ICLR 2026 Papers — Page 6
International Conference on Learning Representations · 5356 papers
Beyond Speedup - Utilizing KV Cache for Sampling and Reasoning
Zeyu XING (The Chinese University of Hong Kong), Sinno Jialin Pan (The Chinese University of Hong Kong)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper explores transforming KV cache in LLMs from a mere acceleration mechanism into a low-cost reusable representation for chain-of-embedding self-evaluation and adaptive fast-slow thinking switching.
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)
Data 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.
Beyond Structure: Invariant Crystal Property Prediction with Pseudo-Particle Ray Diffraction
Bin Cao (Guangzhou Municipal Key Laboratory of Materials Informatics, Advanced Materials Thrust), Tongyi ZHANG
Graph Neural NetworkGraphPhysics Related
🎯 What it does: Propose PRDNet, combining graph neural networks with a learnable pseudo-particle diffraction module to achieve high-precision prediction of crystal structural properties.
Beyond Student: An Asymmetric Network for Neural Network Inheritance
Yiyun Zhou (Zhejiang University), Jingyuan Chen (Zhejiang University)
Knowledge 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)
RetrievalLarge 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
GenerationData 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.
Beyond the Heatmap: A Rigorous Evaluation of Component Impact in MCTS-Based TSP Solvers
Xuanhao Pan (Chinese University of Hong Kong), Tianshu Yu (Chinese University of Hong Kong)
OptimizationHyperparameter SearchGraphBenchmark
🎯 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.
Beyond the Known: An Unknown-Aware Large Language Model for Open-Set Text Classification
Xi Chen (University of Science and Technology of China), Hui Xiong (Hong Kong University of Science and Technology)
ClassificationLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Developed UnLLM, an unknown-aware open-set text classification framework based on large language models.
Beyond Uniformity: Regularizing Implicit Neural Representations through a Lipschitz Lens
Julian McGinnis (Technical University of Munich), Benedikt Wiestler (Technical University of Munich)
RestorationSuper ResolutionImageMeshBiomedical DataComputed Tomography
🎯 What it does: Proposes an adjustable Lipschitz budget framework that regularizes implicit neural representations (INR) using task-related budgets and non-uniform allocation, validated on tasks such as shape representation, lung deformation registration, and image inpainting.
Beyond Uniformity: Sample and Frequency Meta Weighting for Post-Training Quantization of Diffusion Models
Cuong Pham (Monash University), Thanh-Toan Do (Monash University)
GenerationComputational EfficiencyMeta LearningDiffusion modelImage
🎯 What it does: Propose a post-training quantization (PTQ) method based on meta-learning, which jointly optimizes the noise estimation network of diffusion models using sample weights and frequency weights.
Beyond URLs: Metadata Diversity and Position for Efficient LLM Pretraining
Dongyang Fan (EPFL), Martin Jaggi (EPFL)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Investigate the acceleration effects of various document metadata (such as URL, fine-grained quality scores, fine-grained domain information, etc.) in LLM pre-training;
Beyond Visual Reconstruction Quality: Object Perception-aware 3D Gaussian Splatting for Autonomous Driving
Renzhi Wang (University of Alberta), Qing Guo (NKIARI)
Object DetectionAutonomous DrivingConvolutional Neural NetworkGaussian SplattingImage
🎯 What it does: Propose a 3D Gaussian dispersion reconstruction method oriented towards perceptual stability, enabling the reconstructed images in autonomous driving scenarios to maintain visual quality while ensuring consistent outputs from the same perception model on original and reconstructed images.
BeyondBench: Contamination-Resistant Evaluation of Reasoning in Language Models
Gaurav Srivastava (Virginia Tech), Xuan Wang (Virginia Tech)
Data-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;
BézierFlow: Learning Bézier Stochastic Interpolant Schedulers for Few-Step Generation
Yunhong Min (Korea Advanced Institute Of Science And Technology), Minhyuk Sung (Korea Advanced Institute Of Science And Technology)
GenerationDiffusion modelFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Proposed a lightweight training framework called B'ezierFlow, which utilizes a random interpolation scheduler (SI Scheduler) parameterized by Bézier curves to learn the optimal sampling trajectory of pre-trained diffusion and flow models, significantly improving generation quality within only a few steps (≤10 NFE).
