International Conference on Learning Representations · 2207 papers
Beyond Markovian: Reflective Exploration via Bayes-Adaptive RL for LLM Reasoning
Shenao Zhang (Northwestern University), Yunxuan Li (Google)
CodeLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes BARL, a Bayesian reinforcement learning (RL) algorithm for LLM reflective exploration, which can dynamically update beliefs about the environment and proactively reflect to improve reasoning strategies during inference.
Beyond Match Maximization and Fairness: Retention-Optimized Two-Sided Matching
Ren Kishimoto (Institute of Science Tokyo), Yuta Saito (Hanuku-kaso, Co., Ltd.)
CodeRecommendation SystemOptimizationGraphTabular
🎯 What it does: This paper proposes and evaluates an algorithm called MRet based on dynamic learning to rank (LTR), aiming to maximize user retention on two-sided matching platforms, rather than solely pursuing the number of matches or fairness.
Beyond Noisy-TVs: Noise-Robust Exploration Via Learning Progress Monitoring
Zhibo Hou (University of California, Merced), Wan Du (University of California, Merced)
CodeReinforcement LearningImage
🎯 What it does: Proposed a noise-robust exploration method based on Learning Progress Monitoring (LPM) and validated its effectiveness in multiple environments.
Beyond Pass@ 1: Self-Play with Variational Problem Synthesis Sustains RLVR
Xiao Liang (University Of California), Weizhu Chen (Microsoft)
CodeData SynthesisTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes an online self-play variant problem synthesis (SVS) strategy, which uses the model itself to generate and solve challenging training sample variants, thereby achieving data augmentation without external labels in RLVR training while maintaining stable policy entropy.
Beyond Penalization: Diffusion-based Out-of-Distribution Detection and Selective Regularization in Offline Reinforcement Learning
Qingjun Wang (Tongji University), Guang Chen (Shanghai Innovation Institute)
CodeAnomaly DetectionReinforcement LearningDiffusion model
🎯 What it does: Proposed DOSER, a framework based on diffusion models for out-of-distribution (OOD) detection and selective regularization in offline reinforcement learning.
🎯 What it does: Proposed a learning-based decentralized attention retrieval framework LDAR, which can adaptively select retrieval intervals from similarity distributions to reduce interference with large language models.
Beyond Raw Detection Scores: Markov-Informed Calibration for Boosting Machine-Generated Text Detection
Chenwang Wu (Hong Kong Baptist University), Defu Lian (University of Science and Technology of China)
CodeAnomaly DetectionText
🎯 What it does: Proposes a lightweight calibration method based on Markov random fields, leveraging the proximity similarity and initial instability of token-level detection scores to enhance the effectiveness of detecting machine-generated text.
Beyond Real: Imaginary Extension of Rotary Position Embeddings for Long-Context LLMs
Xiaoran Liu (Fudan University), Xipeng Qiu (Fudan University)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Reintroduce the neglected imaginary part information based on RoPE rotational position encoding, constructing dual attention heads that simultaneously include real and imaginary parts, forming RoPE++.
Haike Xu (Massachusetts Institute of Technology), Tong Chen (University of Washington)
CodeRetrievalLarge Language ModelContrastive LearningTextMultimodalityBenchmark
🎯 What it does: Proposed a new retrieval pipeline called Reranker-Guided-Search (RGS), which improves retrieval accuracy under a limited reranker budget by performing greedy search on document similarity graphs and leveraging reranker preferences to select documents requiring re-ranking.
🎯 What it does: Proposed two multi-objective multigraph path planning models based on graph neural networks: GMS-EB, which directly performs autoregressive edge selection on the multigraph, and GMS-DH, which first conducts non-autoregressive pruning followed by autoregressive path construction.
Beyond Softmax and Entropy: Convergence Rates of Policy Gradients with $\boldsymbol{f}$-SoftArgmax Parameterization $\&$ Coupled Regularization
Safwan Labbi (Centre de Mathématiques Appliquées, French National Center for Scientific Research, École Polytechnique, Institut Polytechnique de Paris), Eric Moulines (Mohamed bin Zayed University of Artificial Intelligence)
CodeOptimizationReinforcement Learning
🎯 What it does: Propose the f-softargmax strategy parameterization coupled with the corresponding f-divergence regularization, derive explicit convergence rates and sample complexity of stochastic policy gradients under no preprocessing conditions, and conduct theoretical analysis and experimental validation on discrete MDPs.
