International Conference on Learning Representations Β· 2207 papers
AlphaSAGE: Structure-Aware Alpha Mining via GFlowNets for Robust Exploration
Binqi Chen (Peking University), Ming Zhang (Peking University)
CodeGraph Neural NetworkReinforcement LearningFlow-based ModelTabularTime SeriesFinance Related
π― What it does: Proposes AlphaSAGE, a framework for automatically discovering high-quality, structurally diverse alphas, leveraging structure-aware AST encoding and generative flow networks (GFlowNet) to sample multi-modal alphas, and achieving more robust trading signals through multi-dimensional rewards and dynamic linear combination.
AlphaSteer: Learning Refusal Steering with Principled Null-Space Constraint
Leheng Sheng (National University Of Singapore), Tat-Seng Chua (National University Of Singapore)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Developed an activation-guided method called AlphaSteer, which utilizes a learnable transformation matrix and dynamically steers the activation of malicious prompts toward rejection directions within the zero space of well-activated states, while maintaining the functionality of normal prompts unaffected.
Ambig-SWE: Interactive Agents to Overcome Underspecificity in Software Engineering
Sanidhya Vijayvargiya (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
CodeAI Code AssistantLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: Built the Ambig-SWE evaluation framework to systematically assess LLMs' capabilities in detecting missing information, clarifying questions, and completing tasks interactively in software engineering scenarios.
AMiD: Knowledge Distillation for LLMs with $\alpha$-mixture Assistant Distribution
Donghyeok Shin (Korea Advanced Institute of Science and Technology), Il-chul Moon
CodeKnowledge DistillationLarge Language ModelText
π― What it does: Proposed the Ξ±-mixture assistant distribution and the AMiD framework for knowledge distillation in large language models, unifying and extending existing assistant distribution and divergence methods.
π― What it does: Proposed an Alignment-Aware Masked Learning (AML) training strategy for Referring Image Segmentation (RIS), which enhances model alignment and segmentation performance by computing pixel-level visual-language alignment and masking unreliable pixels.
AMPED: Adaptive Multi-objective Projection for balancing Exploration and skill Diversification
Geonwoo Cho (Gwangju Institute of Science and Technology), Sundong Kim (Gwangju Institute of Science and Technology)
CodeReinforcement LearningContrastive Learning
π― What it does: Propose the AMPED method to simultaneously balance exploration and skill diversity during the skill learning phase, and achieve adaptive skill deployment through a skill selector during the fine-tuning phase.
An Agentic Framework with LLMs for Solving Complex Vehicle Routing Problems
Ni Zhang (Singapore Management University), Yew-Soon Ong (Nanyang Technological University)
CodeOptimizationTransformerLarge Language ModelAgentic AIGraphBenchmark
π― What it does: Propose a fully automated, external-module-free LLM agent framework called AFL for end-to-end solving complex vehicle routing problems.
An Efficient SE(p)-Invariant Transport Metric Driven by Polar Transport Discrepancy-based Representation
Junyi Lin (Renmin University of China), Cheng Meng (Renmin University of China)
CodeClassificationGenerationDrug DiscoveryGenerative Adversarial NetworkPoint CloudBiomedical Data
π― What it does: Proposed a new SE(p) invariant optimal transport metric called SEINT, and introduced unsupervised, training-free representations named Polar Transport Discrepancy (PTD) and Distance-convoluted PTD (DcPTD), while applying them as regularization terms in molecular generation models.
An Information-Theoretic Parameter-Free Bayesian Framework for Probing Labeled Dependency Trees from Attention Score
Hongxu Liu (Nanyang Technological University), Fangxiang Feng (Beijing University of Posts and Telecommunications)
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: Propose an untrained network, information-theoretic Bayesian framework (IPBP), which directly reconstructs labeled syntactic dependency trees by estimating the mutual information between attention scores and dependencies.