BFM-Zero: A Promptable Behavioral Foundation Model for Humanoid Control Using Unsupervised Reinforcement Learning
Yitang Li (Tsinghua University), Guanya Shi (Carnegie Mellon University)
Robotic IntelligenceReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: This paper proposes and implements a behavior foundation model called BFM‑Zero based on unsupervised reinforcement learning, which can accomplish various whole-body control tasks (motion tracking, reward optimization, target reaching, etc.) on the real Unitree G1 humanoid robot without retraining or planning;
Bi-Criteria Metric Distortion
Kiarash Banihashem (University of Maryland), Max Springer (Princeton University)
Optimization
🎯 What it does: The study selects a fixed number of candidates in multi-voting to achieve optimal or approximate representative selection by leveraging order information;
Bi-directional Bias Attribution: Debiasing Large Language Models without Modifying Prompts
Yujie Lin (Xiamen University), Jinsong Su (Xiamen University)
Explainability and InterpretabilityLarge Language ModelTextBenchmark
🎯 What it does: In large language models, first automatically discover adjective/noun terms that can induce bias (i.e., stereotype cues), then use two attribution methods based on integrated gradients—forward and backward—to locate bias-causing neurons in the projection layer, and intervene by fixing the activation values of these neurons to a constant, thereby achieving bias mitigation; the entire process does not require fine-tuning the model or modifying user prompts;
Bi-Lipschitz Autoencoder With Injectivity Guarantee
Qipeng Zhan (University of Pennsylvania), Li Shen (University of Pennsylvania)
Representation 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).
Bi-LoRA: Efficient Sharpness-Aware Minimization for Fine-Tuning Large-Scale Models
Yuhang Liu (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)
OptimizationComputational 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)
Safty 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)
Data 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)
TransformerLarge 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);
BiasScope: Towards Automated Detection of Bias in LLM-as-a-Judge Evaluation
Peng Lai (Southern University of Science and Technology), Guanhua Chen (Southern University of Science and Technology)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the BIASSCOPE framework to automatically discover potential biases in the LLM-as-a-Judge evaluation process, and construct a more challenging JudgeBench-Pro benchmark based on this;
BideDPO: Conditional Image Generation with Simultaneous Text and Condition Alignment
Dewei Zhou, Yi Yang (Zhejiang University)
GenerationVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose a bidirectional decoupled Direct Preference Optimization (BideDPO) framework to simultaneously satisfy text and conditional inputs when conflicts occur between text and conditions, enhancing controllability in conditional image generation.
Bidirectional Predictive Coding
Gaspard Oliviers (University of Oxford), Rafal Bogacz (University of Oxford)
ClassificationGenerationImageMultimodality
🎯 What it does: Proposed and implemented a bidirectional predictive coding (bPC) model that can perform both generative and discriminative reasoning within the same network architecture, and conducted systematic evaluations on supervised, unsupervised, and hybrid tasks.
BigMaQ: A Big Macaque Motion and Animation Dataset Bridging Image and 3D Pose Representations
Lucas Martini (University of Tbingen), Martin A. Giese (University of Tbingen)
RecognitionPose EstimationTransformerVideoMeshBenchmark
🎯 What it does: This paper constructs the BigMaQ dataset, providing multi-view videos of macaques in over 750 scenarios along with corresponding 3D skeletal joint rotations and mesh surface models.
Bilateral Information-aware Test-time Adaptation for Vision-Language Models
Jingwei Sun (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Domain AdaptationTransformerSupervised Fine-TuningContrastive LearningImageText
🎯 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.
Bilevel Optimization with Lower-Level Uniform Convexity: Theory and Algorithm
Yuman Wu (George Mason University), Mingrui Liu (George Mason University)
OptimizationText
🎯 What it does: This paper proposes a new two-layer optimization method, conducting theoretical analysis and algorithm design for the uniform convexity of the lower-level function, especially developing the UniBiO algorithm to solve two-layer optimization problems under uniform convexity conditions.
