ICLR 2025 Papers — Page 3
International Conference on Learning Representations · 3704 papers
AI2TALE: An Innovative Information Theory-based Approach for Learning to Localize Phishing Attacks
Van Nguyen (Monash University), Carsten Rudolph (Monash University)
Anomaly DetectionExplainability and InterpretabilityText
🎯 What it does: This paper proposes a deep learning framework called AI2TALE, based on information theory and the information bottleneck, for locating and explaining phishing attacks in emails under weak supervision;
AIMS.au: A Dataset for the Analysis of Modern Slavery Countermeasures in Corporate Statements
Adriana Eufrosina Bora (Mila Quebec AI Institute), Kerrie Mengersen (Queensland University of Technology)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The largest modern slavery statement sentence-level annotated dataset AIMS.au has been constructed and made public, covering 5,731 statements and over 800,000 sentences; baseline experiments under zero-shot and supervised learning are also provided.
Aioli: A Unified Optimization Framework for Language Model Data Mixing
Mayee F Chen, Christopher Re
OptimizationLarge Language ModelText
🎯 What it does: A unified optimization framework called AIOLI is proposed for data mixing in language models, aimed at optimizing the combination ratio of training data to improve model performance.
Air Quality Prediction with Physics-Guided Dual Neural ODEs in Open Systems
Jindong Tian (East China Normal University), Bin Yang (East China Normal University)
Graph Neural NetworkContrastive LearningTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes Air-DualODE, which utilizes a physics-guided dual neural ODE model to simultaneously model physical (diffusion-convection) dynamics and data-driven unknown dynamics in an open air quality system, achieving accurate PM2.5 predictions through fusion and alignment in the latent space.
AIR-BENCH 2024: A Safety Benchmark based on Regulation and Policies Specified Risk Categories
Yi Zeng (Virginia Tech), Bo Li (University of Chicago)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: AIR-BENCH 2024 has been proposed and released—a language model safety benchmark based on 8 government regulations and 16 corporate policies, comprising a total of 314 fine-grained risk categories and 5,694 diverse prompts, aimed at unifying the assessment of model compliance with regulations and policies.
ALBAR: Adversarial Learning approach to mitigate Biases in Action Recognition
Joseph Fioresi (University of Central Florida), Mubarak Shah (University of Central Florida)
RecognitionAnomaly DetectionConvolutional Neural NetworkGenerative Adversarial NetworkContrastive LearningVideo
🎯 What it does: A label-free bias elimination framework based on adversarial learning, ALBAR, has been developed specifically for addressing background and foreground bias issues in video action recognition.
Alchemy: Amplifying Theorem-Proving Capability Through Symbolic Mutation
Shaonan Wu (Xi'an Jiaotong University), Ping Wei (Xi'an Jiaotong University)
Large Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: By performing symbolic mutation on existing theorems in the Lean theorem proving environment (using the tactics rw and apply), approximately 6.3 million new formal theorems were generated, and these theorems were constructed into a large-scale corpus along with their corresponding proofs for continuous pre-training and supervised fine-tuning of LLMs, thereby enhancing their performance in theorem proving tasks.
Algorithmic Stability Based Generalization Bounds for Adversarial Training
Runzhi Tian (University of Ottawa), Yongyi Mao (University of Ottawa)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper analyzes the algorithm stability of adversarial training and provides a general upper bound on the generalization error based on a perturbation mapping's scalability parameter.
Aligned Better, Listen Better for Audio-Visual Large Language Models
Yuxin Guo (University of Chinese Academy of Sciences), Wei Zou (University of Chinese Academy of Sciences)
ClassificationRecognitionTransformerLarge Language ModelVision Language ModelVideoMultimodalityAudio
🎯 What it does: Dolphin is proposed, a large language model for audio and video that achieves fine-grained alignment in both space and time, and constructs a 5.24M question-answer aligned AVU dataset, enhancing audio and video understanding and dialogue capabilities.
Aligned Datasets Improve Detection of Latent Diffusion-Generated Images
Anirudh Sundara Rajan (University of Wisconsin Madison), Yong Jae Lee (University of Wisconsin Madison)
GenerationDiffusion modelAuto EncoderImage
🎯 What it does: Using the LDM's autoencoder for single-step reconstruction of real images, generating forged images that are highly aligned with real images in terms of size, semantics, tone, etc., and then training a fake image detector with these aligned data.
Aligned LLMs Are Not Aligned Browser Agents
Priyanshu Kumar (Carnegie Mellon University), Zifan Wang (Scale AI)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This study investigates how the security of LLMs deployed as browser proxies declines when trained to reject harmful instructions, and constructs the BrowserART red team toolkit to evaluate browser proxies.
Aligning Generative Denoising with Discriminative Objectives Unleashes Diffusion for Visual Perception
Ziqi Pang (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
SegmentationGenerationDepth EstimationDiffusion modelImageMultimodality
🎯 What it does: In response to the application of diffusion models in visual perception tasks, a method is proposed to align the generative denoising process through learning objectives, training data, and interactive user interfaces, significantly improving the performance of depth estimation, referential image segmentation, and general perception tasks.