An Open-Ended Benchmark and Formal Framework for Adjuvant Research with MLLM
yi chen, Cheng-Lin Liu (Chinese Academy of Sciences)
CodeDrug DiscoveryTransformerLarge Language ModelTextBenchmark
π― What it does: Proposed the first open-ended question-answering benchmark and formal description framework for adjuvants, and conducted systematic evaluations of various multimodal large language models (MLLMs);
π― What it does: This paper proposes a new framework based on terminal entropy regularized stochastic control (TEC), which solves the rate-distortion (RD) function of continuous sources under mean squared error (MSE) distortion using diffusion processes, and provides a numerical estimation method called R2D2;
An Orthogonal Learner for Individualized Outcomes in Markov Decision Processes
Emil Javurek (LMU Munich), Stefan Feuerriegel (LMU Munich)
CodeMeta LearningReinforcement LearningBenchmark
π― What it does: Propose a new DR Q-learner for estimating the Q-function in Markov Decision Processes (MDPs) from observational data, combining double robustness, Neyman-orthogonality, and quasi-likelihood efficiency.
Analyzing and Evaluating Unbiased Language Model Watermark
Yihan Wu (University of Maryland), Heng Huang (University of Maryland)
CodeGenerationTransformerLarge Language ModelTextBenchmark
π― What it does: Proposed an open-source benchmark called UWBENCH specifically for evaluating unbiased watermarks, including theoretical analysis and experimental evaluation
Analyzing the Training Dynamics of Image Restoration Transformers: A Revisit to Layer Normalization
MinKyu Lee (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
CodeRestorationSuper ResolutionTransformerImage
π― What it does: Analyzes the training dynamics of image restoration Transformers, revealing that traditional LayerNorm causes feature amplitude explosion and sudden drop in channel entropy, leading to the proposal of an i-LN normalization scheme tailored for image restoration tasks.
He Zhu (Southern University of Science and Technology), Guanhua Chen (Shanghai Artificial Intelligence Laboratory)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningTextBiomedical Data
π― What it does: Proposed and evaluated an Anchored Supervised Fine-Tuning (ASFT) method based on reward-weighted regression to enhance the generalization of large language models (LLMs) while maintaining the efficiency of supervised fine-tuning (SFT).
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical DataBenchmarkChain-of-Thought
π― What it does: Proposed AnesSuite, which includes a cross-lingual evaluation benchmark called AnesBench, along with three training datasets (AnesCorpus, AnesQA, AnesR1), and trained the Morpheus series of LLMs as a baseline for anesthesia reasoning.
Antislop: A Comprehensive Framework for Identifying and Eliminating Repetitive Patterns in Language Models
Samuel J Paech (Liquid AI), Ravid Shwartz-Ziv (New York University)
CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes the Antislop framework to address the problem of repeated words and phrases (called slop) in the outputs of large language models (LLMs). The framework includes: 1) Antislop Sampler β which detects and suppresses already-occurred slop during inference through backtracking; 2) an automated pipeline β which generates slop fingerprints and automatically constructs training samples by comparing the model with human baselines; 3) Final Token Preference Optimization (FTPO) β which precisely adjusts the logits of individual tokens during training to enable the model to naturally avoid slop.
Jaeyeon Kim (Harvard University), Michael Samuel Albergo (Harvard University)
CodeGenerationTransformerDiffusion modelText
π― What it does: Proposed FlexMDM, a discrete diffusion model capable of handling variable-length sequences, demonstrating its effectiveness in pre-training and large-scale fine-tuning
π― What it does: Propose a generic framework called Any-Subgroup Equivariant Networks (ASEN), achieving equivariance for any subgroup by introducing spin symmetry breaking in the base model, with the same network capable of handling multiple data types and tasks;
π― What it does: Proposed a generic feature upsampling method called AnyUp, which can upsample features extracted from any visual encoder during inference;
AP-OOD: Attention Pooling for Out-of- Distribution Detection
Claus Hofmann (Johannes Kepler University), Werner Zellinger (Johannes Kepler University)
CodeAnomaly DetectionTransformerTextAudio
π― What it does: Proposed a semi-supervised OOD detection method called AP-OOD based on attention pooling, which can utilize token-level information to detect OOD in natural language text under unsupervised or few abnormal sample conditions.
APT: Towards Universal Scene Graph Generation via Plug-in Adaptive Prompt Tuning
Ruikun Luo (National Engineering Research Center for Big Data Technology and System), Xiaoyu Xia (Royal Melbourne Institute of Technology)
CodeGenerationTransformerPrompt EngineeringVision Language ModelMultimodality
π― What it does: Proposes a pluggable Adaptive Prompt Tuning (APT) module that converts frozen language model semantic features into context-aware dynamic representations to enhance scene graph generation.