Bilinear representation mitigates reversal curse and enables consistent model editing
Dong-Kyum Kim (MPI-SP), Meeyoung Cha (MPI-SP)
Representation LearningTransformerLarge Language ModelGraph
🎯 What it does: Investigating how language models overcome the curse of inversion and achieve consistent model editing by learning bilinear relationship structures.
BindWeave: Subject-Consistent Video Generation via Cross-Modal Integration
Zhaoyang Li (University of Science and Technology of China), Zehuan Yuan (ByteDance)
GenerationTransformerLarge Language ModelVision Language ModelDiffusion modelAuto EncoderImageVideoTextMultimodality
🎯 What it does: Propose the BindWeave framework to achieve high consistency between agent identities and actions during multi-agent video generation;
Binomial Gradient-Based Meta-Learning for Enhanced Meta-Gradient Estimation
Yilang Zhang (University of Minnesota), Georgios B. Giannakis (University of Minnesota)
ClassificationOptimizationMeta LearningImage
🎯 What it does: Propose BinomGBML, a gradient estimation method based on binomial expansion, to improve the accuracy of GBML's meta-gradient estimation and reduce computational complexity.
BioBO: Biology-informed Bayesian Optimization for Perturbation Design
Yanke Li (Johnson & Johnson Innovative Medicine), Rui Liao (Johnson & Johnson Innovative Medicine)
OptimizationExplainability and InterpretabilityBiomedical Data
🎯 What it does: Propose the BioBO framework, integrating multi-modal gene embeddings with enrichment analysis priors into Bayesian optimization for efficient design of gene knockout experiments.
BioCAP: Exploiting Synthetic Captions Beyond Labels in Biological Foundation Models
Ziheng Zhang (Ohio State University), Jianyang Gu (Ohio State University)
ClassificationRetrievalExplainability 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.
Biologically Plausible Learning via Bidirectional Spike-Based Distillation
Yifei Wang (Fudan University), Xiaoqing Zheng (Fudan University)
Explainability and InterpretabilityKnowledge DistillationRecurrent Neural NetworkSpiking Neural NetworkAuto EncoderContrastive LearningImageTextTime Series
🎯 What it does: Proposed and implemented Bidirectional Spike-based Distillation (BSD), a biologically interpretable neural network learning algorithm that jointly trains and mutually distills features through forward and backward synaptic networks.
BioMD: All-atom Generative Model for Biomolecular Dynamics Simulation
Bin Feng, Yu Li (International Digital Economy Academy)
Drug DiscoveryGraph Neural NetworkTransformerFlow-based ModelBiomedical Data
🎯 What it does: Proposes BioMD, a full-atom generative model based on hierarchical prediction and interpolation, for simulating long-timescale protein-ligand dynamic trajectories.
BioTamperNet: Affinity-Guided State-Space Model Detecting Tampered Biomedical Images
Soumyaroop Nandi (University of Southern California), Prem Natarajan (University of Southern California)
Anomaly 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.
BioX-Bridge: Model Bridging for Unsupervised Cross-Modal Knowledge Transfer across Biosignals
Chenqi Li (University of Oxford), Tingting Zhu (University of Oxford)
ClassificationDomain AdaptationTime SeriesBiomedical DataElectrocardiogram
🎯 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.
Birch SGD: A Tree Graph Framework for Local and Asynchronous SGD Methods
Alexander Tyurin (Applied AI Institute), Danil Sivtsov (Applied AI Institute)
OptimizationImageText
🎯 What it does: Propose the Birch SGD framework, modeling distributed SGD methods as weighted directed trees and leveraging the geometric properties of trees for unified analysis, based on which at least six out of eight new methods are proven to achieve optimal computational time;
BIRD-INTERACT: Re-imagining Text-to-SQL Evaluation via Lens of Dynamic Interactions
Nan Huo (University of Hong Kong), Reynold Cheng (University of Hong Kong)
AI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Constructed a text-to-SQL benchmark named BIRD-INTERACT that covers the full spectrum of CRUD operations and supports multi-turn interactions.
