ICLR 2025 Papers — Page 4
International Conference on Learning Representations · 3704 papers
Autonomous Evaluation of LLMs for Truth Maintenance and Reasoning Tasks
Rushang Karia (Arizona State University), Siddharth Srivastava (Arizona State University)
Large Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the AutoEval framework, which utilizes grammar generation and formal equivalence verification to achieve human-free evaluation of LLM in truth maintenance and reasoning tasks.
Autoregressive Pretraining with Mamba in Vision
Sucheng Ren (Johns Hopkins University), Cihang Xie (UC Santa Cruz)
Object DetectionSegmentationTransformerContrastive LearningImage
🎯 What it does: This paper proposes a training method for visual models based on autoregressive pre-training—ARM—and applies it to the Mamba architecture, significantly enhancing Mamba's performance and scalability in large-scale visual tasks.
Autoregressive Video Generation without Vector Quantization
Haoge Deng (Beijing University of Posts and Telecommunications), Xinlong Wang (Beijing Academy of Artificial Intelligence)
GenerationData SynthesisTransformerOptical FlowImageVideoText
🎯 What it does: A non-quantized autoregressive video and image generation model named NOVA is proposed, capable of performing various tasks such as text-to-image, text-to-video, and image-to-video within the same framework.
AutoUAD: Hyper-parameter Optimization for Unsupervised Anomaly Detection
Wei Dai (Chinese University of Hong Kong), Jicong Fan (Chinese University of Hong Kong)
Anomaly DetectionOptimizationHyperparameter SearchAuto EncoderTabular
🎯 What it does: This paper addresses the challenges of hyperparameter tuning and model selection in unsupervised anomaly detection (UAD) by proposing three internal evaluation metrics: Relative Top Median (RTM), Expected Anomaly Gap (EAG), and Normalized Pseudo-Difference (NPD), and utilizes Bayesian optimization to automatically search for the best hyperparameters and models.
AvatarGO: Zero-shot 4D Human-Object Interaction Generation and Animation
Yukang Cao (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisOptimizationLarge Language ModelDiffusion modelGaussian SplattingImageText
🎯 What it does: Generating and animating 4D full-body human-computer interaction scenes through text prompts
AVHBench: A Cross-Modal Hallucination Benchmark for Audio-Visual Large Language Models
Kim Sung-Bin (POSTECH), Tae-Hyun Oh (KAIST)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodalityBenchmarkAudio
🎯 What it does: This paper proposes AVHBench, a benchmark specifically designed to evaluate the cross-modal hallucinations of audio-visual large language models (AV-LLM). It constructs a dataset of 2,136 videos containing 5,302 question-answer pairs and 1,106 multimodal descriptions through a semi-automated annotation process.
B-STaR: Monitoring and Balancing Exploration and Exploitation in Self-Taught Reasoners
Weihao Zeng (Hong Kong University of Science and Technology), Junxian He (Hong Kong University of Science and Technology)
TransformerSupervised Fine-TuningReinforcement LearningTextMultimodality
🎯 What it does: The B-STAR framework is proposed to automatically monitor and balance the two main factors of exploration (diversity) and exploitation (reward discrimination) during the self-learning process, thereby enhancing the performance of self-improving models.
BaB-ND: Long-Horizon Motion Planning with Branch-and-Bound and Neural Dynamics
Keyi Shen (University of Illinois Urbana-Champaign), Yunzhu Li (Columbia University)
OptimizationRobotic IntelligenceGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A GPU-accelerated branch-and-bound framework is proposed for long-term motion planning based on neural network dynamics models.
Backdooring Vision-Language Models with Out-Of-Distribution Data
Weimin Lyu (Stony Brook University), Chao Chen (Stony Brook University)
GenerationKnowledge DistillationAdversarial AttackTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Conducting backdoor attacks on visual-language models (VLM) by utilizing only out-of-domain (OOD) data available during training to attack image-to-text generation tasks.
Backtracking Improves Generation Safety
Yiming Zhang (Carnegie Mellon University), Eric Michael Smith (Meta)
GenerationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A 'backtracking' mechanism is proposed, allowing the language model to revoke generated content and regenerate safe responses by inserting a special token [RESET] after generating unsafe text.
Bad-PFL: Exploiting Backdoor Attacks against Personalized Federated Learning
Mingyuan Fan (East China Normal University), Cen Chen (East China Normal University)
Federated LearningAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A new attack method named Bad-PFL is designed for backdoor attacks in personalized federated learning (PFL), which implants backdoors using natural feature triggers during client training and dynamically generates triggers through a generator to ensure their persistent presence in personalized models.
BadJudge: Backdoor Vulnerabilities of LLM-As-A-Judge
Terry Tong (University of California), Muhao Chen (University of California)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: The study demonstrates a backdoor attack mechanism targeting the LLM-as-a-Judge automatic evaluation system, proving that even contaminating just 1% of the training data can significantly enhance the evaluation scores of the attacker's model.
BadRobot: Jailbreaking Embodied LLM Agents in the Physical World
Hangtao Zhang (Huazhong University of Science and Technology), Leo Yu Zhang (Griffith University)
Adversarial AttackRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes the BADROBOT attack paradigm, demonstrating how to jailbreak embedded LLM robots in a no-box manner, enticing them to perform harmful physical actions.