π― What it does: Propose ABSignSGD, a block-coordinate based optimizer using sign gradients, to achieve memory and runtime efficiency for full-parameter fine-tuning of large language models without sacrificing performance.
π― What it does: Unify the controllable generation of spatial attributes such as viewpoint, field of view (FOV), and resolution, supporting various image shapes including perspective, panoramic, and fisheye views;
π― What it does: Proposes a test-time adaptation method called Progressive Embedding Alignment (PEA), which corrects domain drift through layer-wise covariance alignment without using backpropagation.
π― What it does: Proposed and systematically evaluated multiple EEG base models (including the newly constructed ST-EEGFormer) under multi-task and multi-evaluation protocols, comparing their performance with classical CNNs and traditional non-neural decoders;
Are Global Dependencies Necessary? Scalable Time Series Forecasting via Local Cross-Variate Modeling
Kun Liu (East China Normal University), Cen Chen (East China Normal University)
CodeConvolutional Neural NetworkAuto EncoderTime Series
π― What it does: Propose a time series prediction framework VPNet based on local cross-variable interaction, mapping inputs to a high-dimensional variate-patch field through a Patch-level autoencoder, then performing linear complexity local convolution processing with VarTCNBlock, and finally outputting prediction results.
Are LLMs Really Not Knowledgeable? Mining the Submerged Knowledge in LLMs' Memory
Xingjian Tao (Hong Kong University of Science and Technology (Guangzhou)), Jing Tang (Hong Kong University of Science and Technology (Guangzhou))
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: This paper analyzes the output distribution of LLMs to reveal the gap between knowledge storage and expression in question-answering tasks, and proposes the Hits@k metric to evaluate potential knowledge.
ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping
Shuang Chen (University of California Los Angeles), Nanyun Peng (University of California Los Angeles)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Proposed and implemented a framework named ARES, aiming to enable multimodal large-scale reasoning models (MLRMs) to adaptively allocate reasoning steps based on task difficulty, avoiding excessive reasoning on simple tasks and insufficient reasoning on complex tasks.
π― What it does: Propose Aria, an automated formal agent capable of recursively decomposing mathematical statements and automatically generating Lean 4 formalizations through Graph-of-Thought and Retrieval-Augmented Generation (RAG), with self-reflection in the compilation loop; simultaneously design AriaScorer for term-level semantic checks, retrieving Mathlib definitions and comparing them with the original text to improve accuracy.
ARROW: An Adaptive Rollout and Routing Method for Global Weather Forecasting
Jindong Tian (East China Normal University), Bin Yang (East China Normal University)
CodeTransformerReinforcement LearningMixture of ExpertsTime SeriesPhysics Related
π― What it does: Propose a global weather prediction framework ARROW based on adaptive rolling and routing, incorporating a multi-time-step prediction model and an adaptive rolling scheduler.
ASCIIEval: Benchmarking Models' Visual Perception in Text Strings via ASCII Art
Qi Jia (Shanghai Artificial Intelligence Laboratory National University of Singapore), Guangtao Zhai (Shanghai Artificial Intelligence Laboratory National University of Singapore)
CodeClassificationRecognitionTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelTextMultimodalityBenchmark
π― What it does: This paper proposes the ASCIIEval benchmark, which evaluates the text-visual perception capabilities of LLMs and MLLMs using ASCII art;
ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack
Yein Park (Korea University), Jaewoo Kang (Korea University)
CodeSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyRepresentation LearningAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningTextSequentialBenchmark
π― What it does: Locate and repair security vulnerabilities in LLMs during tense attacks through Transformer circuit analysis, achieving precise secure alignment via channel-level activation scaling and preventive fine-tuning.
ASIDE: Architectural Separation of Instructions and Data in Language Models
Egor Zverev (Institute of Science and Technology Austria), Christoph H. Lampert (Institute of Science and Technology Austria)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Propose and evaluate an architectural modification named ASIDE, which enhances the security of large language models (LLMs) by applying a fixed orthogonal rotation to data tokens in the token embedding layer, explicitly distinguishing instructions and data in the embedding space.