Aligning Human Motion Generation with Human Perceptions
Haoru Wang (Peking University), Yizhou Wang (Peking University)
GenerationData SynthesisPose EstimationTransformerDiffusion modelContrastive LearningVideoMultimodality
🎯 What it does: This paper proposes a motion quality assessment framework based on large-scale human perception evaluation data, constructs the MotionPercept dataset, and trains the MotionCritic model for automated evaluation and enhancement of human motion generation quality.
Aligning Language Models with Demonstrated Feedback
Omar Shaikh (Stanford University), Diyi Yang (Stanford University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: The DITTO method is proposed, which directly generates preference comparison data for language models through a small number (<10) of user demonstrations, achieving personalized alignment.
Aligning Visual Contrastive learning models via Preference Optimization
Amirabbas Afzali (Massachusetts Institute of Technology), Simon Gottschalk (Leibniz University Hannover)
RetrievalOptimizationAdversarial AttackTransformerContrastive LearningImage
🎯 What it does: By applying preference optimization (PO) methods (DPO, IPO, KTO) to contrastive learning models, this paper achieves behavior alignment and robustness enhancement of image-text retrieval models (such as CLIP) in scenarios such as adversarial text attacks and gender bias.
ALLaM: Large Language Models for Arabic and English
M Saiful Bari (National Center for AI), Haidar Khan (National Center for AI)
TransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: By extending the tokenizer and vocabulary of Llama-2 and continuing pre-training on a large-scale English-Arabic mixed corpus, or training from random initialization on 4T English followed by mixing in 1.2T English-Arabic corpus, the ALLaM series (7B, 13B, 34B, 70B) large language models were constructed. Subsequently, SFT and DPO were used for alignment and preference training.
Almost Optimal Batch-Regret Tradeoff for Batch Linear Contextual Bandits
Zihan Zhang (University of Washington), Yuan Zhou (Tsinghua University)
Reinforcement Learning
🎯 What it does: This study investigates the optimal batch-regret trade-off for batch linear contextual multi-armed bandits and presents a batch learning algorithm that achieves approximately optimal regret.
AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models
Junfeng Fang (University of Science and Technology of China), Tat-Seng Chua (National University of Singapore)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes AlphaEdit, a method for editing language model knowledge by projecting parameter perturbations into the null space that retains knowledge, balancing knowledge updating and retention;
Ambient Diffusion Posterior Sampling: Solving Inverse Problems with Diffusion Models Trained on Corrupted Data
Asad Aali (Stanford University), Jon Tamir
RestorationCompressionDiffusion modelImageMagnetic Resonance Imaging
🎯 What it does: This paper proposes Ambient Diffusion Posterior Sampling (A-DPS), which utilizes a diffusion model trained on linearly corrupted data for posterior sampling to solve inverse problems.
Amortized Control of Continuous State Space Feynman-Kac Model for Irregular Time Series
Byoungwoo Park (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
ClassificationOptimizationComputational EfficiencyRecurrent Neural NetworkTransformerTime SeriesBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies a continuous state space model capable of handling irregular time series—ACSSM. It proposes efficient simulation and inference of posterior trajectories through multi-boundary Doob’s h-transformation and variational inference (SOC).
Amulet: ReAlignment During Test Time for Personalized Preference Adaptation of LLMs
Zhaowei Zhang (Peking University), Yaodong Yang (Peking University)
Recommendation SystemOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: During the inference phase of LLM, the decoding process of each token is treated as an online learning problem, utilizing user-provided preference prompts to achieve real-time adaptive alignment without training.
An Asynchronous Bundle Method for Distributed Learning Problems
Daniel Cederberg (Stanford University), Mikael Johansson (KTH)
OptimizationFederated LearningTabular
🎯 What it does: An asynchronous Bundle method is proposed, utilizing the multi-tangent model of each worker node to solve subproblems on the parameter server, thereby achieving iteration updates in distributed learning without synchronization and without the need for maximum delay information.
An Auditing Test to Detect Behavioral Shift in Language Models
Leo Richter (University College London), Matt Kusner
Large Language ModelText
🎯 What it does: A sequential hypothesis testing method is proposed for monitoring changes in language model behavior, capable of detecting shifts in model behavior distribution over time under black-box access.
An Effective Manifold-based Optimization Method for Distributionally Robust Classification
Jiawei Huang (University of Science and Technology of China), Hu Ding (University of Science and Technology of China)
ClassificationOptimizationContrastive LearningImageBiomedical Data
🎯 What it does: A distributionally robust optimization method based on data manifolds (MWDRO) is proposed, which automatically extracts the tangent space of the manifold using a game mechanism of contrastive learning and Jacobian regularization, thereby achieving distributionally robust training in uncertain sets defined by geometric Wasserstein distance.
An Effective Theory of Bias Amplification
Arjun Subramonian (University of California Los Angeles), Elvis Dohmatob (Meta)
OptimizationData-Centric LearningGaussian SplattingTabular
🎯 What it does: This paper constructs an accurate theoretical model within the framework of high-dimensional ridge regression (including random projection) to analyze how different model designs and data distribution properties lead to bias amplification and minority group bias.
