KD method, which splits, samples, and adaptively fuses the teacher's top items, significantly enhancing distillation effectiveness.
π― What it does: A new graph Transformer architecture is studied for converting relational databases into heterogeneous temporal graphs for end-to-end learning.
π― What it does: Proposed Relational Transformer (RT), a foundational model capable of zero-shot prediction on new relational databases without requiring task- or dataset-specific fine-tuning.
π― What it does: Propose a multi-view clustering framework RAV that integrates cross-view relation alignment with perspective-adaptive weighted label contrastive learning, aiming to simultaneously preserve neighborhood structure consistency and avoid semantic conflicts caused by low-similarity views.
π― What it does: Propose a reinforcement learning framework (Relative Value Learning, RV) that uses an antisymmetric function to learn the value difference between state pairs, thereby removing the meaningless constant in absolute value;
π― What it does: To address the problem of visual forgery localization, the RelayFormer framework is proposed, which can uniformly process images and videos of different resolutions.
Reliable Probabilistic Forecasting of Irregular Time Series through Marginalization-Consistent Flows
Vijaya Krishna Yalavarthi (University of Hildesheim), Lars Schmidt-Thieme (University of Hildesheim)
CodeTransformerFlow-based ModelTime SeriesBiomedical DataElectronic Health Records
π― What it does: Proposed a hybrid separable flow model called MOSES for reliable probabilistic prediction of irregular time series, ensuring marginal consistency while also capturing joint distribution expressions.
CodeSafty and PrivacyComputational EfficiencyConvolutional Neural NetworkTransformerImage
π― What it does: This paper proposes a novel Machine Unlearning method called MU-Mis, which directly suppresses the contribution of samples during training by minimizing the input sensitivity gap between the target class logit and irrelevant class logit of the forgotten samples, thereby achieving model unlearning after training;
REMem: Reasoning with Episodic Memory in Language Agent
Yiheng Shu (Ohio State University), Yu Su (Ohio State University)
CodeTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
π― What it does: Proposes the REMem two-stage framework for constructing time-aware event memory graphs and achieving recall and reasoning in language agents through tool-driven multi-step retrieval.
Michael Amir (University of Cambridge), Amanda Prorok (University of Cambridge)
CodeRobotic IntelligenceTime SeriesSequential
π― What it does: Designed and implemented a detectable watermarking technique for robot control strategies under remote observation, addressing the physical observation gap problem.
RepIt: Steering Language Models with Concept-Specific Refusal Vectors
Vincent Siu (University of California, Santa Cruz), Chenguang Wang (University of California, Santa Cruz)
CodeSafty and PrivacyExplainability and InterpretabilityRepresentation LearningTransformerTextBenchmark
π― What it does: Propose the REPIT framework to precisely isolate and edit rejection behaviors for specific concepts in language models, allowing 'unbanning' of specific categories such as weapon information while maintaining rejection of other harmful concepts.
π― What it does: Proposes a self-supervised representation alignment method called SRA, which does not require external representation components, and utilizes the step-by-step denoising process of the diffusion Transformer to achieve self-guided internal representations;
Representing local protein environments with machine learning force fields
Meital Bojan (IST Austria), Alexander Bronstein (IST Austria)
CodeRepresentation LearningProtein Structure PredictionGraph Neural NetworkBiomedical Data
π― What it does: Construct a general representation of protein local environments using atomic embeddings from pre-trained machine learning force fields (MLFF), validated on downstream tasks such as pKa prediction, chemical shift prediction, and secondary structure identification.
RePrompt: Reasoning-Augmented Reprompting for Text-to-Image Generation via Reinforcement Learning
Mingrui Wu (Xiamen University), Rongrong Ji (Microsoft)
CodeGenerationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringImageTextMultimodalityChain-of-Thought
π― What it does: Proposes the RePrompt framework based on reinforcement learning, which significantly improves the semantic and visual alignment in text-to-image tasks by training language models to generate enhanced prompts containing reasoning trajectories.
π― What it does: Proposed a framework (FORMED) that repurposes a generic time series foundation model for medical time series classification by freezing the pre-trained backbone and introducing adaptable channel embeddings, label queries, and shared decoding attention to flexibly handle varying channels, lengths, and class numbers;
Jiahao Shi (Purdue University), Tianyi Zhang (Purdue University)
CodeSafty and PrivacyAI Code AssistantTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose a retrieval-enhanced secure code generation framework called RESCUE, which automatically builds a hierarchical security knowledge base and enhances LLM secure code generation through multi-dimensional retrieval.
π― What it does: Proposes the ResiliBench benchmark to evaluate the workflow execution capability of LLMs in real-world environments with fluctuating tool reliability and instruction quality.
Resisting Contextual Interference in RAG via Parametric-Knowledge Reinforcement
Chenyu Lin (Baidu Inc), Zhonghou Lv
CodeGenerationRetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
π― What it does: Propose the Knowledgeable-R1 framework, training LLMs in retrieval-augmented generation (RAG) tasks by dynamically balancing parameter knowledge (PK) and retrieved context (CK) through joint sampling and reinforcement learning, thereby mitigating context interference.