ASMIL: Attention-Stabilized Multiple Instance Learning for Whole-Slide Imaging
Linfeng Ye (University of Toronto), Konstantinos N. Plataniotis (University of Toronto)
CodeClassificationTransformerBiomedical Data
π― What it does: This paper proposes a multi-instance learning framework called ASMIL based on attention stabilization for weakly supervised diagnosis on whole slide images (WSI);
ASSESS: A Semantic and Structural Evaluation Framework for Statement Similarity
Xiaoyang Liu (Shanghai Jiao Tong University), Tao Luo (Shanghai Jiao Tong University)
CodeClassificationTextBenchmark
π― What it does: Developed a framework called ASSESS for automatically evaluating the similarity of formal statements, and proposed the TransTED Similarity measure that integrates semantic transformations.
AstaBench: Rigorous Benchmarking of AI Agents with a Scientific Research Suite
Jonathan Bragg, Daniel S Weld (Asta Team, Allen Institute for AI)
CodeLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
π― What it does: Developed AstaBench, a complete evaluation framework containing over 2400 fine-grained scientific research tasks, a reproducible Asta environment, a unified tool interface, and an agent-baselines kit, conducting large-scale evaluations on 57 agents.
Astra: General Interactive World Model with Autoregressive Denoising
Yixuan Zhu (Tsinghua University), Jie Zhou (Tsinghua University)
CodeGenerationData SynthesisTransformerMixture of ExpertsVision-Language-Action ModelDiffusion modelScore-based ModelWorld ModelVideo
π― What it does: Built an interactive general-purpose world model Astra, which uses an autoregressive denoising framework to generate long-term, controllable videos from initial frames and action sequences.
π― What it does: Proposed the AsyncBEV module, which aligns multi-modal features by predicting BEV space flow under sensor time asynchrony, thereby enhancing the robustness of 3D object detection.
Asynchronous Denoising Diffusion Models for Aligning Text-to-Image Generation
Zijing Hu (Zhejiang University), Kun Kuang (Zhejiang University)
CodeGenerationDiffusion modelImageText
π― What it does: Achieve asynchronous denoising by assigning different time steps to each pixel and using a variable concave time schedule, thereby improving text-to-image alignment.
ATEX-CF: Attack-Informed Counterfactual Explanations for Graph Neural Networks
Yu Zhang (Aalborg University), Cuneyt Gurcan Akcora (University of Central Florida)
CodeExplainability and InterpretabilityAdversarial AttackGraph Neural NetworkGraph
π― What it does: Investigated the integration of adversarial attacks with causal counterfactual explanations, proposing the ATEX-CF framework to generate interpretable explanations involving edge additions and deletions.
CodeGenerationTransformerLarge Language ModelReinforcement LearningText
π― What it does: Train a test case generator ATGEN using adversarial reinforcement learning to generate high-quality test cases as it continuously encounters more challenging bugs.
π― What it does: This paper proposes the concept of atomic HIN and performs attribute atomization on data based on the entity-attribute duality principle, constructing the most expressive HIN; subsequently, a task-specific schema refinement method is designed, optimizing the graph structure through binary selection of node types and relationship types, and systematically searching using genetic algorithms; experiments are conducted on eight benchmark datasets using simplified RGCN (sRGCN) and more advanced HGNN to verify its effectiveness.
ATPO: ADAPTIVE TREE POLICY OPTIMIZATION FOR MULTI-TURN MEDICAL DIALOGUE
Ruike Cao (University of Science and Technology of China), Li Xiao (University of Science and Technology of China)
CodeOptimizationDrug DiscoveryReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical DataBenchmark
π― What it does: Proposed the ATPO (Adaptive Tree Policy Optimization) algorithm, which optimizes information acquisition strategies in multi-round medical dialogues through uncertainty-guided tree search.
Attack-Resistant Watermarking for AIGC Image Forensics via Diffusion-based Semantic Deflection
Qingyu Liu (Zhejiang University), Zhibo Wang (Zhejiang University)
CodeAnomaly DetectionSafty and PrivacyDiffusion modelImage
π― What it does: Proposed a native watermarking framework PAI that requires no training and can be directly embedded into diffusion models for AIGC image copyright protection, ownership verification, attack detection, and semantic-level tampering localization.
Attention as a Compass: Efficient Exploration for Process-Supervised RL in Reasoning Models
Runze Liu (Tsinghua University), Kun Gai (Kuaishou Technology)
CodeReinforcement LearningTextBenchmark
π― What it does: Propose the AttnRL framework, which utilizes attention scores to guide process supervision reinforcement learning for efficient exploration and training in large reasoning models.