Beyond Static Vision: Scene Dynamic Field Unlocks Intuitive Physics Understanding in Multi-modal Large Language Models
Nanxi Li (Shanghai Jiao Tong University), Yong-Lu Li (Shanghai Innovation Institute)
CodeData SynthesisSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: Propose two low-level physical perception tasks (Next Frame Selection and Temporal Coherence Verification) to systematically evaluate MLLMs' understanding of intuitive physical dynamics.
CodeKnowledge DistillationMixture of ExpertsImageTextMultimodality
🎯 What it does: Propose InherNet in knowledge distillation, directly inheriting the teacher network's structure and knowledge through low-rank SVD decomposition, rather than training the student network
Beyond Text-Only: Towards Multimodal Table Retrieval in Open-World
Da Li (State Key Laboratory Of AI Safety), Xueqi Cheng (State Key Laboratory Of AI Safety)
CodeRetrievalLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Constructed and publicly released the image-based multimodal table retrieval benchmark TaR-ViR, transforming traditional text-based table retrieval tasks into image retrieval problems;
Beyond Text-to-Image: Liberating Generation with a Unified Discrete Diffusion Model
Qingyu Shi (Peking University), Shuicheng YAN
CodeGenerationData SynthesisTransformerVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Proposes Muddit, a unified discrete diffusion model that achieves efficient parallel generation for text-to-image, image-to-text, and visual question answering.
🎯 What it does: Systematic evaluation of a TSP solver under the 'Heatmap+MCTS' framework, quantifying the impact of MCTS hyperparameters and heatmap complexity on solution quality, and proposing a parameter-free heatmap GT-Prior based on k-Nearest Neighbor and a standardized MCTS hyperparameter tuning pipeline.
BeyondBench: Contamination-Resistant Evaluation of Reasoning in Language Models
Gaurav Srivastava (Virginia Tech), Xuan Wang (Virginia Tech)
CodeData-Centric LearningLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposed and implemented BEYONDBENCH, a dynamic reasoning evaluation framework based on algorithm generation, verifiability, and Token-awareness, covering 44 algorithm tasks, 117 variants, and including three difficulty levels: easy, medium, and hard;
Bi-Lipschitz Autoencoder With Injectivity Guarantee
Qipeng Zhan (University of Pennsylvania), Li Shen (University of Pennsylvania)
CodeRepresentation LearningAuto EncoderImage
🎯 What it does: This paper addresses the local optimum problem caused by non-injectivity of autoencoders in low-dimensional manifold learning, proposing an injective and robust Bi-Lipschitz Autoencoder (BLAE).
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelImageText
🎯 What it does: Propose Bi-LoRA, a dual low-rank adapter that enhances generalization performance in large model fine-tuning by decoupling SAM's sharpness optimization and task adaptation through auxiliary adversarial LoRA modules, enabling single-step parallel updates while maintaining low memory and time costs.
Bias Similarity Measurement: A Black-Box Audit of Fairness Across LLMs
Hyejun Jeong (University of Massachusetts Amherst), Amir Houmansadr (University of Massachusetts Amherst)
CodeSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the Bias Similarity Measurement (BSM) framework, conducting inter-comparison of fairness among 30 LLMs on over 1 million prompts.
BiasBusters: Uncovering and Mitigating Tool Selection Bias in Large Language Models
Thierry Blankenstein (University of Oxford), Adel Bibi (University of Oxford)
CodeData SynthesisExplainability and InterpretabilityData-Centric LearningLarge Language ModelTextBenchmark
🎯 What it does: Investigate the selection bias of large language models when invoking external tools, propose an evaluation benchmark, analyze the root causes of bias, and provide a lightweight mitigation solution.