BIRD: Behavior Induction via Representation-structure Distillation
Galen Pogoncheff (University of California, Santa Barbara), Michael Beyeler (University of California, Santa Barbara)
Safty and PrivacyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackContrastive LearningImageText
🎯 What it does: The BIRD framework is proposed by aligning the internal representation structure of the teacher model, transferring alignment behaviors (e.g., robustness, safety) from small models to large models.
Bird's-eye-view Informed Reasoning Driver
Yinuo Wang (Tsinghua University), Siyuan Cheng (Huawei Inc.)
Autonomous DrivingSupervised Fine-TuningVision Language ModelImage
🎯 What it does: Propose the BIRDriver framework, integrating VLM with motion planners to generate keypoint-guided trajectory planning using single-frame bird's-eye view (BEV).
Black-Box Privacy Attacks on Shared Representations in Multitask Learning
John Abascal (Northeastern University), Matthew Jagielski (Google Deepmind)
Safty 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.
BLADE: Block-Sparse Attention Meets Step Distillation for Efficient Video Generation
Youping Gu (Zhejiang University), Bohan Zhuang (Zhejiang University)
GenerationComputational EfficiencyKnowledge DistillationTransformerVideoText
🎯 What it does: This paper proposes an efficient video generation framework called BLADE based on sparse attention and data-free distillation.
Block Recurrent Dynamics in Vision Transformers
Mozes Jacobs (Harvard University), T. Anderson Keller
ClassificationSegmentationDepth EstimationCompressionTransformerImage
🎯 What it does: This paper proposes Block-Recurrent Hypothesis (BRH), suggesting that the depth of a trained ViT can be divided into a few adjacent stages (blocks), where each stage repeatedly applies a small number of parameter-shared blocks to reconstruct the internal representations of the full network; and constructs the Raptor model (Recurrent Approximations to Phase-structured TransfORmers) to implement this rewriting.
Block-Sample MAC-Bayes Generalization Bounds
Matthias Frey (University of Melbourne), Michael Gastpar
🎯 What it does: This paper proposes a new Block-Sample MAC-Bayes Bound, which bounds the expected generalization error by dividing the training set into multiple blocks and using the KL divergence between the conditional posterior and prior of each block.
Block-wise Adaptive Caching for Accelerating Diffusion Policy
Kangye Ji (Tsinghua University), Zhi Wang (Tsinghua University)
Computational EfficiencyRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelImageVideoMultimodality
🎯 What it does: Provide a training-free acceleration plugin for diffusion policies based on Transformers
BOAD: Discovering Hierarchical Software Engineering Agents via Bandit Optimization
Iris Xu (Massachusetts Institute Of Technology), Zhang-Wei Hong (Mit-Ibm Watson Ai Lab)
OptimizationAI Code AssistantLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Discover hierarchical multi-agent systems through automated search to address GitHub software engineering problems
BoGrape: Bayesian optimization over graphs with shortest-path encoded
Yilin Xie (Imperial College London), Calvin Tsay (Imperial College London)
OptimizationDrug DiscoveryGraphBiomedical Data
🎯 What it does: Propose a graph structure optimization framework BoGrape based on Bayesian optimization, achieving global acquisition function optimization through the shortest path graph kernel and mixed-integer programming (MIP);
BOLT: Decision‑Aligned Distillation and Budget-Aware Routing for Constrained Multimodal QA on Robots
Tengjun Ni (University of Technology Sydney), Wenjie Zhang (University of New South Wales)
Knowledge DistillationRobotic IntelligenceTransformerSupervised Fine-TuningMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 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.
Bongard-RWR+: Real-World Representations of Fine-Grained Concepts in Bongard Problems
Szymon Pawlonka (Warsaw University of Technology), Jacek Mańdziuk (Warsaw University of Technology)
Data SynthesisTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes the Bongard-RWR+ dataset, leveraging photorealistic images generated by VLMs to realize abstract concepts in Bongard problems, significantly enhancing the data scale;
Boolean Satisfiability via Imitation Learning
Zewei Zhang (McMaster University), Xiangyu Xu (Xi'an Jiaotong University)
OptimizationTransformerBenchmark
🎯 What it does: Designed a branching strategy called ImitSAT for CDCL SAT solvers based on imitation learning, trained by refining expert KeyTrace sequences.