Attributing Response to Context: A JensenβShannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented Generation
Ruizhe Li (University of Aberdeen), Emine Yilmaz (University College London)
CodeExplainability and InterpretabilityComputational EfficiencyTextRetrieval-Augmented Generation
π― What it does: Proposes an inference-time context attribution method called ARC-JSD based on Jensen-Shannon Divergence, which can quickly identify the most critical retrieved sentences for generation results in retrieval-augmented generation (RAG), and applies this method to interpret the internal mechanisms of RAG.
π― What it does: Propose a lightweight framework AttriCtrl that can control the aesthetic attribute intensity of diffusion models based on numerical values.
ATTS: Asynchronous Test-Time Scaling via Conformal Prediction
Jing Xiong (University of Hong Kong), Ngai Wong (University of Hong Kong)
CodeComputational EfficiencyText
π― What it does: Proposes the ATTS (Asynchronous Test-Time Scaling) framework, which accelerates large-scale LLM inference through asynchronous inference and appropriate rejection sampling.
Auditing Black-Box LLM APIs with a Rank-Based Uniformity Test
Xiaoyuan Zhu (University of Southern California), Willie Neiswanger (University of Southern California)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
π― What it does: To address the model replacement issue in black-box LLM APIs, the Rank-Based Uniformity Test (RUT) is proposed for model equivalence testing.
AUHead: Realistic Emotional Talking Head Generation via Action Units Control
Jiayi Lyu (University of the Chinese Academy of Sciences), Tat-Seng Chua (National University of Singapore)
CodeGenerationTransformerLarge Language ModelDiffusion modelVideoMultimodalityChain-of-ThoughtAudio
π― What it does: This paper proposes a two-stage audio-driven speaker facial animation generation framework called AUHead. It first extracts fine-grained action unit (AU) sequences from speech using an audio language model (ALM), and then uses these AUs as conditional inputs to a diffusion model to generate speaker videos that are emotionally expressive, identity-preserving, and precisely synchronized.
Aurelius: Relation Aware Text-to-Audio Generation At Scale
Yuhang He (Microsoft Research), Vibhav Vineet (Microsoft Research)
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelFlow-based ModelTextAudio
π― What it does: This paper proposes the Aurelius framework, which constructs a large-scale audio event set (AudioEventSet) and relation set (AudioRelSet) to provide data and evaluation platforms for relation-aware text-to-audio generation.
Aurora: Towards Universal Generative Multimodal Time Series Forecasting
Xingjian Wu (East China Normal University), Chenjuan Guo (East China Normal University)
CodeGenerationTransformerVision Language ModelFlow-based ModelMultimodalityTime SeriesBenchmark
π― What it does: Built a cross-domain multimodal time series foundation model Aurora, supporting zero-shot prediction and generative probabilistic prediction.
Auto-RT: Automatic Jailbreak Strategy Exploration for Red-Teaming Large Language Models
yanjiang liu, Le Sun (University of Chinese Academy of Sciences)
CodeAdversarial AttackTransformerLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: Proposes the AUTO-RT framework, which leverages reinforcement learning to automatically explore jailbreak strategies in large language models, achieving higher exploitability and severity in vulnerability discovery.
π― What it does: Proposes AutoDA-Timeseries, a general-purpose automatic data augmentation framework for time series that can learn and optimize time series augmentation strategies in a single-stage, end-to-end manner;
AutoDV: An End-to-End Deep Learning Model for High-Dimensional Data Visualization
Wei Dai (Zhejiang University of Technology), Jicong Fan (Chinese University of Hong Kong)
CodeComputational EfficiencyRepresentation LearningData-Centric LearningGraph Neural NetworkTransformerImageTabularBiomedical Data
π― What it does: Propose an end-to-end high-dimensional data visualization model AutoDV, which directly outputs 2D/3D embeddings using graph Transformer and invariant loss without parameter tuning or retraining.