BiasFreeBench: a Benchmark for Mitigating Bias in Large Language Model Responses
Xin Xu (University Of California San Diego), Zexue He (University Of California San Diego)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a unified bias mitigation benchmark, BIASFREEBENCH, to directly evaluate the fairness, safety, and anti-stereotyping level of responses generated by large language models (LLMs);
🎯 What it does: This paper proposes BITTA, a bilateral information-aware test-time adaptation framework for vision-language models, aiming to enhance model robustness under distribution drift by learning from low-entropy samples and unlearning from high-entropy samples.
BioCAP: Exploiting Synthetic Captions Beyond Labels in Biological Foundation Models
Ziheng Zhang (Ohio State University), Jianyang Gu (Ohio State University)
CodeClassificationRetrievalExplainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Construct the BIOCAP multimodal foundation model by jointly training on the TreeOfLife-10M dataset, combining images, species labels, and instantiated descriptive captions generated by MLLM guided by Wikipedia visual information and classification-level format examples, to enhance understanding and retrieval of biological images.
BioTamperNet: Affinity-Guided State-Space Model Detecting Tampered Biomedical Images
Soumyaroop Nandi (University of Southern California), Prem Natarajan (University of Southern California)
CodeAnomaly DetectionTransformerGenerative Adversarial NetworkBiomedical Data
🎯 What it does: Proposed BioTamperNet, a unified dual-tower architecture for detecting copy-paste regions (external copy and internal copy) and sharp transitions in medical images, capable of simultaneously locating copy sources and targets.
🎯 What it does: Propose the BioX-Bridge framework, which achieves unsupervised cross-modal knowledge transfer on unlabeled new biosignal modalities through a lightweight bridge network.
Black-Box Privacy Attacks on Shared Representations in Multitask Learning
John Abascal (Northeastern University), Matthew Jagielski (Google Deepmind)
CodeSafty and PrivacyRepresentation LearningAdversarial AttackImageText
🎯 What it does: This paper investigates privacy leakage in shared representations within multi-task learning (MTL), proposes a black-box task inference threat model, and demonstrates an attack method capable of determining whether a specific task participated in training by querying only the embedding vectors of the shared encoder.
🎯 What it does: Propose the BOLT framework in robotic scenarios, first using Option-Level Decision Distillation (ODD) to transfer the multi-choice decision surface of large models to small models, then during inference dynamically deciding whether to activate high-resolution re-evaluation, type-matching retrieval, or question decomposition on a per-instance basis through a budget-aware router (bTTA), to enhance decision quality under strict latency, memory, and energy consumption budgets.
🎯 What it does: Designed a branching strategy called ImitSAT for CDCL SAT solvers based on imitation learning, trained by refining expert KeyTrace sequences.
🎯 What it does: Propose and verify a technique called Boomerang Distillation, which utilizes a pre-trained large teacher model. First, the teacher's layers are trimmed in blocks and a smaller student model is trained using knowledge distillation. Subsequently, during inference, without further training, the student model is interleaved with the teacher layers to create a series of intermediate-sized models;
Ningfeng Yang (University of British Columbia), Tor M. Aamodt (University of British Columbia)
CodeComputational EfficiencyText
🎯 What it does: Proposed a low-precision quantization method called Bell Box Quantization (BBQ), aiming to achieve computational efficiency while maintaining information-theoretic optimality (ITO).
Boosting Medical Visual Understanding From Multi-Granular Language Learning
Zihan Li (University of Washington), Paul Kinahan (University of Washington)
CodeClassificationRecognitionTransformerVision Language ModelContrastive LearningBiomedical Data
🎯 What it does: Propose the Multi-Granular Language Learning (MGLL) framework, which leverages multi-granular text information and soft label supervision to enhance visual-text alignment in medical imaging, achieving consistency across multi-label and cross-granular tasks, and applying the visual encoder to various medical vision downstream tasks.
Boosting Multi-Domain Reasoning of LLMs via Curvature-Guided Policy Optimization
Xize Liang (University of Science and Technology of China), Jianye HAO
CodeOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Designed and validated a curvature-guided policy optimization framework CGPO to enhance the reasoning capabilities of large language models (LLMs) in multi-domain reinforcement learning (RL) environments.