An Efficient Framework for Crediting Data Contributors of Diffusion Models
MingYu Lu, Su-In Lee (University of Washington)
Computational EfficiencyData-Centric LearningSupervised Fine-TuningDiffusion modelImage
🎯 What it does: This paper proposes an efficient framework that estimates the Shapley values of contributors to diffusion models using sparse fine-tuning, aiming for fair attribution of data contributors.
An Empirical Analysis of Uncertainty in Large Language Model Evaluations
Qiujie Xie (Zhejiang University), Linyi Yang (University College London)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper conducts a systematic empirical analysis of uncertainty in the evaluation of large language models and proposes the Uncertainty Perception Evaluator ConfiLM to enhance evaluation performance in out-of-distribution (OOD) scenarios.
An Engorgio Prompt Makes Large Language Model Babble on
Jianshuo Dong (Tsinghua University), Han Qiu (Tsinghua University)
Computational EfficiencyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A cost attack named Engorgio is proposed and implemented for autoregressive large language models, which generates malicious prompts to induce the model to output abnormally long texts, thereby increasing inference time and computational resource consumption.
An Evolved Universal Transformer Memory
Edoardo Cetin (Sakana AI), Yujin Tang (Sakana AI)
TransformerVideoText
🎯 What it does: This study proposes Neural Attention Memory Models (NAMMs), which adaptively prune the KV cache of the Transformer through evolutionary learning to enhance the performance and efficiency of long-context tasks.
An Exploration with Entropy Constrained 3D Gaussians for 2D Video Compression
Xiang Liu (Tsinghua University), Shu-Tao Xia (Tsinghua University)
CompressionGaussian SplattingOptical FlowVideo
🎯 What it does: Applying 3D Gaussian Splatting to video compression, we propose the Toast-Like Sliding Window (TSW) orthogonal projection and construct an end-to-end video compression framework GSVC that supports streaming decoding.
An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels
Duy Kien Nguyen, Xinlei Chen (Meta AI)
ClassificationGenerationTransformerDiffusion modelImage
🎯 What it does: The study directly applies Transformer to each pixel, eliminating the traditional 16×16 patch division and locality prior, and validates its feasibility on three major tasks: classification, regression, and unsupervised pre-training, as well as diffusion model generation.
An Information Criterion for Controlled Disentanglement of Multimodal Data
Chenyu Wang (Massachusetts Institute of Technology), Caroline Uhler (Massachusetts Institute of Technology)
Explainability and InterpretabilityRepresentation LearningDrug DiscoveryContrastive LearningMultimodalityBenchmark
🎯 What it does: This paper proposes a self-supervised multimodal representation learning framework DISENTANGLEDSSL, aimed at decoupling shared information from modality-specific information to enhance interpretability and robustness.
An Intelligent Agentic System for Complex Image Restoration Problems
Kaiwen Zhu (Shanghai Jiao Tong University), Chao Dong (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
RestorationTransformerLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelImage
🎯 What it does: Proposes the AgenticIR agent system, which achieves perception-scheduling-execution-reflection-rescheduling in five stages through the interaction of LLM and VLM, dynamically calling a single degradation recovery tool to solve complex image restoration tasks.
An Online Learning Theory of Trading-Volume Maximization
Tommaso Cesari (University of Ottawa), Roberto Colomboni (Politecnico di Milano)
OptimizationFinance Related
🎯 What it does: This paper introduces a new objective in the OTC market - maximizing trading volume, and studies how brokers set quotes to achieve this goal within an online learning framework.
An Optimal Discriminator Weighted Imitation Perspective for Reinforcement Learning
Haoran Xu (University of Texas at Austin), Amy Zhang (UMass Amherst)
Reinforcement LearningTabular
🎯 What it does: A new reinforcement learning method called Iterative Dual Reinforcement Learning (IDRL) is proposed, which addresses reinforcement learning problems from the perspective of optimal discriminator weighted imitation learning.
An Undetectable Watermark for Generative Image Models
Sam Gunn (University of California Berkeley), Dawn Song (University of California Berkeley)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: The first undetectable watermark scheme (PRC watermark) is proposed, which can embed watermarks in the latent space of diffusion models without affecting image quality and achieve robust detection.
ANaGRAM: A Natural Gradient Relative to Adapted Model for efficient PINNs learning
Nilo Schwencke (INRIA Saclay), Cyril Furtlehner (INRIA Saclay)
OptimizationComputational EfficiencyTabularBenchmarkPhysics Related
🎯 What it does: ANaGRAM is proposed, a low-complexity natural gradient optimization method in PINNs, treating PINNs as least squares regression and proving that the natural gradient is equivalent to the Green's function;
AnalogGenie: A Generative Engine for Automatic Discovery of Analog Circuit Topologies
Jian Gao (Northeastern University), Xuan Zhang (Northeastern University)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningGraph
🎯 What it does: This paper presents AnalogGenie—a GPT-based generative engine for the automatic discovery of multi-type, scalable analog circuit topologies.