Autoencoding-Free Context Compression for LLMs via Contextual Semantic Anchors
Xin Liu (Northeastern University), JingBo Zhu
CodeCompressionRepresentation LearningTransformerLarge Language ModelText
π― What it does: Propose a non-autoencoding context compression method called Semantic-Anchor Compression (SAC), which selects anchor words directly from the original context and uses bidirectional attention to aggregate global information into the key-value representations of these anchor words;
AutoEP: LLMs-Driven Automation of Hyperparameter Evolution for Metaheuristic Algorithms
Zhenxing Xu (National University of Defense Technology), Ji Wang (Peking University)
CodeOptimizationHyperparameter SearchTransformerLarge Language ModelChain-of-Thought
π― What it does: Designed the AutoEP framework, leveraging large language models (LLM) and real-time exploratory landscape analysis (ELA) to achieve zero-training dynamic hyperparameter tuning;
Automated Formalization via Conceptual Retrieval-Augmented LLMs
Wangyue Lu (Northeastern University), Ge Yu (Northeastern University)
CodeAI Code AssistantTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Proposed and implemented the CRAMF (Concept-driven Retrieval-Augmented Mathematical Formalization) framework, enhancing the performance of LLMs in automated formalization by retrieving formal definitions of mathematical concepts.
Automated Stateful Specialization for Adaptive Agent Systems
Myan Vu (University of Auckland), Mayank Goel (University of Auckland)
CodeLarge Language ModelAgentic AIText
π― What it does: Developed the ASPEC framework to realize stateful lifecycle management of expert agents, including automatically evolved discovery of expert agents, experience-based cultivation, and efficient reuse of expert knowledge across queries through a 'retain-then-escalate' control strategy.
π― What it does: This paper proposes an end-to-end generative model called Skip-BART, which directly generates the hue and brightness of stage lighting from music, achieving complete automatic control from music to lighting;
AutoMetrics: Approximate Human Judgments with Automatically Generated Evaluators
Michael J Ryan, Diyi Yang (Stanford University)
CodeExplainability and InterpretabilityData-Centric LearningLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: The AutoMetrics framework automatically generates interpretable, task-specific evaluation metrics using a small amount of human feedback, and optimizes them through regression to be highly correlated with human judgments.
AutoQD: Automatic Discovery of Diverse Behaviors with Quality-Diversity Optimization
Saeed Hedayatian (University of Southern California), Stefanos Nikolaidis (University of Southern California)
CodeOptimizationReinforcement LearningBenchmark
π― What it does: Developed AutoQD, which automatically generates behavioral descriptors and combines them with QD algorithms to discover diverse and high-quality reinforcement learning strategies.
Autoregressive Models Rival Diffusion Models at ANY-ORDER Generation
Tianqi Du (Peking University), Yisen Wang (Peking University)
CodeGenerationTransformerLarge Language ModelDiffusion modelText
π― What it does: Proposed the A3 framework, which enables arbitrary order and subset autoregressive modeling, and implemented two-stream attention networks and progressive training;
Autoregressive-based Progressive Coding for Ultra-Low Bitrate Image Compression
Ziyuan Zhang (Tsinghua University), Han Qiu (Tsinghua University)
CodeCompressionVision Language ModelImage
π― What it does: Proposed an autoregressive progressive coding (ARPC) based on the visual autoregressive model (VAR) for ultra-low bitrate image compression;
AutoSP: Unlocking Long-Context LLM Training Via Compiler-Based Sequence Parallelism
Ahan Gupta (University of Illinois Urbana-Champaign), Minjia Zhang (University of Illinois Urbana-Champaign)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose AutoSP, an automatic sequence parallelism and activation checkpointing optimization based on the PyTorch 2.0 compiler, to increase the trainable length of long-context LLMs.
π― What it does: Conduct a large-scale empirical study on training methods for animal vocalization encoders, evaluating the impact of different models, data mixing strategies, and training protocols on multi-task performance (classification, detection, individual identification, spectrogram discovery), and propose an optimal training workflow.
π― What it does: Investigate the gradient geometry of diffusion models, proposing a continual learning framework that combines rank-1 Fisher-based EWC with generative replay.
CodeAnomaly DetectionTransformerLarge Language ModelTextBenchmark
π― What it does: Propose a training-agnostic, weight matrix-based LLM fingerprinting method that can detect whether a model is based on an existing base model.
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelText
π― What it does: Proposes BA-LoRA, a parameter-efficient fine-tuning method that mitigates 'catastrophic forgetting' during the transfer of large language models through output space regularization.
BAPO: Stabilizing Off-Policy Reinforcement Learning for LLMs via Balanced Policy Optimization with Adaptive Clipping
Zhiheng Xi (Fudan University), Xuanjing Huang (Fudan University)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningText
π― What it does: Proposed and validated the BAPO (Balanced Policy Optimization with Adaptive Clipping) algorithm, specifically addressing gradient instability, sharp entropy decline, and training collapse caused by data obsolescence in large language models during offline reinforcement learning.