Boomerang Distillation Enables Zero-Shot Model Size Interpolation
Sara Kangaslahti (Harvard University), David Alvarez-Melis (Harvard University)
Computational EfficiencyKnowledge DistillationTransformerText
🎯 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;
Boosted Trees on a Diet: Compact Models for Resource-Constrained Devices
Nina Herrmann (University of Munster), Fabian Gieseke (University of Munster)
Computational EfficiencyTabular
🎯 What it does: Implementing a lightweight boosting decision tree model on resource-constrained IoT devices
Boosting Entropy with Bell Box Quantization
Ningfeng Yang (University of British Columbia), Tor M. Aamodt (University of British Columbia)
Computational 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 for Predictive Sufficiency
Abbavaram Gowtham Reddy (Cispa Helmholtz Center For Information Security), Rebekka Burkholz (University Of Amsterdam)
Domain AdaptationExplainability and InterpretabilityTabular
🎯 What it does: Studied the out-of-distribution (OOD) generalization performance of boosting methods under hidden confounding shift environments, and proposed an information-theoretic metric, α-predictive sufficiency, to explain their effectiveness.
Boosting Medical Visual Understanding From Multi-Granular Language Learning
Zihan Li (University of Washington), Paul Kinahan (University of Washington)
ClassificationRecognitionTransformerVision 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
OptimizationTransformerLarge 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.
Boosting Open Set Recognition Performance through Modulated Representation Learning
Amit Kumar Kundu (University of Maryland), Joseph JaJa (University of Maryland)
RecognitionRepresentation LearningContrastive LearningImage
🎯 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.
Bootstrapping MLLM for Weakly‑Supervised Class‑Agnostic Object Counting
Xiaowen Zhang (Tongji University), Miaojing Shi (Tongji University)
Object DetectionLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Proposes WS-COC, a weakly supervised, class-agnostic object counting framework based on multimodal large language models.
BoRA: Towards More Expressive Low-Rank Adaptation with Block Diversity
Shiwei Li (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
Computational EfficiencyRepresentation LearningText
🎯 What it does: Proposed a new low-rank adaptation method called BoRA, which enhances the expressiveness of LoRA weights by introducing block diversity.
BoreaRL: A Multi-Objective Reinforcement Learning Environment for Climate-Adaptive Boreal Forest Management
Kevin Bradley Dsouza (University of Waterloo), Yuri Leonenko (University of Waterloo)
OptimizationReinforcement LearningTime Series
🎯 What it does: Developed BoreaRL, a multi-objective reinforcement learning environment for climate-adaptive coniferous forest management;
BOTS: A Unified Framework for Bayesian Online Task Selection in LLM Reinforcement Finetuning
Qianli Shen (Alibaba Group), Jingren Zhou (Alibaba Group)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes an online task selection framework called BOTS for reinforcement fine-tuning of large language models.
Bottlenecked Transformers: Periodic KV Cache Consolidation for Generalised Reasoning
Adnan Oomerjee (University College London), Jun Wang (University College London)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a mechanism called Bottlenecked Transformer, which periodically rewrites the KV cache during Transformer LLM inference. By merging and reorganizing memory through a cache processor after each thinking step, the model enhances its inference capabilities.
Bound by semanticity: universal laws governing the generalization-identification tradeoff
Marco Nurisso (Politecnico di Torino), Giovanni Petri (Northeastern University)
ClassificationRecognitionConvolutional 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)
Explainability 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.
Bradley-Terry and Multi-Objective Reward Modeling Are Complementary
Zhiwei Zhang (Pennsylvania State University), Suhang Wang (Pennsylvania State University)
OptimizationRepresentation LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningTextBenchmark
🎯 What it does: Proposes the SMORM framework, which jointly trains Bradley-Terry single-objective and multi-objective reward models, enhancing the robustness and resistance to reward hacking of reward models in out-of-distribution (OOD) scenarios.