Analysis of Linear Mode Connectivity via Permutation-Based Weight Matching: With Insights into Other Permutation Search Methods
Akira Ito (NTT Social Informatics Laboratories), Atsutoshi Kumagai (NTT Computer and Data Science Laboratories)
Convolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper analyzes the search for permutations between models through Weight Matching (WM) to achieve Linear Mode Connectivity (LMC), and explores its impact on model merging, Activation Matching (AM), and the Straight-Through Estimator (STE).
Analytic DAG Constraints for Differentiable DAG Learning
Zhen Zhang (Australian Institute for Machine Learning), Javen Qinfeng Shi (Australian Institute for Machine Learning)
OptimizationGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper proposes the use of analytic functions to construct differentiable DAG constraints, establishing a theoretical connection between analytic functions and DAG constraints. It utilizes the closure properties of analytic functions (differentiation, addition, multiplication) to design a series of high-order DAG constraints. Additionally, efficient algorithms for evaluating these constraints are provided and implemented within a path tracking optimization framework.
Analyzing and Boosting the Power of Fine-Grained Visual Recognition for Multi-modal Large Language Models
Hulingxiao He (Peking University), Yuxin Peng (Peking University)
RecognitionTransformerLarge Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a multimodal large language model named Finedefics, specifically improved for fine-grained visual recognition (FGVR);
Analyzing Neural Scaling Laws in Two-Layer Networks with Power-Law Data Spectra
Roman Worschech (Institut für Theoretische Physik, Universität Leipzig), Bernd Rosenow (Institut für Theoretische Physik, Universität Leipzig)
Tabular
🎯 What it does: Under the student-teacher framework, a theoretical analysis of two-layer networks is conducted using statistical mechanics methods, resulting in an analytical expression for the generalization error when the data covariance spectrum follows a power-law distribution; the study also investigates the learning phase, plateau length, and the transition from exponential decay to power-law decay under linear and nonlinear activation functions.
AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents
Christopher Rawles (Google DeepMind), Oriana Riva (Google)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAgentic AIMultimodalityBenchmark
🎯 What it does: Created AndroidWorld, a reproducible and dynamically parameterizable automated agent evaluation environment based on Android, providing 116 tasks covering 20 real Android applications;
Animate Your Thoughts: Reconstruction of Dynamic Natural Vision from Human Brain Activity
Yizhuo Lu (Chinese Academy of Sciences), Huiguang He (Chinese Academy of Sciences)
GenerationData SynthesisTransformerDiffusion modelContrastive LearningVideoMultimodalityMagnetic Resonance Imaging
🎯 What it does: Designed and implemented the Mind-Animator two-stage model, which first decodes semantic, structural, and motion features from fMRI signals, and then inputs these three features into an inflated Stable Diffusion model to generate dynamic natural videos without external video training.
Animate-X: Universal Character Image Animation with Enhanced Motion Representation
Shuai Tan (Ant Group), Ming Yang (Ant Group)
GenerationPose EstimationTransformerDiffusion modelVideoBenchmark
🎯 What it does: A general character animation framework named Animate-X is proposed, capable of generating high-quality videos under reference images and target pose sequences, particularly supporting animations for human, anthropomorphized, and non-human characters.
AniSDF: Fused-Granularity Neural Surfaces with Anisotropic Encoding for High-Fidelity 3D Reconstruction
Jingnan Gao (Shanghai Jiao Tong University), Yichao Yan (Shanghai Jiao Tong University)
GenerationData SynthesisNeural Radiance FieldPoint CloudMesh
🎯 What it does: AniSDF is proposed, a neural SDF representation that integrates coarse and fine resolution with a physics-based dual radiance field for achieving high-fidelity 3D reconstruction and novel view rendering.
AnoLLM: Large Language Models for Tabular Anomaly Detection
Che-Ping Tsai (Amazon), Wei Ding (Amazon)
Anomaly DetectionTransformerLarge Language ModelTabular
🎯 What it does: Proposes AnoLLM, a framework that serializes tabular data into text and utilizes large language models for unsupervised anomaly detection.
Answer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questions
Sarah Wiegreffe (Allen Institute for AI), Ashish Sabharwal (Allen Institute for AI)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the answering mechanism of Transformer language models in formatted multiple-choice questions (MCQA), identifying the causal roles of key layers and attention heads, and revealing the timing of the model's learning of formatted answering through synthetic tasks.
Anti-Exposure Bias in Diffusion Models
Junyu Zhang (Central South University), Chang Xu (University of Sydney)
GenerationData SynthesisPrompt EngineeringDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: The exposure bias caused by training-sampling inconsistency in diffusion models is identified, and a lightweight prompt prediction model is proposed to generate bias-correcting prompts at each step to correct the sampling trajectory, thereby improving generation quality.
Any-step Dynamics Model Improves Future Predictions for Online and Offline Reinforcement Learning
Haoxin Lin (Nanjing University), Yang Yu (Nanjing University)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: Proposes an Arbitrary Step Dynamic Model (ADM) and uses this model in online (ADMPO-ON) and offline (ADMPO-OFF) reinforcement learning frameworks, significantly reducing multi-step cumulative errors.