π― What it does: This paper formally defines and analyzes the 'loss of plasticity' (LoP) problem in deep networks within non-stationary environments through dynamical systems theory, proposing that LoP can be viewed as gradient descent being trapped in low-dimensional invariant manifolds in the parameter space. It identifies two types of traps: frozen units and cloned units, and proves that gradient optimization cannot escape from them. Meanwhile, it reveals the tension between low-rank feature compression and plasticity loss. Finally, experiments are conducted to verify the theory and explore methods such as normalization, noise injection, and Dropout to escape or prevent LoP.
π― What it does: Propose a batch pruning method (B-PAS) based on activation stability, dynamically identifying and removing batches whose learning contributions have approached zero during training to accelerate training.
Bayesian Attention Mechanism: A Probabilistic Framework for Positional Encoding and Context Length Extrapolation
Arthur S. Bianchessi (Pontificia Universidade Catolica do Rio Grande do Sul), Lucas S. KupssinskΓΌ (Pontificia Universidade Catolica do Rio Grande do Sul)
CodeRepresentation LearningTransformerLarge Language ModelTextBenchmark
π― What it does: Propose the Bayesian Attention Mechanism (BAM), treating positional encoding as a Bayesian prior, and introduce a learnable Generalized Gaussian Prior (GGD-BAM) for long-context inference.
Bayesian Post Training Enhancement of Regression Models with Calibrated Rankings
Kevin Tirta Wijaya (Max Planck Institute for Informatics), Vahid Babaei (Fraunhofer SCAI)
CodeOptimizationData-Centric LearningDrug DiscoveryLarge Language ModelTabularBiomedical DataAgriculture Related
π― What it does: By building upon existing regression models and employing Bayesian inference to integrate the Gaussian likelihood of the regressor with the Bradley-Terry likelihood equipped with temperature calibration and soft gating, prediction accuracy is enhanced without requiring retraining.
π― What it does: Designed and implemented a robust collaborative multi-agent reinforcement learning framework based on Bayesian decision-making (BATPAL), which discretizes continuous attack types into subsets based on severity against baseline policies, and uses external constraint PPO to learn adversarial policies, ultimately achieving an adaptive robust policy against unknown attacks.
π― What it does: Proposed a quasi-dynamic program embedding method (Behavioral Embeddings), which applies a set of designed optimization probes on program LLVM IR to quantify the program's response to different optimization sequences, forming a behavioral spectrum. Continuous response vectors are discretized using Product Quantization, and a multi-task Transformer model, PQ-BERT, is then used to learn deep syntax, ultimately obtaining program representations applicable for compiler optimization prediction.
Benchmarking Bias Mitigation Toward Fairness Without Harm from Vision to LVLMs
Xuwei Tan (Ohio State University), Xueru Zhang (Ohio State University)
CodeData-Centric LearningVision Language ModelImageMultimodalityBenchmark
π― What it does: Proposed the NH-Fair benchmark, which evaluates the performance of visual and multimodal models in fairness without harm using a unified experimental protocol.
π― What it does: Systematic evaluation of 26 clinical tasks across 12 public databases, comparing the performance of 8 ECG baseline models with two supervised baselines.
π― What it does: Proposed the FG-BMK fine-grained vision-language model evaluation benchmark, containing 1.01 million questions and 280,000 images, covering two evaluation paradigms: human-oriented and machine-oriented;
π― What it does: Proposed a multi-turn, multi-step tool usage evaluation benchmark called WildToolBench based on real user behavior, and systematically assessed the tool usage capabilities of 57 mainstream LLMs.
Cristina GonzΓ‘lez (Universidad de los Andes), Pablo Arbelaez
CodeSegmentationLarge Language ModelVision Language ModelContrastive LearningImageMultimodalityBenchmark
π― What it does: Proposed a dedicated evaluation protocol for open segmentation and rebaselined existing methods; simultaneously developed the first multimodal large language model OPAL that uses contrastive learning to align visual regions with text descriptions.