Brain-IT: Image Reconstruction from fMRI via Brain-Interaction Transformer
Roman Beliy (Weizmann Institute of Science), michal Irani
RestorationTransformerDiffusion modelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: The paper proposes a Brain-IT framework for fMRI image reconstruction, which utilizes clustering of brain functional regions and a Transformer model to directly generate local image features, combined with a low-level DIP branch and a high-level diffusion branch to achieve high-fidelity image reconstruction.
Brain-Semantoks: Learning Semantic Tokens of Brain Dynamics with a Self-Distilled Foundation Model
Sam Gijsen (University of Tübingen), Kerstin Ritter (University of Tübingen)
Knowledge DistillationRepresentation LearningTransformerBiomedical DataMagnetic Resonance Imaging
🎯 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.
Branch and Bound Search for Exact MAP Inference in Credal Networks
Radu Marinescu (IBM Research), Alexander G. Gray
OptimizationGraph
🎯 What it does: This paper proposes a depth-first branch-and-bound search algorithm for exact MAP inference, capable of solving maximax and maximin MAP problems in exact networks.
Branched Schrödinger Bridge Matching
Sophia Tang (University of Pennsylvania), Pranam Chatterjee (University of Pennsylvania)
OptimizationPoint CloudBiomedical DataPhysics RelatedStochastic Differential Equation
🎯 What it does: Designed and implemented Branched Schrödinger Bridge Matching (BranchSBM) to learn branching trajectories from a single initial distribution to multiple target distributions.
BranchGRPO: Stable and Efficient GRPO with Structured Branching in Diffusion Models
Yuming Li (Peking University), Shanghang Zhang (Peking University)
GenerationReinforcement LearningDiffusion modelImageVideoStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose BranchGRPO, a tree-structured GRPO training framework that reconstructs the inference process of diffusion models through branch replay;
Breaking Agent Backbones: Evaluating the Security of Backbone LLMs in AI Agents
Julia Bazinska (Lakera AI), Niklas Pfister (Lakera AI)
Safty and PrivacyTransformerLarge Language ModelAgentic AIBenchmark
🎯 What it does: Propose the 'threat snapshots' framework and the b3 benchmark for systematic evaluation of the security of AI agent backbone LLMs.
Breaking and Fixing Defenses Against Control Flow Hijacking in Multi-Agent Systems
Rishi Dev Jha, Vitaly Shmatikov (Cornell University)
Anomaly DetectionTransformerLarge Language ModelGraphBenchmark
🎯 What it does: This paper studies control flow hijacking attacks in multi-agent systems and proposes the CONTROLVALVE defense scheme
Breaking Barriers: Do Reinforcement Post Training Gains Transfer To Unseen Domains?
Chuxuan Hu (University of Illinois at Urbana-Champaign), Daniel Kang (University of Illinois at Urbana-Champaign)
Domain AdaptationTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: A two-phase evaluation of the transfer performance of reinforced post-training (RPT) for large language models across different domains was conducted. The study systematically compared the performance of 18 publicly available RPT models and their baseline models on multi-domain benchmarks, and tested their cross-domain transferability through single-domain RPT experiments.
Breaking Gradient Temporal Collinearity for Robust Spiking Neural Networks
Desong Zhang (University of Exeter), Geyong Min (University of Exeter)
ClassificationAdversarial AttackSpiking Neural NetworkImageSequential
🎯 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).
Breaking Safety Paradox with Feasible Dual Policy Iteration
Yujie Yang (Tsinghua University), Shengbo Eben Li (Tsinghua University)
Safty and PrivacyReinforcement LearningBenchmark
🎯 What it does: This paper first identifies the 'safety paradox' in safe reinforcement learning, and then proposes a Feasible Dual Policy Iteration (FDPI) algorithm. By training an auxiliary policy that maximizes constraint violations to generate more violation samples, and using importance sampling to correct data distribution, the algorithm significantly improves safety and reward.