🎯 What it does: Proposes a set of temperature schedules (e.g., Negative Cosine Schedule) for dynamically adjusting the model's temperature in open-set recognition (OSR) tasks, achieving a balance between instance-level and class-level features during training, and embedding the schedule into existing CE, SupCon, and ARPL losses without additional overhead.
🎯 What it does: Proposed a new low-rank adaptation method called BoRA, which enhances the expressiveness of LoRA weights by introducing block diversity.
Bound by semanticity: universal laws governing the generalization-identification tradeoff
Marco Nurisso (Politecnico di Torino), Giovanni Petri (Northeastern University)
CodeClassificationRecognitionConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: Investigated the inevitable trade-off between generalization and recognition in intelligent systems under limited semantic resolution, and provided a general Pareto frontier.
Bounds of Chain-of-Thought Robustness: Reasoning Steps, Embed Norms, and Beyond
Dingzirui Wang (Harbin Institute of Technology), Yang Deng (Singapore Management University)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper theoretically analyzes the upper bound of robustness of chain-of-thought (CoT) under input perturbations and applies this analysis to the linear self-attention model, verifying the negative correlation between the norms of input embeddings and hidden states and robustness.
🎯 What it does: This study constructs BrainSemantoks, a self-distilled foundational model for fMRI, which learns abstract representations of brain dynamics through a semantic tokenizer and is pre-trained accordingly.
🎯 What it does: This paper proposes a Gradient Temporal Collinearity (GTC) metric to analyze the reasons behind poor robustness in direct encoding, and based on this, designs a Structured Temporal Orthogonal De-correlation (STOD) method to enhance the robustness of Spiking Neural Networks (SNNs).
🎯 What it does: Propose a closed-loop trajectory planning method called BridgeDrive based on the diffusion bridge strategy, which uses anchors (typical expert driving trajectories) to guide the diffusion process and realizes real-time planning in the CARLA simulation environment.
Bridging Draft Policy Misalignment: Group Tree Optimization for Speculative Decoding
Shijing Hu, Pan Zhou
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose Group Tree Optimization (GTO), which trains the draft model with tree-level reward training, aligning the tree strategy during inference with the training objectives, thereby improving the inference speed of large language models.
Bridging Fairness and Explainability: Can Input-Based Explanations Promote Fairness in Hate Speech Detection?
Yifan Wang (Saarland University), Vera Demberg (Saarland University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: In the hate speech detection task, the relationship between input-based explanations (rationales) and fairness was systematically quantitatively studied, covering three aspects: bias detection, model selection, and bias mitigation.
🎯 What it does: Proposes a unified framework that simultaneously accomplishes expressive piano performance rendering (EPR) and automatic piano transcription (APT), achieving content-style decoupling;
Bridging Successor Measure and Online Policy Learning with Flow Matching-Based Representations
Haosen Shi (Chinese University of Hong Kong), Sinno Jialin Pan (Ant International)
CodeReinforcement LearningFlow-based Model
🎯 What it does: Proposes the Successor Flow Features (SF2) framework, which leverages flow matching techniques to learn the Successor Measure and projects it into compact time-invariant state-action features, thereby achieving representation learning and policy optimization in online reinforcement learning.
Bridging the performance-gap between target-free and target-based reinforcement learning
Théo Vincent (DFKI), Carlo D'Eramo (University of Würzburg)
CodeReinforcement LearningImageVideoText
🎯 What it does: Propose a structure where only the final linear layer of the Q-network is copied as the target network, while sharing the remaining parameters with the online network; combine iterative Q-learning to perform multi-head parallel learning of multi-step Bellman iterations on the same network, thereby improving sample efficiency while maintaining low memory usage.
Bringing Stability to Diffusion: Decomposing and Reducing Variance of Training Masked Diffusion Models
Mengni Jia (University of Cambridge), guanjunjiang
CodeGenerationDiffusion modelTextMultimodality
🎯 What it does: This paper proposes variance decomposition during training of Masked Diffusion Models (MDMs) and designs six variance reduction methods based on this, with the core methods being P-POTS (Pareto optimal time step sampling) and MIRROR (mirror masking), significantly improving training stability and performance.