Benchmarking Stochastic Approximation Algorithms for Fairness-Constrained Training of Deep Neural Networks
Andrii Kliachkin (Czech Technical University in Prague), Jakub Marecek (Czech Technical University in Prague)
CodeOptimizationTabularBenchmark
π― What it does: Propose a fairness-constrained deep learning training benchmark based on US census data, and implement and compare several stochastic approximation algorithms.
π― What it does: Proposed a full-binary error propagation algorithm (BEP) to enable end-to-end training of multi-layer binary neural networks (MLP and RNN).
Best-of-Infinity: Asymptotic Performance of Test-Time LLM Ensembling
Junpei Komiyama (Mohamed bin Zayed University of Artificial Intelligence), Masafumi Oyamada (NEC Corporation)
CodeOptimizationTransformerLarge Language ModelText
π― What it does: Investigated and implemented the limit (best-of-β) performance of majority voting based on large language models (LLM), and proposed adaptive sampling methods and weighted LLM integration strategies to approach this limit under limited inference budgets.
David Woodruff, Samson Zhou (Texas A&M University)
CodeOptimizationBenchmark
π― What it does: Studied the distributed expert online learning problem under a coordinator model, proposing a communication-reward trade-off protocol for general βp loss, achieving near-optimal expected reward while significantly reducing communication volume.
Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs
Mohammad Tavakoli (University of Alberta), J Ross Mitchell
CodeLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose the BEAM benchmark and the LIGHT framework to evaluate and enhance the memory and reasoning capabilities of large language models (LLMs) in extremely long conversations.
Beyond Accuracy: Are Time Series Foundation Models Well-Calibrated?
Coen Adler (University of California), Padhraic Smyth (University of California)
CodeTransformerTime SeriesBenchmark
π― What it does: This study systematically evaluates the probability calibration performance of current mainstream time series foundation models (TSFM) in zero-shot prediction scenarios.
Beyond Aggregation: Guiding Clients in Heterogeneous Federated Learning
Zijian Wang (Renmin University of China), Qiong Zhang (Renmin University of China)
CodeFederated LearningConvolutional Neural NetworkImageBiomedical Data
π― What it does: Proposes FedDRM, a new framework in federated learning that shifts the server's role from passive aggregation to active guidance, learning local models on each client while routing new tasks to the most suitable clients based on their features.
Beyond Fixed: Training-Free Variable-Length Denoising for Diffusion Large Language Models
Jinsong Li (Chinese University of Hong Kong), Dahua Lin (Chinese University of Hong Kong)
CodeGenerationAI Code AssistantTransformerLarge Language ModelDiffusion modelText
π― What it does: Proposes a training-agnostic two-stage dynamic length extension method called DAEDAL to address the limitations of fixed-length generation in Diffusion LLM;
π― What it does: Proposed an implicit neural response function (NRF) based on coordinates, modeling fMRI brain signals as a continuous function of visual stimuli and MNI spatial coordinates, enabling cross-subject transfer and arbitrary resolution queries.
Beyond Linear Probes: Dynamic Safety Monitoring for Language Models
James Oldfield (Queen Mary University of London), Fazl Barez (University of Oxford)
CodeSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyText
π― What it does: Proposed a Truncated Polynomial Classifier (TPC) for dynamic safety monitoring in the activation space of large language models, supporting variable computational budgets and input-driven cascading evaluations;
π― What it does: Propose a new spiking neuron model DLIF, incorporating a biologically validated two-linear dendritic integration rule, providing theoretical proof and numerical validation.
Beyond Magic Words: Sharpness-Aware Prompt Evolving for Robust Large Language Models with TARE
Guancheng Wan (University of California Los Angeles), Wei Wang (University of California Los Angeles)
CodeOptimizationLarge Language ModelPrompt EngineeringText
π― What it does: Propose a black-box gradient-free prompt search framework TARE and its adaptive version ATARE based on text sharpness, aiming to find efficient and robust prompts within the semantic neighborhood.
Beyond Magnitude: Leveraging Direction of RLVR Updates for LLM Reasoning
Kexin Huang (University of Science and Technology of China), Jingren Zhou (Independent Researcher)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: Investigated the importance of update direction in RLVR and proposed using token-level logp differences (Ξlogβ p) to diagnose and enhance LLM's reasoning capabilities.
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningTextGraph
π― What it does: Proposes the Structured Walk framework Walk2Pers, modeling user preference evolution through action-conditioned geometric steps and dual memory channels, and achieving personalized summarization based on this.