Resp-Agent: An Agent-Based System for Multimodal Respiratory Sound Generation and Disease Diagnosis
Pengfei ZHANG, Li Liu (Hong Kong University of Science and Technology)
CodeClassificationGenerationData SynthesisAnomaly DetectionTransformerLarge Language ModelAgentic AIFlow-based ModelMultimodalityBiomedical DataAudio
π― What it does: Proposed a closed-loop multimodal system, Resp-Agent, integrating breath sound synthesis and diagnosis to achieve controllable breath sound generation and disease classification;
π― What it does: Propose a generic denoising and view synthesis framework called ReSplat, which can achieve multi-view image restoration and novel view rendering under any degraded input;
ResT: Reshaping Token-Level Policy Gradients for Tool-Use Large Language Models
Zihan Lin (MAIS and NLPR, Institute of Automation, Chinese Academy of Sciences), Ran He (Meituan)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
π― What it does: Proposed a tool calling method based on reinforcement learning called ResT, significantly improving the performance of large language models in multi-round tool usage tasks through entropy-aware reshaping of token-level policy gradients and lightweight curriculum learning.
Resurfacing the Instance-only Dependent Label Noise Model through Loss Correction
Mustafa Enes AydΔ±n, Alexander Bertrand (KU Leuven)
CodeClassificationImageTabularAudio
π― What it does: In binary classification tasks, the authors propose an instance-aware loss correction method based on risk equivalence to address label noise.
ResWorld: Temporal Residual World Model for End-to-End Autonomous Driving
Jinqing Zhang (Beihang University), Yunhong Wang (Beihang University)
CodeAutonomous DrivingWorld ModelImageBenchmark
π― What it does: Propose a temporal residual-based world model called ResWorld, which focuses on dynamic object prediction using residuals to improve end-to-end autonomous driving planning;
Retaining Suboptimal Actions to Follow Shifting Optima in Multi-Agent Reinforcement Learning
Yonghyeon Jo (Ulsan National Institute of Science and Technology), Seungyul Han (Ulsan National Institute of Science and Technology)
CodeReinforcement LearningBenchmark
π― What it does: Proposed the Successive Sub-value Q-learning (S2Q) framework, which utilizes multiple sub-value functions to continuously track sub-optimal actions, addressing the issue of value functions dynamically changing during training.
Rethinking Code Similarity for Automated Algorithm Design with LLMs
Rui Zhang (City University of Hong Kong), Zhichao Lu (City University of Hong Kong)
CodeOptimizationAI Code AssistantTransformerLarge Language ModelSequential
π― What it does: Proposed a method called BehaveSim for algorithm similarity measurement based on problem-solving behavior trajectory (PSTrajectory), and applied it to the scenarios of large language model automatic algorithm design (LLMAAD) and algorithm analysis.
π― What it does: Propose a Progressive Neural Collapse (ProNC) framework for continuous learning tasks, which enhances feature representations and reduces catastrophic forgetting by dynamically generating and aligning simple ETF objectives.
Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods
Wanru Zhao (University of Cambridge), Nicholas D. Lane (University of Cambridge)
CodeData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Propose the ADAPT framework, which dynamically adjusts the importance of training samples through online sample reweighting, replacing traditional offline screening/mixing processes, to achieve adaptive data processing for large language model (LLM) training and instruction tuning;
CodeGenerationTransformerVision Language ModelDiffusion modelImageText
π― What it does: This paper re-evaluates the global text modulation mechanism in diffusion transformers, proposing a training-free 'modulation guidance' method that utilizes pooled CLIP embeddings for controllable guidance in the modulation space;
π― What it does: Propose the Latent Reasoning Tuning (LRT) framework, which uses a lightweight reasoning network to generate implicit reasoning representations, replacing traditional long-text reasoning trajectories.
Rethinking LoRA for Privacy-Preserving Federated Learning in Large Models
Jin Liu (Xidian University), Junkang Liu (Xidian University)
CodeFederated LearningSafty and PrivacyImageText
π― What it does: This paper proposes a privacy-preserving federated learning framework called LA-LoRA, which alternately updates the two low-rank matrices of LoRA within each local training round and applies low-pass filtering to gradients to mitigate issues of gradient coupling, noise amplification, and sharp global convergence.
Rethinking Pareto Frontier: On the Optimal Trade-offs in Fair Classification
Junyi Chai (Purdue University), Xiaoqian Wang (Purdue University)
CodeClassificationSupervised Fine-TuningImageTabularBiomedical Data
π― What it does: This paper proposes a model-specific (MS) Pareto-optimal fairness-accuracy trade-off and the balance between fairness and accuracy, and designs a group-dependent bias-aware last-layer fine-tuning framework to improve the equilibrium between fairness and accuracy;
π― What it does: In a large-scale parallel reinforcement learning environment, a method called Coupled Policy Optimization (CPO) is proposed, which regulates diversity among multiple policies by incorporating KL constraints and adversarial rewards within a leader-follower framework, thereby achieving more efficient and stable sample collection and learning.
Rethinking Reasoning in Document Ranking: Why Chain-of-Thought Falls Short
Xuan Lu (Shanghai Jiao Tong University), Xiaoyu Shen (Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: Systematically evaluated the effectiveness of chained reasoning in document re-ranking, comparing point-wise and list-wise re-ranking models, using SFT and GRPO training.
Rethinking Residual Errors in Compensation-based LLM Quantization
Shuaiting Li (Zhejiang University), Kejie Huang (Zhejiang University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Redefined the residual error of GPTQ and GPTAQ, expanding the error source to internal errors caused by weight compensation and incorporating it into weight updates;
π― What it does: Propose the DCFlow framework to achieve unsupervised cross-modal optical flow estimation, integrating separated optimization and cross-modal consistency constraints;
Retrieval-of-Thought: Efficient Reasoning via Reusing Thoughts
Ammar Ahmed (University of Minnesota), Ali Anwar (University of Minnesota)
CodeComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes the Retrieval-of-Thought (RoT) framework, which constructs a thinking graph with sequential and semantic edges, dynamically retrieves and combines previous reasoning steps during inference to generate reusable thinking templates, thereby improving the efficiency of large language model reasoning.