Anyprefer: An Agentic Framework for Preference Data Synthesis
Yiyang Zhou (University of North Carolina at Chapel Hill), Huaxiu Yao (University of North Carolina at Chapel Hill)
Data SynthesisOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelTextMultimodalityMagnetic Resonance Imaging
🎯 What it does: Designed the Anyprefer framework for the automatic synthesis of high-quality preference data to enhance the alignment performance of foundational models with human preferences.
AnyTouch: Learning Unified Static-Dynamic Representation across Multiple Visuo-tactile Sensors
Ruoxuan Feng (Wuhan University of Science and Technology), Di Hu (Wuhan University of Science and Technology)
ClassificationRepresentation LearningContrastive LearningImageVideoTextMultimodality
🎯 What it does: The AnyTouch framework is proposed, which combines static images and dynamic videos to learn a unified multi-sensor tactile representation, and collects the TacQuad multimodal alignment dataset.
APE: Faster and Longer Context-Augmented Generation via Adaptive Parallel Encoding
Xinyu Yang (Carnegie Mellon University), Beidi Chen (Carnegie Mellon University)
GenerationTransformerTextRetrieval-Augmented Generation
🎯 What it does: This paper studies the application of parallel encoding in Context-Augmented Generation (CAG) and proposes the Adaptive Parallel Encoding (APE) method, which enables parallel encoding to align attention distribution, thereby maintaining performance close to that of sequential encoding.
API Pack: A Massive Multi-Programming Language Dataset for API Call Generation
Zhen Guo (Massachusetts Institute of Technology), Rameswar Panda (IBM)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper creates a large-scale multilingual API call instruction-code pair dataset named API Pack (over 1.1 million instances) and fine-tunes various LLMs using this dataset, significantly improving performance in generating new API calls.
Apollo-MILP: An Alternating Prediction-Correction Neural Solving Framework for Mixed-Integer Linear Programming
Haoyang Liu (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
OptimizationGraph Neural Network
🎯 What it does: The Apollo-MILP framework is proposed, which continuously improves prediction quality and gradually fixes reliable variables by alternately executing prediction (using GNN to predict variable values) and correction (using trust region search to obtain reference solutions) in each iteration, thus achieving efficient solving of mixed-integer linear programming problems.
Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding
Eric Lei (University of Pennsylvania), Shirin Saeedi Bidokhti (University of Pennsylvania)
CompressionConvolutional Neural NetworkFlow-based ModelImagePhysics Related
🎯 What it does: A nonlinear transform coding framework based on lattice quantization (Lattice Transform Coding, LTC) is proposed, addressing the suboptimal issues caused by scalar quantization in traditional neural compression.
Approximating Full Conformal Prediction for Neural Network Regression with Gauss-Newton Influence
Dharmesh Tailor (University of Amsterdam), Christos Louizos (Qualcomm AI Research)
Tabular
🎯 What it does: An Approximate Full Transductive Consistency Prediction (ACP-GN) is proposed, which can generate adaptive prediction intervals for regression tasks without retraining the neural network.
Approximation algorithms for combinatorial optimization with predictions
Antonios Antoniadis (University of Twente), Moritz Venzin (Bocconi University)
OptimizationGraph
🎯 What it does: This paper proposes a general framework for learning-enhanced approximation algorithms that can significantly improve the approximation ratio of classical combinatorial optimization algorithms while maintaining near-linear time complexity, even utilizing (or even infeasible) predictions;
ARB-LLM: Alternating Refined Binarizations for Large Language Models
Zhiteng Li (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a binary post-training quantization method for large language models called ARB-LLM, which utilizes Alternating Refinement Binarization (ARB), Calibration Data Augmentation (ARB-X), and Row-Column Bidirectional Scaling (ARB-RC), and improves weight grouping and Column Group Bitmap (CGB) to enhance quantization performance.
Are Large Vision Language Models Good Game Players?
Xinyu Wang (University of Adelaide), Qi Wu (University of Adelaide)
TransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper presents LVLM‑Playground—a game-based evaluation framework that comprehensively assesses the cognitive and reasoning abilities of visual language models through six types of games (chess, Sudoku, Minesweeper) across four tasks (perception, question answering, rule following, end-to-end gameplay).
Are Transformers Able to Reason by Connecting Separated Knowledge in Training Data?
Yutong Yin (Northwestern University), Zhaoran Wang (Northwestern University)
TransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: Designed the FTCT (Fragmented at Training, Chained at Testing) synthetic dataset, and trained a model using Transformer to verify its ability to combine fragmented knowledge pieces from training into a complete reasoning chain during testing.
Aria-MIDI: A Dataset of Piano MIDI Files for Symbolic Music Modeling
Louis Bradshaw (Queen Mary University of London), Simon Colton (Queen Mary University of London)
ClassificationSegmentationGenerationConvolutional Neural NetworkLarge Language ModelVideoAudio
🎯 What it does: The Aria-MIDI dataset was constructed using web crawling + language models, audio classifiers, and automated source separation + AMT processes, containing over 1 million piano MIDI files and approximately 100,000 hours of high-quality transcriptions.