Breaking Scale Anchoring: Frequency Representation Learning for Accurate High-Resolution Inference from Low-Resolution Training
Wenshuo Wang, Fan Zhang (South China University of Technology)
Super ResolutionRepresentation LearningConvolutional Neural NetworkTransformerTime SeriesPhysics Related
🎯 What it does: Investigated the scale anchoring problem in zero-shot super-resolution spatiotemporal prediction, and proposed a frequency representation learning (FRL) method compatible with multiple networks to alleviate this issue.
Breaking the Correlation Plateau: On the Optimization and Capacity Limits of Attention-Based Regressors
Jingquan Yan (University of Texas at Arlington), Junzhou Huang (University of Texas at Arlington)
OptimizationTransformerMultimodalityTabular
🎯 What it does: Investigated the phenomenon of PCC plateau in attention-based regression models under joint MSE+PCC training, and provided theoretical analysis and improvement methods
Breaking the SFT Plateau: Multimodal Structured Reinforcement Learning for Chart-to-Code Generation
Lei Chen, Lin Ma
AI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality
🎯 What it does: This paper constructs a 3M chart-code pair dataset and proposes a multi-modal structured reinforcement learning (MSRL) method, significantly breaking through the performance bottlenecks of SFT and improving the accuracy of chart code generation.
Breaking the Total Variance Barrier: Sharp Sample Complexity for Linear Heteroscedastic Bandits with Fixed Action Set
Heyang Zhao (University of California, Los Angeles), Quanquan Gu (University of California, Los Angeles)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the linear heteroscedastic noise multi-armed bandit problem under a fixed action set, proposes two variance-adaptive exploration algorithms (VAEE and VAGD), and provides upper and lower bounds on their simple regret.
BRIDGE: Bi-level Reinforcement Learning for Dynamic Group Structure in Coalition Formation Games
Shuqing Shi (King's College London), Yali Du (King's College London)
OptimizationReinforcement LearningBenchmark
🎯 What it does: Proposed a two-layer reinforcement learning framework called BRIDGE based on deep reinforcement learning for coalition formation games with dynamic group structures, aiming to efficiently explore potential coalition structures to achieve optimal allocation.
BridgeDrive: Diffusion Bridge Policy for Closed-Loop Trajectory Planning in Autonomous Driving
Shu Liu (Bosch (China) Investment Ltd), Hao Yang (Bosch (China) Investment Ltd)
Autonomous DrivingTransformerDiffusion modelOrdinary Differential Equation
🎯 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 Degradation Discrimination and Generation for Universal Image Restoration
JiaKui Hu (Peking University), Yanye Lu (Peking University)
RestorationConvolutional Neural NetworkDiffusion modelContrastive LearningImage
🎯 What it does: Designed the BDG framework, integrating degradation discrimination with diffusion generation to achieve universal image restoration.
Bridging Draft Policy Misalignment: Group Tree Optimization for Speculative Decoding
Shijing Hu, Pan Zhou
OptimizationComputational 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 Explainability and Embeddings: BEE Aware of Spuriousness
Cristian Daniel Paduraru (Bitdefender), Elena Burceanu (Bitdefender)
Explainability and InterpretabilityLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the BEE framework, which detects and names pseudo correlations learned during training by leveraging the model weight space and embedding geometry;
Bridging Fairness and Explainability: Can Input-Based Explanations Promote Fairness in Hate Speech Detection?
Yifan Wang (Saarland University), Vera Demberg (Saarland University)
Explainability 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.
Bridging Generalization Gap of Heterogeneous Federated Clients Using Generative Models
Ziru Niu (RMIT University), A. K. Qin (Swinburne University of Technology)
GenerationData SynthesisFederated LearningConvolutional Neural NetworkImage
🎯 What it does: Propose the FedVTC framework, which uses variational transposed convolution to generate synthetic data, enhancing generalization for model-heterogeneous FL clients without requiring public data.
Bridging Input Feature Spaces Towards Graph Foundation Models
Moshe Eliasof (University of Cambridge), Carola-Bibiane Schönlieb (University of Cambridge)
Domain AdaptationRepresentation LearningDrug DiscoveryGraph Neural NetworkImageGraphBiomedical Data
🎯 What it does: Developed a framework called ALL-IN, which uses random projection to map arbitrary-dimensional, semantically inconsistent node features into a shared space, and calculates node covariance to construct input feature-agnostic graph representations, enabling cross-dataset transfer.