BrowseNet: Graph-Based Associative Memory for Contextual Information Retrieval
PAVAN KUMAR S (Indian Institute of Technology Madras), Nirav Pravinbhai Bhatt (Indian Institute of Technology Madras)
CodeRetrievalLarge Language ModelGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Construct a graph structure based on entity associations (Graph-of-Chunks) and achieve retrieval and memory for multi-hop question answering through query subgraph exploration.
🎯 What it does: Proposes Block-Wise Caching (BWCache), a training-free acceleration method that caches blocks and dynamically reuses intermediate features in Diffusion Transformer (DiT).
Cache What Lasts: Token Retention for Memory-Bounded KV Cache in LLMs
Ngoc Bui (Yale University), Rex Ying (JPMorgan Chase AI Research)
CodeOptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposed a KV cache eviction strategy called TRIM-KV based on retention gates, which dynamically retains the most important tokens under a fixed memory budget.
Cache-to-Cache: Direct Semantic Communication Between Large Language Models
Tianyu Fu (Tsinghua University), Yu Wang (Chinese University of Hong Kong)
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose Cache-to-Cache (C2C) and demonstrate its capability to directly share KV-Cache for semantic communication in multi-LLM systems, improving accuracy and speed.
Cactus: Accelerating Auto-Regressive Decoding with Constrained Acceptance Speculative Sampling
Yongchang Hao (University of Alberta), Lili Mou (University of Alberta)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a constraint-based optimization method for accelerating autoregressive decoding called Cactus, which improves upon traditional Speculative Sampling.
cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning
Maksim Kolodiazhnyi (Lomonosov Moscow State University), Danila Rukhovich (Institute of Mechanics Armenia)
CodeGenerationAI Code AssistantLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityPoint Cloud
🎯 What it does: Developed a multi-modal CAD reconstruction model capable of receiving point cloud, image, and text inputs simultaneously, and outputting executable Python CAD code
Calibrating Verbalized Confidence with Self-Generated Distractors
Victor Wang (University of Texas at Austin), Elias Stengel-Eskin (University of Texas at Austin)
CodeExplainability and InterpretabilityTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: Propose the DINCO method, which normalizes the verbalized confidence of LLMs through self-generated distractors and NLI weights, combined with self-consistency to calibrate the model's confidence estimates.
CALM: Co-evolution of Algorithms and Language Model for Automatic Heuristic Design
Ziyao Huang (City University of Hong Kong), Jianping Wang (City University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTabularBenchmark
🎯 What it does: Propose the CALM framework, achieving co-evolution between language models and heuristic search, automatically generating and iteratively optimizing heuristic algorithms for various combinatorial optimization problems.
Can SAEs reveal and mitigate racial biases of LLMs in healthcare?
Hiba Ahsan (Northeastern University), Byron C Wallace
CodeExplainability and InterpretabilityTransformerLarge Language ModelAuto EncoderBiomedical DataElectronic Health RecordsChain-of-Thought
🎯 What it does: This paper uses a sparse autoencoder (SAE) to interpret the internal activations of the Gemma-2 LLM, revealing that the model associates Black identity with negative concepts such as crime and drugs in clinical text. The causal impact of this association on prediction outcomes is demonstrated through steering of the corresponding SAE latent variables, while the effectiveness of mitigating bias by zeroing out these latent variables is evaluated.
🎯 What it does: This study proposes a method that replaces the nonlinear activations (GeLU, softmax) in Transformers with learnable piecewise linear functions, subsequently optimizes the architecture in two stages, and uses this method to evaluate the transferability between different tasks.
Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective.