Arithmetic Transformers Can Length-Generalize in Both Operand Length and Count
Hanseul Cho (KAIST), Chulhee Yun (KAIST)
Transformer
🎯 What it does: This paper studies the length generalization problem of Transformers in arithmetic tasks and proposes a combination of scratchpad and multi-layer Position Coupling to address the generalization challenges of multi-operand addition and double-length multiplication.
Arithmetic Without Algorithms: Language Models Solve Math with a Bag of Heuristics
Yaniv Nikankin (Technion - Israel Institute of Technology), Yonatan Belinkov (Technion - Israel Institute of Technology)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates how large language models work in arithmetic reasoning tasks, revealing that they do not learn complete algorithms or merely memorize answers, but instead achieve arithmetic calculations through a set of sparse MLP neurons that collaboratively implement multiple heuristics.
ARLON: Boosting Diffusion Transformers with Autoregressive Models for Long Video Generation
Zongyi Li (Huazhong University of Science and Technology), Furu Wei (Microsoft Corporation)
GenerationData SynthesisTransformerDiffusion modelVideoText
🎯 What it does: The ARLON framework is constructed, generating coarse-grained video discrete codes through an autoregressive Transformer, and then utilizing the Diffusion Transformer (DiT) to produce high-quality long videos, achieving efficient generation from text to long video.
Articulate-Anything: Automatic Modeling of Articulated Objects via a Vision-Language Foundation Model
Long Le (University of Pennsylvania), Eric Eaton (University of Pennsylvania)
GenerationRetrievalRobotic IntelligenceTransformerReinforcement LearningAgentic AIVision Language ModelImageVideoTextMesh
🎯 What it does: Automatically generate interactive 3D digital twin models from text, images, or videos, achieving full automation of object assembly.
Artificial Kuramoto Oscillatory Neurons
Takeru Miyato (University of Tubingen), Max Welling (University of Amsterdam)
Object DetectionRepresentation LearningAdversarial AttackConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper proposes and implements Artificial Kuramoto Oscillating Neurons (AKOrN), embedding multidimensional vectorized Kuramoto synchronization dynamics into conventional network layers such as convolution and attention, to construct an iterative dynamic network.
As Simple as Fine-tuning: LLM Alignment via Bidirectional Negative Feedback Loss
Xin Mao (Nanyang Technological University), Anh Tuan Luu (Nanyang Technological University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a new loss function for value alignment of large language models—Bidirectional Negative Feedback (BNF) loss, aimed at addressing the hyperparameter sensitivity and instability issues of traditional DPO series methods.
Ask, and it shall be given: On the Turing completeness of prompting
Ruizhong Qiu (University of Illinois Urbana Champaign), Hanghang Tong (University of Illinois Urbana Champaign)
TransformerPrompt EngineeringChain-of-Thought
🎯 What it does: Proves that a single finite-size Transformer combined with prompts can achieve any computable function, achieving Turing completeness.
AssembleFlow: Rigid Flow Matching with Inertial Frames for Molecular Assembly
Hongyu Guo (National Research Council Canada University of Ottawa), Shengchao Liu (Université de Montréal)
GenerationData SynthesisOptimizationGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: A generative model named AssembleFlow has been developed to maintain molecular rigidity during the molecular assembly process, utilizing an inertial frame and SE(3) transformations to achieve rigid translations and rotations.
Associative memory and dead neurons
Vladimir Fanaskov (AIRI Skolkovo Institute of Science and Technology), Ivan Oseledets (AIRI Skolkovo Institute of Science and Technology)
🎯 What it does: This study investigates the vulnerability of dead neurons in the associative memory model based on the Krotov & Hopfield framework, and proposes an improved dynamical system and energy function to eliminate flat energy directions and enhance the reliability of stability analysis.
ASTrA: Adversarial Self-supervised Training with Adaptive-Attacks
Prakash Chandra Chhipa (Lulea Tekniska Universitet), Marcus Liwicki (University of Central Florida)
Representation LearningAdversarial AttackConvolutional Neural NetworkReinforcement LearningContrastive LearningImage
🎯 What it does: A self-supervised adaptive adversarial training framework, ASTrA, has been designed and validated. It utilizes a learnable attack strategy network to dynamically generate the optimal PGD attack for each sample through a reward mechanism in a single training phase, thereby enhancing self-supervised adversarial robustness.
AstroCompress: A benchmark dataset for multi-purpose compression of astronomical data
Tuan Truong (University of California Irvine), Joshua S. Bloom (University of California Berkeley)
CompressionDiffusion modelImageBenchmark
🎯 What it does: This study proposes the AstroCompress dataset and benchmark, aimed at advancing lossless neural compression techniques for astronomical images; it also compares the compression rates and runtime performance of various traditional and neural compression algorithms.