Balanced Neural ODEs: nonlinear model order reduction and Koopman operator approximations
Julius Aka (University of Augsburg), Lars Mikelsons (University of Augsburg)
Auto EncoderTime SeriesOrdinary Differential Equation
🎯 What it does: A fast proxy model with adjustable complexity is constructed by combining VAE and Neural ODE, achieving model reduction without prior dimensionality and approximating the Koopman operator;
Balanced Ranking with Relative Centrality: A multi-core periphery perspective
Chandra Sekhar Mukherjee (University of Southern California), Jiapeng Zhang (University of Southern California)
Graph Neural NetworkGraphBiomedical Data
🎯 What it does: Proposes and implements an unsupervised vertex ranking method based on relative centrality, primarily achieving ranking balance through local normalization of traditional centrality metrics;
Balancing Act: Diversity and Consistency in Large Language Model Ensembles
Ahmed Abdulaal (UCL), Amrutha Saseendran (AstraZeneca)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper systematically studies the trade-off between diversity and consistency in the integration of large language models (LLMs) and proposes a unified framework and Dynamic Mixture of Agents (DMoA) strategy;
Balancing Bias in Two-sided Markets for Fair Stable Matchings
Siyuan Wu (University of Macau), Panagiotis Karras (University of Copenhagen and Aarhus University)
Optimization
🎯 What it does: An algorithm called ISORROPIA is proposed to efficiently solve the Balanced Stable Marriage (BSM) problem in practice.
BALROG: Benchmarking Agentic LLM and VLM Reasoning On Games
Davide Paglieri (University College London), Tim Rocktäschel
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelTextMultimodalityBenchmark
🎯 What it does: The BALROG benchmark is proposed, integrating six long-term, complex reinforcement learning games (BabyAI, TextWorld, Crafter, Baba Is AI, MiniHack, NetHack) to evaluate the intelligence capabilities of large language models and visual language models in long-sequence decision-making tasks.
BAMDP Shaping: a Unified Framework for Intrinsic Motivation and Reward Shaping
Aly Lidayan (University of California), Stuart Russell (University of California)
Meta LearningRecurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: This paper proposes a unified framework (BAMDP Shaping) that views all intrinsic motivation and reward shaping as reward shaping on Bayes-Adaptive MDP (BAMDP). It provides theoretical analysis and empirical validation, demonstrating how to design potential functions (BAMPF) to accelerate exploration while avoiding reward hacking.
Bandit Learning in Matching Markets with Indifference
Fang Kong (Southern University of Science and Technology), Shuai Li (Shanghai Jiao Tong University)
Recommendation SystemOptimizationReinforcement Learning
🎯 What it does: This paper studies how participants in a matching market can learn preferences and achieve stable matching through a multi-armed bandit approach when preferences on the other side are unknown and may be the same (indifference). It proposes an arm-guided adaptive exploration algorithm, AE-AGS.
BANGS: Game-theoretic Node Selection for Graph Self-Training
Fangxin Wang (University of Illinois Chicago), Philip S. Yu (University of Illinois Chicago)
Graph Neural NetworkGraph
🎯 What it does: Proposes the BANGS framework, which selects pseudo-label nodes in a combinatorial manner during the graph self-training process to enhance the performance of GNNs.
Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression
Jingcun Wang (Technical University of Darmstadt), Grace Li Zhang (Technical University of Darmstadt)
CompressionTransformerLarge Language ModelText
🎯 What it does: The weight matrix of the pre-trained LLM shares a common basis vector across layers, achieving efficient compression;
Bayesian Analysis of Combinatorial Gaussian Process Bandits
Jack Sandberg (Chalmers University of Technology and University of Gothenburg), Morteza Haghir Chehreghani (Chalmers University of Technology and University of Gothenburg)
Recommendation SystemAutonomous DrivingOptimizationGraph
🎯 What it does: This paper addresses the problem of combinatorial variable Gaussian process semi-reward, deriving the Bayesian cumulative loss upper bounds for three algorithms: GP-UCB, GP-BayesUCB, and GP-TS, and applies this framework to online energy-saving navigation for electric vehicles.
Bayesian Experimental Design Via Contrastive Diffusions
Jacopo Iollo (Université Grenoble Alpes), Florence Forbes (Université Grenoble Alpes)
OptimizationDiffusion modelImage
🎯 What it does: A Bayesian experimental design method based on contrastive diffusion, CoDiff, is proposed, which can estimate the EIG gradient in a single sampling-optimization loop and is compatible with traditional density sampling and diffusion generative models.
Bayesian Image Regression with Soft-thresholded Conditional Autoregressive Prior
Yuliang Xu (Duke University), Jian Kang (University of Michigan)
Biomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper addresses the high-dimensional sparse regression problem of functional magnetic resonance imaging (fMRI) data by proposing a new prior called the Soft-Thresholded Conditional Auto-Regressive (ST-CAR) prior, which aims to achieve spatial smoothing, induce sparsity, and enable adaptive learning without relying on predefined correlation structures.