Ruichuan An, Jiang Bian (Microsoft Research Asia)
CodeClassificationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposes AesEval-Bench, a comprehensive benchmark for graphic design aesthetic evaluation, encompassing four dimensions, twelve indicators, and three quantifiable tasks (aesthetic judgment, region selection, and precise localization). Subsequently, the authors conduct systematic evaluations of various proprietary, open-source, and reasoning-enhanced visual language models (VLMs), finding that they achieve approximately 72–73% accuracy in aesthetic judgment but perform significantly worse in region selection and precise localization, particularly with IoU below 0.2 for localization. To improve performance, the authors construct the AesEval-Train training set, generating task labels and reasoning paths through two methods: 'human-in-the-loop VLM annotation' and 'metric-associated reasoning.' They then perform full-parameter fine-tuning on Qwen2.5-VL-7B, significantly enhancing performance across all three tasks.
Can we generate portable representations for clinical time series data using LLMs?
Zongliang Ji (University of Toronto), Rahul G Krishnan
CodeDomain AdaptationSafty and PrivacyRepresentation LearningTransformerLarge Language ModelTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This study investigates whether large language models (LLMs) can generate portable clinical time series data representations for effective deployment of prediction models across different hospitals.
🎯 What it does: Proposed Canonical Tree Cover Neural Networks (CTNNs), which utilize multiple minimum spanning trees to cover the graph in order to obtain expressive and reversible representations that preserve graph distances;
Capability-Based Scaling Trends for LLM-Based Red-Teaming
Alexander Panfilov (ELLIS Institute), Jonas Geiping (ELLIS Institute)
CodeSafty and PrivacyAdversarial AttackLarge Language ModelTextBenchmark
🎯 What it does: This paper investigates the scaling trends of LLM-based red team attacks with the capability gap through large-scale experiments (600+ attacker-target combinations);
Capacity-Aware Inference: Mitigating the Straggler Effect in Mixture of Experts
Shwai He (University of Maryland), Ang Li (University of Maryland)
CodeOptimizationComputational EfficiencyMixture of ExpertsTextMultimodality
🎯 What it does: By introducing capacity-aware Token Drop and Expanded Drop during the MoE inference phase, dynamically limiting expert load and utilizing the idle capacity of underloaded experts, thus alleviating the Straggler effect, improving inference efficiency while maintaining model performance.
CardioComposer: Leveraging Differentiable Geometry for Compositional Control of Anatomical Diffusion Models
Karim Kadry (Massachusetts Institute of Technology), Elazer R Edelman
CodeGenerationDiffusion modelAuto EncoderBiomedical Data
🎯 What it does: Utilizing differentiable geometric constraints to generate controllable three-dimensional multi-class anatomical label maps in unconditional diffusion models according to shape, size, and position
🎯 What it does: Proposed a batch normalization method called CaRe-BN specifically designed for spiking neural networks (SNN) in reinforcement learning scenarios, addressing the gradient instability caused by inaccurate estimation under non-stationary distributions in traditional BN.
CARL: Camera-Agnostic Representation Learning for Spectral Image Analysis
Alexander Baumann (Siemens AG), Slobodan Ilic (Siemens AG)
CodeSegmentationAutonomous DrivingConvolutional Neural NetworkTransformerContrastive LearningImageMultimodalityBiomedical Data
🎯 What it does: Proposes CARL, a camera-agnostic spectral image representation learning framework that learns general features across RGB, MS, and HSI cameras, and achieves cross-modal knowledge transfer through self-supervised pre-training.
CARPRT: Class-Aware Zero-Shot Prompt Reweighting for Vision-Language Model
Ruijiang Dong (University Of Melbourne), Masashi Sugiyama (Riken Center For Advanced Intelligence Project)
CodeClassificationPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: In zero-shot classification for black-box vision-language models (VLMs), a training-agnostic class-aware prompt reweighting method (CARPRT) is proposed, which automatically infers prompt weights for each class by leveraging only the similarity scores between unlabeled images and pre-trained VLMs, thereby enhancing zero-shot inference performance.