Asymmetric Factorized Bilinear Operation for Vision Transformer
Junjie Wu (Tianjin University), Qinghua Hu (Tianjin University)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: An Asymmetric Factorized Bilinear Operation (AFBO) is proposed to replace the Feed-Forward Network (FFN) in Vision Transformers (ViT), achieving a better performance-complexity trade-off through second-order statistical spatial-channel factorized bilinear operations (SCFBO) and structurally sparse channel mapping.
Asymptotic Analysis of Two-Layer Neural Networks after One Gradient Step under Gaussian Mixtures Data with Structure
Samet Demir (Koc University), Zafer Dogan (Koc University)
OptimizationGenerative Adversarial NetworkImage
🎯 What it does: This study investigates the training and generalization performance of a two-layer neural network on Gaussian mixture data after one step of gradient descent in high-dimensional high-ratio limits.
Asynchronous Federated Reinforcement Learning with Policy Gradient Updates: Algorithm Design and Convergence Analysis
Guangchen Lan (Purdue University), Christopher Brinton
Federated LearningReinforcement LearningSequential
🎯 What it does: An asynchronous federated reinforcement learning framework AFedPG is proposed, which updates the global policy through multi-agent collaboration using policy gradients.
Asynchronous RLHF: Faster and More Efficient Off-Policy RL for Language Models
Michael Noukhovitch (Mila Quebec AI Institute), Aaron Courville (Mila Quebec AI Institute)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes an asynchronous RLHF framework that first uses LLM to generate samples and then trains offline, achieving parallel acceleration of training and generation;
Atlas Gaussians Diffusion for 3D Generation
Haitao Yang (University of Texas at Austin), Qixing Huang (University of Texas at Austin)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderGaussian SplattingPoint CloudMesh
🎯 What it does: Proposes the Atlas Gaussians representation method and constructs a VAE + Latent Diffusion pipeline to achieve high-quality 3D generation under both unconditional and text-conditioned scenarios.
Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation
Yikun Zhang (Peking University), Yu Rong (Tencent AI Lab)
GenerationRetrievalDrug DiscoveryTransformerTextMultimodality
🎯 What it does: An end-to-end molecular-text cross-modal representation learning framework called Atomas is proposed, utilizing hierarchical adaptive alignment between molecular SMILES and text descriptions to achieve molecular retrieval, property prediction, and molecular generation tasks.
AtomSurf: Surface Representation for Learning on Protein Structures
Vincent Mallet (Mines Paris PSL Research University), Maks Ovsjanikov (Ecole Polytechnique)
Protein Structure PredictionGraph Neural NetworkDiffusion modelGraphBiomedical DataBenchmark
🎯 What it does: In the study of protein structure learning, the effectiveness of surface representation is researched and evaluated, improving DiffusionNet and constructing a surface graph hybrid network.
Attention as a Hypernetwork
Simon Schug (ETH Zurich), Razvan Pascanu (Google DeepMind)
TransformerText
🎯 What it does: Reformulating multi-head attention as a hypernetwork reveals that low-dimensional latent codes are used to configure key-query specific linear value networks.
Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers
Shijie Chen (Ohio State University), Yu Su (Ohio State University)
RetrievalExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A non-generative zero-shot re-ranking method called ICR based on the attention distribution of large language models (LLM) is proposed, which aggregates and calibrates the attention changes of query words on document words to directly obtain document relevance scores.
Attention layers provably solve single-location regression
Pierre Marion (Institute of Mathematics EPFL), Claire Boyer (Université Paris-Saclay)
TransformerSequential
🎯 What it does: Proposed a single position regression task and designed an attention-based predictor.
Attention with Markov: A Curious Case of Single-layer Transformers
Ashok Vardhan Makkuva (École Polytechnique Fédérale de Lausanne), Michael Gastpar (École Polytechnique Fédérale de Lausanne)
OptimizationTransformerSequential
🎯 What it does: This study investigates the learning capability of a single-layer Transformer under first-order Markov chain inputs, constructs a theoretical framework, and analyzes its loss landscape, proving the existence of global and local optima.
AttriBoT: A Bag of Tricks for Efficiently Approximating Leave-One-Out Context Attribution
Fengyuan Liu (University of Toronto), Colin Raffel (University of Toronto)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes the AttriBoT method, providing efficient context attribution (LOO error) calculation, suitable for large-scale LLMs;
Attribute-based Visual Reprogramming for Vision-Language Models
Chengyi Cai (University of Melbourne), Feng Liu (University of Melbourne)
ClassificationRecognitionTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: Attribute-driven visual reprogramming (AttrVR) is proposed on visual-language models like CLIP, achieving few-shot classification by adding trainable noise to the input image and optimizing guided by descriptive attributes (DesAttrs) and distinctive attributes (DistAttrs) generated by LLM.
Attributing Culture-Conditioned Generations to Pretraining Corpora
Huihan Li (University of Southern California), Xiang Ren (University of Southern California)
GenerationTransformerLarge Language ModelText
🎯 What it does: Analyzes the cultural bias in the generation of large language models under cultural conditions and proposes the MEMOED framework to attribute whether generated symbols come from the memory of pre-trained data.