Bayesian Optimization of Antibodies Informed by a Generative Model of Evolving Sequences
Alan Nawzad Amin (New York University), Andrew Gordon Wilson (New York University)
OptimizationDrug DiscoveryLarge Language ModelReinforcement LearningSequentialBiomedical Data
🎯 What it does: A Clone Information Bayesian Optimization (CloneBO) method based on the antibody maturation process of the immune system is proposed for efficiently optimizing antibody sequences in the laboratory.
Bayesian Optimization via Continual Variational Last Layer Training
Paul Brunzema (RWTH Aachen University), James Harrison (Google DeepMind)
OptimizationTabularBenchmark
🎯 What it does: This paper proposes a Bayesian optimization surrogate model based on the Variational Bayesian Last Layer (VBLL) network, and achieves efficient model training and updating through an online continual learning algorithm;
Bayesian Regularization of Latent Representation
Chukwudi Paul Obite (Arizona State University), Shiwei Lan (Arizona State University)
Representation LearningTabular
🎯 What it does: A novel latent variable model QEP-LVM based on the Q-index process is proposed for learning latent representations, aimed at improving the visualization of data structures and the efficiency of model building.
Bayesian Treatment of the Spectrum of the Empirical Kernel in (Sub)Linear-Width Neural Networks
Ouns El Harzli (University of Oxford), Bernardo Cuenca Grau (University of Oxford)
Gaussian SplattingTabularPhysics Related
🎯 What it does: This study investigates Bayesian Neural Networks (BNN) in the limit of infinite samples, width, and input dimensions, deriving predictor integral formulas under linear and sublinear widths by relating the modified NNGP kernel to the spectral distribution of random matrices.
Bayesian WeakS-to-Strong from Text Classification to Generation
Ziyun Cui (Tsinghua University), Chao Zhang (Tsinghua University)
ClassificationGenerationReinforcement LearningText
🎯 What it does: Bayesian inference is performed on a set of weak models to enhance the transfer effect from weak to strong models, and this framework is extended from text classification to text generation; after training the strong model, conservative DPO is further used to improve preference learning.
Be More Diverse than the Most Diverse: Optimal Mixtures of Generative Models via Mixture-UCB Bandit Algorithms
Parham Rezaei (Sharif University of Technology), Cheuk Ting Li (Chinese University of Hong Kong)
GenerationData SynthesisOptimizationMixture of ExpertsImage
🎯 What it does: This study proposes a new mixed generative model selection mechanism aimed at improving the diversity and quality of generated data by mixing multiple generative models.
BEEM: Boosting Performance of Early Exit DNNs using Multi-Exit Classifiers as Experts
Divya Jyoti Bajpai (Indian Institute of Technology Bombay), Manjesh Kumar Hanawal (Indian Institute of Technology Bombay)
Computational EfficiencyKnowledge DistillationTransformerMixture of ExpertsImageText
🎯 What it does: The BEEM framework is proposed, treating multi-exit classifiers as experts, utilizing weighted confidence and prediction consistency to achieve early exit, thereby enhancing the inference speed and accuracy of Transformer models.
Behavioral Entropy-Guided Dataset Generation for Offline Reinforcement Learning
Wesley A. Suttle (U.S. Army Research Laboratory), Carlos Nieto-Granda (U.S. Army Research Laboratory)
Data SynthesisReinforcement Learning
🎯 What it does: A method for generating offline reinforcement learning datasets based on behavioral entropy is proposed, aiming to systematically generate diverse datasets that cover complex, high-dimensional state spaces.
Benchmarking Agentic Workflow Generation
Shuofei Qiao (Zhejiang University), Huajun Chen (Zhejiang University)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextGraphBenchmark
🎯 What it does: This paper presents WORFBENCH—a unified workflow generation benchmark that covers four types of task scenarios (problem solving, function calling, embodied planning, and open-ended planning) and models complex workflows using directed acyclic graphs (DAGs). It also designs the WORFEVAL evaluation protocol, which utilizes subsequence/subgraph matching algorithms to accurately quantify the generated node chains against the workflow graphs. A systematic evaluation of 18 large language models (including both closed and open-source models) is conducted, exploring the impact of workflows on downstream task performance and inference time, along with a quantitative analysis of the bottlenecks in graph structure generation.
Benchmarking LLMs' Judgments with No Gold Standard
Shengwei Xu (University of Michigan), Yuqing Kong (Peking University)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Two gold-standard-free evaluation metrics, GEM and GEM-S, are proposed, which estimate the mutual information between candidate and reference answers using generative language models, thereby quantifying the semantic information content of LLM-generated text. Based on this, GRE-bench is designed to evaluate LLM performance in subjective tasks such as academic peer review.
Benchmarking Multimodal Retrieval Augmented Generation with Dynamic VQA Dataset and Self-adaptive Planning Agent
Yangning Li (Tsinghua University), Philip S. Yu (Alibaba Group)
GenerationRetrievalTransformerLarge Language ModelMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper studies the challenges of multimodal retrieval-augmented generation (mRAG) in dynamic visual question answering (VQA) and proposes an adaptive retrieval planning agent called OmniSearch.
Benchmarking Predictive Coding Networks -- Made Simple
Luca Pinchetti (University of Oxford), Tommaso Salvatori (VERSES AI Research Lab)
ClassificationGenerationOptimizationComputational EfficiencyConvolutional Neural NetworkImageBenchmark
🎯 What it does: Developed and released a JAX-based PCX library for efficient implementation of predictive coding networks, providing a unified benchmark task and dataset for large-scale experiments.