Bridging Kolmogorov Complexity and Deep Learning: Asymptotically Optimal Description Length Objectives for Transformers
Peter Shaw (Google DeepMind), Kristina Toutanova (Google DeepMind)
CompressionComputational EfficiencyRepresentation LearningTransformerSequential
🎯 What it does: This paper proposes an 'asymptotically optimal description length' objective based on Kolmogorov complexity and MDL principles, proving that Transformers can achieve this optimal compression when resources approach infinity. Subsequently, a differentiable variational objective (based on an adaptive Gaussian Mixture Prior) is designed to realize this theory. Experiments on a synthetic parity task demonstrate that this objective can induce models with low complexity and strong generalization, but optimization with random initialization struggles to converge.
Bridging ML and algorithms: comparison of hyperbolic embeddings
Dorota Celińska-Kopczyńska (University of Warsaw), Eryk Kopczyński (University of Warsaw)
Representation LearningGraph
🎯 What it does: This paper conducts a systematic experimental comparison of hyperplane embedding methods in the fields of machine learning, network theory, and algorithms, focusing on evaluating the performance and quality of the BFKL algorithm by Bläsius et al. and Lorentz embeddings.
Bridging Past and Future: Distribution-Aware Alignment for Time Series Forecasting
Yifan Hu (Tsinghua University), Liang Sun (Alibaba Group)
Time Series
🎯 What it does: Proposed a time series prediction framework based on prediction-reconstruction alignment called TimeAlign;
Bridging Piano Transcription and Rendering via Disentangled Score Content and Style
Wei Zeng (National University of Singapore), Ye Wang (National University of Singapore)
RecognitionGenerationRepresentation LearningTransformerDiffusion modelTextMultimodalityAudio
🎯 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 Radiology and Pathology Foundation Models via Concept-Based Multimodal Co-Adaptation
Yihang Chen (University of Hong Kong), Lequan Yu (University of Hong Kong)
Domain AdaptationExplainability and InterpretabilitySupervised Fine-TuningPrompt EngineeringContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Propose a cross-modal collaborative adaptation framework CTF based on medical concepts, utilizing clinical concepts to dynamically align and fuse between radiology and pathology models.
Bridging Successor Measure and Online Policy Learning with Flow Matching-Based Representations
Haosen Shi (Chinese University of Hong Kong), Sinno Jialin Pan (Ant International)
Reinforcement 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 Distribution Gap to Harness Pretrained Diffusion Priors for Super-Resolution
JoonKyu Park, Kyoung Mu Lee (Seoul National University)
Super ResolutionDiffusion modelAuto EncoderImage
🎯 What it does: Proposed a framework called DM-SR for super-resolution using pre-trained diffusion models, achieving single-step high-quality super-resolution by training only a single image encoder.
Bridging the Gap Between Promise and Performance for Microscaling FP4 Quantization
Vage Egiazarian (Institute of Science and Technology Austria), Dan Alistarh (Institute of Science and Technology Austria)
Computational EfficiencyTransformerText
🎯 What it does: This paper investigates the practicality of MXFP4 and NVFP4 micro-scale FP4 quantization formats in large language models and proposes a micro-rotation GPTQ (MR-GPTQ) algorithm specifically tailored for these two formats.
Bridging the performance-gap between target-free and target-based reinforcement learning
Théo Vincent (DFKI), Carlo D'Eramo (University of Würzburg)
Reinforcement 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
GenerationDiffusion 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)
RetrievalLarge 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.
BTZSC: A Benchmark for Zero-Shot Text Classification Across Cross-Encoders, Embedding Models, and Rerankers
Ilias Aarab (European Central Bank)
ClassificationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the BTZSC benchmark to uniformly evaluate the zero-shot performance of NLI cross-encoders, text embedding models, re-rankers, and instruction-tuned LLMs across 22 publicly available text classification datasets.