Audio Large Language Models Can Be Descriptive Speech Quality Evaluators
Chen Chen (Nanyang Technological University), EngSiong Chng
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAudio
🎯 What it does: This study investigates the perception and description of speech quality by audio large language models (Audio LLM) and proposes corresponding evaluation datasets and training methods.
AugKD: Ingenious Augmentations Empower Knowledge Distillation for Image Super-Resolution
Yun Zhang (Hong Kong University of Science and Technology), Wenjia Wang (Hong Kong University of Science and Technology)
RestorationSuper ResolutionKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: Utilize unpaired scaling/flipping/color inversion to enhance the generation of auxiliary distillation samples, and improve the effectiveness of knowledge distillation in image super-resolution by combining reversible label consistency regularization.
AuroraCap: Efficient, Performant Video Detailed Captioning and a New Benchmark
Wenhao Chai (University of Washington), Christopher D Manning
GenerationComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: AURORACAP, an efficient fine-grained video description method based on large-scale multimodal models, is proposed, which achieves visual token compression by combining token merging technology.
Auto-GDA: Automatic Domain Adaptation for Efficient Grounding Verification in Retrieval-Augmented Generation
Tobias Leemann (University of Tübingen), Sergul Aydore
RetrievalDomain AdaptationTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: The Auto-GDA framework is proposed, which automatically adapts NLI models through unsupervised synthetic data generation and selection, thereby improving the grounding verification accuracy of retrieval-augmented generation (RAG) systems.
AutoBencher: Towards Declarative Benchmark Construction
Xiang Lisa Li (Stanford University), Tatsunori Hashimoto (Stanford University)
OptimizationSafty and PrivacyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the AutoBencher framework, which uses LLM to automatically generate and optimize evaluation datasets, achieving declarative benchmark construction.
AutoCGP: Closed-Loop Concept-Guided Policies from Unlabeled Demonstrations
Pei Zhou (University of Hong Kong), Yanchao Yang (University of Hong Kong)
Robotic IntelligenceTransformerDiffusion modelVideo
🎯 What it does: Develop a closed-loop concept-guided strategy that automatically learns manipulation concepts from unlabeled demonstrations and uses concept-guided diffusion strategies to complete complex robotic tasks.
Autocorrelation Matters: Understanding the Role of Initialization Schemes for State Space Models
Fusheng Liu (National University of Singapore), Qianxiao Li (National University of Singapore)
Time SeriesSequential
🎯 What it does: This paper studies the initialization scheme of state space models (SSM), particularly the effects of the time scale ∆, the real and imaginary parts of the state matrix W on training stability, long-term memory capability, and optimization conditions;
AutoDAN-Turbo: A Lifelong Agent for Strategy Self-Exploration to Jailbreak LLMs
Xiaogeng Liu (University of Wisconsin-Madison), Chaowei Xiao (University of Wisconsin-Madison)
Adversarial AttackTransformerLarge Language ModelAgentic AIText
🎯 What it does: AutoDAN-Turbo implements black-box jailbreak attacks through lifelong learning agents, capable of automatically discovering, combining, and utilizing various jailbreak strategies from scratch, and supports seamless injection of human-designed strategies.
AutoG: Towards automatic graph construction from tabular data
Zhikai Chen (Michigan State University), George Karypis (Amazon)
Recommendation SystemGraph Neural NetworkLarge Language ModelGraphTabularBenchmarkChain-of-Thought
🎯 What it does: This paper presents AutoG, a framework for automatically constructing graph structures from tabular data based on large language models, and constructs eight benchmark datasets covering multi-table real-world scenarios.
Automated Design of Agentic Systems
Shengran Hu (University of British Columbia), Jeff Clune (University of British Columbia)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelAgentic AITextMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a framework for an Automated Design Agent System (ADAS) and implements the Meta Agent Search algorithm, utilizing large language models to generate and optimize new agents in the code space, gradually building agent profiles.
Automated Filtering of Human Feedback Data for Aligning Text-to-Image Diffusion Models
Yongjin Yang (KAIST), Kimin Lee (KAIST)
GenerationOptimizationData-Centric LearningLarge Language ModelReinforcement LearningDiffusion modelText
🎯 What it does: For the fine-tuning of text-to-image diffusion models with human feedback, an automated filtering algorithm called FiFA is proposed, which retains only a high-quality, information-rich, and diverse subset of data, significantly improving training efficiency and model performance.
Automated Proof Generation for Rust Code via Self-Evolution
Tianyu Chen (Peking University), Lidong Zhou (Microsoft Research)
OptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the SAFE framework for automated proof generation on Rust code, overcoming the challenge of lacking manual proof data.
Automatic Curriculum Expert Iteration for Reliable LLM Reasoning
Zirui Zhao (National University of Singapore), Doyen Sahoo (Salesforce AI Research)
Large Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes an Automated Course Expert Iteration method (AUTO-CEI) aimed at enhancing the reliability of large language models (LLMs) in multi-step reasoning tasks, reducing hallucinations and lazy behaviors.