Benchmarking Vision Language Model Unlearning via Fictitious Facial Identity Dataset
Yingzi Ma (University of Wisconsin Madison), Chaowei Xiao (University of Wisconsin Madison)
Safty and PrivacyTransformerSupervised Fine-TuningVision Language ModelImageTextBenchmark
🎯 What it does: This paper proposes the FIUBENCH benchmark, constructs a fictional facial identity VQA dataset, and designs a two-stage learning-forgetting evaluation pipeline to assess the performance of Vision-Language Models on the 'forgetting' task.
Benign Overfitting in Out-of-Distribution Generalization of Linear Models
Shange Tang (Princeton University), Chi Jin (Princeton University)
Tabular
🎯 What it does: This study investigates the phenomenon of 'benign overfitting' in over-parameterized linear models under covariate shift, and provides non-asymptotic risk upper bounds for ridge regression and principal component regression.
BenTo: Benchmark Reduction with In-Context Transferability
Hongyu Zhao (University of Maryland), Tianyi Zhou (Lehigh University)
OptimizationLarge Language ModelTextBenchmark
🎯 What it does: This paper studies the estimation of task transferability using the untrained In-Context Transferability (ICT) metric, and based on this, constructs the Facility Location problem to implement BENTO, reducing the number of benchmark tasks for LLM evaluation to within 5% while keeping the error below 4%.
Better autoregressive regression with LLMs via regression-aware fine-tuning
Michal Lukasik (Google Research), Sanjiv Kumar (Google Research)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningTabular
🎯 What it does: For fine-tuning decoder-based LLMs for regression tasks, we propose Regression-Aware Fine-Tuning (RAFT) and conduct a systematic comparison with traditional cross-entropy fine-tuning, RAIL decoding, and prediction head methods.
Better Instruction-Following Through Minimum Bayes Risk
Ian Wu (C3 AI), Graham Neubig (Carnegie Mellon University)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: By utilizing minimum Bayes risk (MBR) decoding based on an LLM judge, the generation quality of instruction-following large language models (LLMs) is improved, and the high computational cost during testing is eliminated through DPO self-training on MBR outputs.
Better than Your Teacher: LLM Agents that learn from Privileged AI Feedback
Sanjiban Choudhury (Cornell University), Paloma Sodhi (Cornell University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: By iterative fine-tuning and privileged AI teacher feedback, the performance of LLM agents in decision-making tasks is enhanced, allowing weaker models to surpass strong teacher models and even achieve self-improvement.
Beware of Calibration Data for Pruning Large Language Models
Yixin Ji (Soochow University), Min Zhang (Soochow University)
TransformerLarge Language ModelText
🎯 What it does: This paper studies the importance of calibration data in post-training pruning and proposes a method for self-generating synthetic calibration data (Self-Generating Synthetic Calibration Data, Syn) based on LLM to replace the original training data, thereby enhancing the downstream performance of the pruned model.
Beyond Autoregression: Discrete Diffusion for Complex Reasoning and Planning
Jiacheng Ye (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
Diffusion modelText
🎯 What it does: This paper studies the shortcomings of autoregressive language models in complex reasoning and planning tasks, and proposes a multi-granularity diffusion model (MGDM) based on discrete diffusion to address these issues.
Beyond Autoregression: Fast LLMs via Self-Distillation Through Time
Justin Deschenaux (École Polytechnique Fédérale de Lausanne), Caglar Gulcehre (École Polytechnique Fédérale de Lausanne)
GenerationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelDiffusion modelText
🎯 What it does: A self-timed distillation (SDTT) method is proposed, reducing the inference steps of discrete diffusion language models from thousands to dozens, significantly improving generation speed.
Beyond Canonicalization: How Tensorial Messages Improve Equivariant Message Passing
Peter Lippmann (Heidelberg University), Fred A. Hamprecht (Heidelberg University)
RecognitionSegmentationGraph Neural NetworkPoint Cloud
🎯 What it does: A tensor message passing framework based on local normalization is proposed to enhance equivariant message passing in geometric deep learning, allowing integration into arbitrary architectures.
Beyond Circuit Connections: A Non-Message Passing Graph Transformer Approach for Quantum Error Mitigation
Tianyi Bao (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Graph Neural NetworkTransformerGraphPhysics Related
🎯 What it does: A non-message passing graph transformer GTranQEM is proposed to suppress errors in the measurement results of quantum circuits, modeling global quantum coupling through graph encoding, quantum-specific positional encoding, structural matrix attention bias, and virtual QCR nodes.
Beyond Content Relevance: Evaluating Instruction Following in Retrieval Models
Jianqun Zhou (Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative, Institute of Digital Twin, Eastern Institute of Technology), Xiaoyu Shen (Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative, Institute of Digital Twin, Eastern Institute of Technology)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper constructs a new retrieval evaluation benchmark named InfoSearch, which is used to assess the instruction-following ability of retrieval models on six document-level attributes beyond content relevance (audience, keywords, format, language, length, source), and proposes two new evaluation metrics, SICR and WISE.