🎯 What it does: Propose a training-free method called CASteer, which dynamically suppresses and eliminates unwanted concepts (including abstract and concrete concepts) in the cross-attention layers of diffusion models by leveraging precomputed steering vectors, achieving controllable concept elimination;
Cat-PO: Cross-modal Adaptive Token-rewards for Preference Optimization in Truthful Multimodal LLMs
Zhixiao Zheng (University Of Science And Technology Of China), Zhendong Mao (University Of Science And Technology Of China)
CodeOptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: This paper proposes calculating visual-related rewards for each generated token in a multimodal large language model and embedding these rewards into the direct preference optimization (DPO) loss to finely reduce hallucinations and enhance model truthfulness.
Catching the Details: Self-Distilled RoI Predictors for Fine-Grained MLLM Perception
Yuheng Shi (University of Sydney), Chang Xu (University of Sydney)
CodeSegmentationKnowledge DistillationTransformerVision Language ModelMultimodality
🎯 What it does: Proposed and implemented SD-RPN, a self-distillation based RoI prediction framework, achieving high-precision, unlabeled fine-grained visual perception by generating pseudo labels from internal attention of MLLM.
CaTS: Calibrated Test-Time Scaling for Efficient LLM Reasoning
Chengsong Huang (Washington Univeristy in St. Louis), Jiaxin Huang (Washington Univeristy in St. Louis)
CodeComputational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This study proposes a self-calibrated dynamic sampling framework called CaTS, which adaptively controls multiple sampling methods such as Best-of-N and Self-Consistency during inference using reliable confidence scores obtained from training, thereby improving the inference accuracy of LLMs under the same sampling budget.
Causal-Steer: Disentangled Continuous Style Control without Parallel Corpora
Qingsong Wang (Zhejiang University), Jingyuan Chen (Zhejiang University)
CodeRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposes the Causal-Steer framework, which utilizes LoRA as a causal intervention tool to directly learn and extract pure style vectors from non-parallel data, achieving continuous, bidirectional, and multi-attribute style control of LLMs through modulation of activation layers.
🎯 What it does: Propose Celo2, a meta-learned optimizer with extremely low computational cost (only 4.5 GPU hours), which can transfer from small-scale 8×8 image classification tasks to large-scale pre-training (GPT-3 1.3B, ViT ImageNet) and reinforcement learning (Atari PPO) domains while maintaining stability and outperforming traditional AdamW and previous VeLO.
CERTIFIED VS. EMPIRICAL ADVERSARIAL ROBUSTNESS VIA HYBRID CONVOLUTIONS WITH ATTENTION STOCHASTICITY
Joy Dhar (Indian Institute of Technology Ropar), Nayyar Zaidi (Deakin University)
CodeAdversarial AttackConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: Propose a hybrid convolution and attention stochasticity (HyCAS) network, combining a 1-Lipschitz deterministic backbone with two internal randomization modules to achieve dual improvements in ℓ2 certificates and ℓ∞ empirical robustness.
🎯 What it does: Developed the Continuous Flow Operator (CFO) framework, which directly learns the time-right-hand side of PDEs using flow matching, enabling neural operator learning for continuous-time PDEs without requiring backpropagation through ODE solvers during training.
🎯 What it does: This paper proposes the entity tree-based retrieval-augmented generation algorithm CFT-RAG, which uses an improved Cuckoo Filter to accelerate the retrieval process of Tree-RAG.
🎯 What it does: In the source-agnostic domain adaptation object detection task, the CGSA framework is proposed, achieving object-level decomposition and semantic guidance on target domain images by integrating Hierarchical Slot Awareness and Class-Guided Slot Contrast into the DETR detector, thereby enhancing adaptation performance.
🎯 What it does: Propose a multi-task vehicle routing planning framework based on Chained Context Learning (CCL), which achieves progressive node state updates through dynamic constraint reshaping and trajectory-sharing node re-embedding.
🎯 What it does: Constructed the CHAMMI-75 multi-channel microscope image dataset and trained a general-purpose cell morphology model, MorphEm, using self-supervised learning.
🎯 What it does: Systematically study linear time series forecasting models, analyze the role of characteristic roots in noisy and noise-free scenarios, and propose two root reconstruction strategies.