Beyond correlation: The impact of human uncertainty in measuring the effectiveness of automatic evaluation and LLM-as-a-judge
Aparna Elangovan (Amazon), Dan Roth (Amazon)
Large Language ModelTextBenchmark
🎯 What it does: This study investigates the correlation between automatic evaluation methods and human labels, revealing the impact of human uncertainty on evaluation metrics, and proposes three new methods: stratification by label uncertainty, binned Jensen-Shannon divergence, and perceptual visualization.
Beyond FVD: An Enhanced Evaluation Metrics for Video Generation Distribution Quality
Ge Ya Luo (Mila - Quebec Artificial Intelligence Institute), Christopher Pal (Mila - Quebec Artificial Intelligence Institute)
GenerationData SynthesisAuto EncoderContrastive LearningVideo
🎯 What it does: This paper studies the limitations of video generation evaluation metrics and proposes a new metric, JEDi, based on the V-JEPA feature space and polynomial kernel MMD, and validates its superiority through experiments on multiple datasets.
Beyond Graphs: Can Large Language Models Comprehend Hypergraphs?
Yifan Feng (Tsinghua University), Yue Gao (Tsinghua University)
TransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkChain-of-Thought
🎯 What it does: Designed and released the LLM4Hypergraph benchmark to evaluate the understanding and reasoning capabilities of large language models regarding hypergraph structures, and proposed two new prompting techniques: Hyper-BAG and Hyper-COT.
Beyond Interpretability: The Gains of Feature Monosemanticity on Model Robustness
Qi Zhang (Peking University), Yisen Wang (Peking University)
ClassificationExplainability and InterpretabilityAuto EncoderContrastive LearningImageText
🎯 What it does: This study investigates the impact of monosemanticity on model robustness and demonstrates that monosemantic features can significantly enhance model robustness in various scenarios, such as input noise, label noise, and few-shot fine-tuning, challenging the common accuracy-interpretability trade-off.
Beyond Linear Approximations: A Novel Pruning Approach for Attention Matrix
Yingyu Liang (University of Hong Kong), Yufa Zhou (University of Pennsylvania)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A new weight pruning method for large language models (LLM) is proposed, which directly optimizes the approximation of the attention matrix, addressing the issue that existing methods only focus on linear approximations.
Beyond Mere Token Analysis: A Hypergraph Metric Space Framework for Defending Against Socially Engineered LLM Attacks
Manohar Kaul (Fujitsu Research of India Private Limited), Sadbhavana Babar (Fujitsu Research of India Private Limited)
ClassificationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a defense framework based on hypergraph metric spaces to identify and prevent social engineering attacks on large language models.
Beyond Model Collapse: Scaling Up with Synthesized Data Requires Verification
Yunzhen Feng (Meta), Julia Kempe (New York University)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies how to use a verifier to filter high-quality samples from synthetic data to prevent model collapse when training models with synthetic data.
Beyond Next Token Prediction: Patch-Level Training for Large Language Models
Chenze Shao (Tencent Inc), Jie Zhou (Tencent Inc)
TransformerLarge Language ModelText
🎯 What it does: A patch-level training method is proposed, which significantly reduces the training cost of LLMs by aggregating multiple tokens into patches and predicting the next patch at the patch level.
Beyond Random Augmentations: Pretraining with Hard Views
Fabio Ferreira (University of Freiburg), Frank Hutter (University of Freiburg)
ClassificationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: A hard view selection mechanism is introduced in self-supervised pre-training, where multiple randomly augmented views are generated for each image, and the loss of all view pairs is calculated to select the pair with the highest loss for training;
Beyond Random Masking: When Dropout meets Graph Convolutional Networks
Yuankai Luo (Beihang University), Hao Zhu (Data61 CSIRO)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Analyzed the theoretical mechanism of Dropout in Graph Convolutional Networks (GCN) and experimentally verified its positive impact on oversmoothing and generalization.
Beyond Sequence: Impact of Geometric Context for RNA Property Prediction
Junjie Xu (Pennsylvania State University), Rui Liao (Johnson and Johnson Innovative Medicine)
Hyperparameter SearchProtein Structure PredictionGraph Neural NetworkTransformerPoint CloudGraphBiomedical Data
🎯 What it does: The system evaluates the effect of explicitly incorporating the 2D (graph structure) and 3D (atomic point cloud) geometric information of RNA into attribute prediction models, and provides a novel RNA dataset with 2D/3D annotations.
Beyond Single Concept Vector: Modeling Concept Subspace in LLMs with Gaussian Distribution
Haiyan Zhao (New Jersey Institute of Technology), Mengnan Du (New Jersey Institute of Technology)
TransformerLarge Language ModelText
🎯 What it does: Proposes the Gaussian Concept Subspace (GCS) framework, which characterizes the concept subspace in LLMs using Gaussian distributions instead of a single vector, enhancing the robustness and diversity of concept representations.
Beyond single neurons: population response geometry in digital twins of mouse visual cortex
Dario Liscai (Bocconi University), Alessandro Sanzeni (Bocconi University)
Convolutional Neural NetworkBiomedical Data
🎯 What it does: By training a digital twin model to predict the responses of mouse visual cortex neuron populations, and analyzing their geometric structure and hierarchical differences.
Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks
Rui Hu (Tsinghua University), Longbo Huang (Tsinghua University)
GenerationData SynthesisGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: This paper studies the theory and practice of regression loss in the training of Generative Flow Networks (GFlowNets), proposing a new loss design method based on f-divergence, and improving the training effectiveness of GFlowNet through three new losses: Shifted-Cosh, Linex(1/2), and Linex(1).
Beyond Surface Structure: A Causal Assessment of LLMs' Comprehension ability
Yujin Han (Hong Kong University), Chaochao Lu (Shanghai Artificial Intelligence Laboratory)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A causal evaluation of large language models (LLMs) regarding their understanding of deep structure (core semantics) and surface structure (discourse form) is conducted, proposing quantifiable metrics ADCE (Approximate Direct Causal Effect) and AICE (Approximate Indirect Causal Effect) to measure the model's dependency on different structures, with experimental validation across multiple tasks.
Beyond the convexity assumption: Realistic tabular data generation under quantifier-free real linear constraints
Mihaela C. Stoian (University of Oxford), Eleonora Giunchiglia (Imperial College London)
GenerationData SynthesisGenerative Adversarial NetworkTabular
🎯 What it does: This paper proposes the Disjunctive Refinement Layer (DRL), which can compile quantified first-order linear arithmetic (QFLRA) constraints into differentiable neural network layers, thereby enforcing the satisfaction of non-convex, discrete background knowledge constraints when generating tabular data.
Beyond Worst-Case Dimensionality Reduction for Sparse Vectors
Sandeep Silwal (University of Wisconsin Madison), Qiuyi Zhang (Google DeepMind)
Tabular
🎯 What it does: This paper studies the average case dimensionality reduction of sparse vectors and proposes that a smaller embedding dimension can be achieved on non-negative sparse data compared to traditional birthday paradox mappings.
Beyond-Expert Performance with Limited Demonstrations: Efficient Imitation Learning with Double Exploration
Heyang Zhao (University of California), Quanquan Gu (University of California)
Robotic IntelligenceReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: A dual exploration imitation learning algorithm ILDE is proposed, which integrates expert behavior cloning and two types of exploration rewards, achieving performance beyond that of experts and improving sample efficiency.
Bi-Factorial Preference Optimization: Balancing Safety-Helpfulness in Language Models
Wenxuan Zhang (King Abdullah University of Science and Technology), Adel Bibi (University of Oxford)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A novel supervised alignment framework BFPO is proposed to simultaneously enhance the usefulness and safety of large language models.
Bias Mitigation in Graph Diffusion Models
Meng Yu (Lanzhou University), Kun Zhan (Lanzhou University)
Drug DiscoveryGraph Neural NetworkDiffusion modelGraphStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: To address the inverse starting point bias and exposure bias in graph diffusion models, a network-free modification method is proposed that aligns the forward maximum perturbation distribution through Langevin sampling and performs score correction based on score differences.
Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling
Yuejiang Liu (Stanford University), Chelsea Finn (Stanford University)
Robotic IntelligenceDiffusion modelContrastive LearningMultimodality
🎯 What it does: Proposes an action chunking improvement method based on bidirectional decoding, utilizing multiple sampling and forward-backward consistency evaluation to achieve closed-loop action planning.
BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions
Terry Yue Zhuo (Monash University), Leandro Von Werra
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: A benchmark called BigCodeBench is proposed and constructed, which includes 1,140 Python programming tasks that require multiple tool calls and complex instructions. It also provides a natural language instruction version, BigCodeBench Instruct, for evaluating the code generation capabilities of LLMs.
BigDocs: An Open Dataset for Training Multimodal Models on Document and Code Tasks
Juan A. Rodriguez (ServiceNow), Sai Rajeswar (ServiceNow)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper presents BigDocs-7.5M, an open-source, license-friendly dataset of 7.5 million image-text pairs. Based on this dataset, we conducted fine-tuning and compared it with existing datasets, further constructing 10 new tasks in BigDocs-Bench to evaluate models' capabilities in generating structured code (HTML, LaTeX, SVG, etc.) from images and understanding GUIs.
BiGR: Harnessing Binary Latent Codes for Image Generation and Improved Visual Representation Capabilities
Shaozhe Hao (University of Hong Kong), Kwan-Yee K. Wong (University of Hong Kong)
GenerationRepresentation LearningTransformerDiffusion modelAuto EncoderImage
🎯 What it does: BiGR is a model that utilizes binary latent codes for conditional image generation, while also achieving powerful representation learning capabilities within the same framework.
Bilinear MLPs enable weight-based mechanistic interpretability
Michael T Pearce, Lee Sharkey (Apollo Research)
Explainability and InterpretabilityTransformerAuto EncoderImageText
🎯 What it does: This paper proposes and validates the use of a bilinear MLP (Multilayer Perceptron) with non-linear activation as an interpretable alternative in Transformers, utilizing its third-order tensor structure for feature interaction analysis of weights.
Binary Losses for Density Ratio Estimation
Werner Zellinger (Johannes Kepler University Linz)
Domain AdaptationOptimizationImageText
🎯 What it does: This paper constructs a general framework for binary classification loss functions by providing a given Bregman divergence, and based on this framework, proposes a novel loss function that prioritizes estimating large density ratios.