Characterization and Learning of Causal Graphs with Latent Confounders and Post-treatment Selection from Interventional Data
Gongxu Luo (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
CodeBiomedical Data
🎯 What it does: Proposes a causal structure learning method addressing potential confounding variables and post-treatment selection bias, defining fine-grained interactive Markov equivalence classes (FI-Markov equivalence) and corresponding graphical representations (F-PAG). Subsequently, designs a complete and feasible algorithm F-FCI to simultaneously identify causal relationships, latent confounders, and selection bias from observational and interventional data.
Characterizing Deep Research: A Benchmark and Formal Definition
Abhinav Java (Microsoft Research), Amit Sharma (Microsoft Research)
CodeLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper formally defines the Deep Research (DR) task as requiring both high search volume and non-trivial reasoning, and proposes a statement-based evaluation framework along with the public benchmark LiveDRBench (100 scientific and current affairs-related queries) to objectively assess DR system performance.
Characterizing Human Semantic Navigation in Concept Production as Trajectories in Embedding Space
Felipe Diego Toro-Hernández (Federal University of ABC), Rodrigo M. Cabral-Carvalho (Federal University of ABC)
CodeRetrievalRepresentation LearningTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: View concept generation as a trajectory in the semantic embedding space, and utilize cumulative embeddings to calculate five kinematic and geometric metrics (distance to next, velocity, acceleration, entropy, distance to centroid) to characterize the human semantic retrieval process.
Characterizing the Discrete Geometry of ReLU Networks
Blake B. Gaines (University of Connecticut), Jinbo Bi (University of Connecticut)
CodeExplainability and InterpretabilityImageTabular
🎯 What it does: This paper studies the discrete geometric structure of polyhedral complexes defined by fully connected ReLU networks, providing theoretical upper bounds on the average degree and diameter of the connectivity graph, while proposing a BFS enumeration algorithm based on symbolic sequences to construct the complete complex;
ChartGalaxy: A Dataset for Infographic Chart Understanding and Generation
Zhen Li (Tsinghua University), Shixia Liu (Tsinghua University)
CodeGenerationData SynthesisRetrievalLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed a dataset named ChartGalaxy containing 2.1 million information charts (including 61,833 real and 1,701,356 synthetic charts), and conducted three tasks based on this dataset: chart understanding, code generation, and example-based generation;
Chasing the Tail: Effective Rubric-based Reward Modeling for Large Language Model Post-Training
Junkai Zhang (University of California, Los Angeles), Lifeng Jin (Scale AI, Inc.)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBiomedical DataBenchmarkFinance Related
🎯 What it does: This paper proposes a rubric-based reward model to address the problem of reward over-optimization during the late training of large language models, and improves the accuracy of the reward model in the high-value tail through an iterative differentiated approach.
ChatInject: Abusing Chat Templates for Prompt Injection in LLM Agents
Hwan Chang (Chung-Ang University), Hwanhee Lee (Chung-Ang University)
CodeAdversarial AttackLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the ChatInject attack, leveraging LLM chat templates and multi-round dialogue structures to achieve indirect prompt injection, bypassing instruction levels and security defenses.
🎯 What it does: Proposes the CheckWate framework, which embeds checkerboard watermarks during sampling in graph diffusion models and achieves polynomial-time verification.
ChemEval: A Multi-level and Fine-grained Chemical Capability Evaluation for Large Language Models
Yuqing Huang (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeLarge Language ModelPrompt EngineeringMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed and released ChemEval, a multi-level fine-grained evaluation benchmark covering 62 chemical tasks, including text, images, and spectrograms, addressing the gaps in existing benchmarks regarding chemical depth and multi-modal assessment.
Chessformer: A Unified Architecture for Chess Modeling
Daniel Monroe, Ashton Anderson
CodeExplainability and InterpretabilityKnowledge DistillationTransformerSequential
🎯 What it does: Propose Chessformer, a unified transformer architecture for board games, which enhances the strength of chess engines, accurately simulates moves of human players at different skill levels, and achieves model interpretability.
🎯 What it does: Propose the ChronoPlay framework to achieve automated, continuous generation of game RAG benchmarks, addressing dual dynamics of knowledge evolution and player interest drift, as well as authenticity issues.