BinaryDM: Accurate Weight Binarization for Efficient Diffusion Models
Xingyu Zheng (Beihang University), Michele Magno (ETH Zurich)
GenerationData SynthesisCompressionComputational EfficiencyDiffusion modelImage
🎯 What it does: A weight quantization binarization method specifically designed for diffusion models, BinaryDM, has been developed, which can maintain or improve generation quality at ultra-low bit widths.
BingoGuard: LLM Content Moderation Tools with Risk Levels
Fan Yin (Salesforce), Chien-Sheng Wu (Salesforce)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: BingoGuard has been developed, a content moderation system capable of generating content for LLMs while outputting safety binary labels and multi-level severity labels, along with the corresponding training set BingoGuardTrain and test set BingoGuardTest.
Bio-xLSTM: Generative modeling, representation and in-context learning of biological and chemical sequences
Niklas Schmidinger (ELLIS Unit Linz and LIT AI Lab), Günter Klambauer (ELLIS Unit Linz and LIT AI Lab)
GenerationRepresentation LearningDrug DiscoveryRecurrent Neural NetworkSequentialBiomedical Data
🎯 What it does: Designed and implemented three variants of Bio-xLSTM for DNA, protein, and SMILES sequences, and evaluated them on multiple tasks including generation, representation learning, and context learning.
BioDiscoveryAgent: An AI Agent for Designing Genetic Perturbation Experiments
Yusuf H Roohani, Jure Leskovec (Stanford University)
OptimizationDrug DiscoveryTransformerLarge Language ModelAgentic AIPrompt EngineeringBiomedical Data
🎯 What it does: This paper presents BioDiscoveryAgent, a closed-loop experimental design agent based on large language models and tools, which is used to automatically design gene perturbation experiments and continuously optimize the next batch of experiments based on the results.
Biologically Constrained Barrel Cortex Model Integrates Whisker Inputs and Replicates Key Brain Network Dynamics
Tianfang Zhu (Wuhan National Laboratory for Optoelectronics Huazhong University of Science and Technology), Anan LI
ClassificationSpiking Neural NetworkReinforcement LearningTime Series
🎯 What it does: A biologically constrained rat dorsal whisker cortex model containing 13 types of neuron subtypes and 4218 neurons was constructed and trained. Based on spike input, object classification was achieved by converting physical simulation whisker sweep data into spike input.
Biologically Plausible Brain Graph Transformer
Ciyuan Peng (Federation University Australia), Yaochu Jin (Westlake University)
ClassificationAnomaly DetectionGraph Neural NetworkTransformerContrastive LearningGraphBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: A Transformer-based brain map learning model, BioBGT, has been designed and implemented to enhance the biological plausibility of brain map representations by encoding small-world structures (node importance and functional modules) and is used for brain disease detection tasks.
BIRD: A Trustworthy Bayesian Inference Framework for Large Language Models
Yu Feng (University of Pennsylvania), Dan Roth (University of Pennsylvania)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the BIRD framework, which combines the inductive reasoning of LLMs with Bayesian networks to achieve reliable probability estimation under incomplete information;
BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics
Lukas Rauch (University of Kassel), Christoph Scholz (University of Kassel)
ClassificationConvolutional Neural NetworkTransformerBenchmarkAudio
🎯 What it does: Created BirdSet, a large-scale bird audio classification benchmark dataset, containing approximately 520k recordings, 6,800 hours of training, and 400 hours of testing, supporting multi-label and multi-scenario evaluations;
Bisimulation Metric for Model Predictive Control
Yutaka Shimizu (University of California), Masayoshi Tomizuka (University of California)
OptimizationComputational EfficiencyConvolutional Neural NetworkReinforcement LearningImageVideo
🎯 What it does: A model predictive control method based on bisimulation metric, BS-MPC, is proposed and implemented to improve encoder training, enhance training stability, noise robustness, and computational efficiency.
BitStack: Any-Size Compression of Large Language Models in Variable Memory Environments
Xinghao Wang (Fudan University), Xipeng Qiu (Fudan University)
CompressionTransformerLarge Language ModelText
🎯 What it does: A training-independent decomposition compression method named BitStack is proposed, enabling LLM to dynamically adjust model size at the megabyte level in variable memory environments.
Black Sheep in the Herd: Playing with Spuriously Correlated Attributes for Vision-Language Recognition
Xinyu Tian (Australian National University), Jing Zhang (Australian National University)
RecognitionDomain AdaptationTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: This study addresses the out-of-distribution generalization problem of visual-language models in few-shot scenarios by proposing the SAP method, which automatically identifies and removes 'black sheep' pseudo-related attributes, and designs a pluggable pseudo-attribute masking module (SAS) to reduce the model's dependence on noisy attributes.
Black-Box Detection of Language Model Watermarks
Thibaud Gloaguen (ETH Zurich), Martin Vechev (ETH Zurich)
Large Language ModelText
🎯 What it does: A rigorous statistical detection method is designed for three mainstream watermarking schemes (Red-Green, Fixed Sampling, Cache Enhancement) in a black-box environment;
BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation
Zhengrui Guo (Hong Kong University of Science and Technology), Hao Chen (Hong Kong University of Science and Technology)
Knowledge DistillationMultimodality
🎯 What it does: A neural population dynamics modeling framework called BLEND is proposed, which utilizes behavioral information as privileged knowledge for knowledge distillation.
BlendRL: A Framework for Merging Symbolic and Neural Policy Learning
Hikaru Shindo (Technical University of Darmstadt), Kristian Kersting (Technical University of Darmstadt)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningImage
🎯 What it does: The BlendRL framework is proposed, which integrates neural networks with differentiable logical reasoning to form parallel neural and symbolic policies that are dynamically allocated through a hybrid module, ultimately achieving an RL agent that possesses both low-level reactions and high-level reasoning.
Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models
Marianne Arriola (Cornell University), Volodymyr Kuleshov
GenerationTransformerLarge Language ModelDiffusion modelText
🎯 What it does: A Block Diffusion Language Model (BD3-LM) is proposed, which combines discrete diffusion and autoregression at the block level, supporting arbitrary length generation and utilizing KV caching for efficient inference.
Block Verification Accelerates Speculative Decoding
Ziteng Sun (Google Research), Ananda Theertha Suresh (Google Research)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A Block Verification algorithm is proposed for speculative decoding of large models during inference, improving traditional Token Verification, thereby increasing generation speed while maintaining lossless output distribution.
Block-Attention for Efficient Prefilling
Dongyang Ma (Tencent), Tian Lan
RetrievalComputational EfficiencyTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Proposes the Block-Attention mechanism, which allows for chunking retrieval documents and calculating KV states separately, significantly reducing inference latency and FLOPs in RAG scenarios while maintaining performance close to full attention.
BlueSuffix: Reinforced Blue Teaming for Vision-Language Models Against Jailbreak Attacks
Yunhan Zhao (Fudan University), Yu-Gang Jiang (Fudan University)
Adversarial AttackTransformerLarge Language ModelReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: A black-box blue team defense framework called BlueSuffix is proposed, which defends against cross-modal jailbreak attacks on Vision-Language Models through image denoising diffusion models, LLM text rewriting, and suffixes generated by reinforcement learning.
BodyGen: Advancing Towards Efficient Embodiment Co-Design
Haofei Lu (Tsinghua University), Yuanchun Shi (Tsinghua University)
OptimizationRobotic IntelligenceTransformerReinforcement LearningMultimodality
🎯 What it does: This paper presents BodyGen, an efficient embodiment co-design framework based on reinforcement learning that can simultaneously optimize robot morphology and control strategies.
BOFormer: Learning to Solve Multi-Objective Bayesian Optimization via Non-Markovian RL
Yu Heng Hung, Ping-Chun Hsieh (National Yang Ming Chiao Tung University)
OptimizationTransformerReinforcement LearningGaussian SplattingTabular
🎯 What it does: This paper proposes BOFormer, a multi-objective Bayesian optimization learning acquisition function based on non-Markov deep Q networks and Transformers, addressing the non-Markov identification problem in multi-objective acquisition functions.
Boltzmann priors for Implicit Transfer Operators
Juan Viguera Diez (Chalmers University of Technology), Simon Olsson (Chalmers University of Technology)
GenerationData SynthesisDrug DiscoveryDiffusion modelScore-based ModelTime SeriesSequential
🎯 What it does: Using Boltzmann Prior to enhance the learning of Implicit Transfer Operator, guiding the generation of samples through pre-trained BG and embedding long-term dynamical priors, achieving efficient modeling of long time series in molecular dynamics.
Boltzmann Semantic Score: A Semantic Metric for Evaluating Large Vision Models Using Large Language Models
Ali Khajegili Mirabadi (University of British Columbia), Ali Bashashati (University of British Columbia)
Large Language ModelImageTextBiomedical Data
🎯 What it does: This paper proposes the Boltzmann Semantic Score (BSS) to measure the semantic capability of large visual models (LVMs) and evaluates the BSS of seven LVMs against five large language models (LLMs) on the TCGA dataset.
Boltzmann-Aligned Inverse Folding Model as a Predictor of Mutational Effects on Protein-Protein Interactions
Xiaoran Jiao (Zhejiang University), Chunhua Shen (Zhejiang University)
Protein Structure PredictionSupervised Fine-TuningBiomedical Data
🎯 What it does: A Boltzmann alignment-based inverse folding model has been developed to predict the variation in protein-protein interaction ΔΔG, explicitly considering both bound and unbound states to enhance prediction accuracy.
BOND: Aligning LLMs with Best-of-N Distillation
Pier Giuseppe Sessa (Google DeepMind), Olivier Bachem (Google DeepMind)
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies a RLHF algorithm called BOND that distills the Best-of-N sampling strategy into LLM through distribution matching, and implements a specific version called J-BOND;
BoneMet: An Open Large-Scale Multi-Modal Murine Dataset for Breast Cancer Bone Metastasis Diagnosis and Prognosis
Tiankuo Chu (University of Delaware), Liyun Wang (University of Delaware)
SegmentationGenerationConvolutional Neural NetworkTransformerNeural Radiance FieldGenerative Adversarial NetworkMultimodalityBiomedical DataComputed Tomography
🎯 What it does: BoneMet is proposed and released - the first large-scale, full high-resolution, serialized multimodal mouse bone metastasis imaging dataset, supporting the diagnosis, prognosis, and imaging processing research of B-cell bone metastasis.