These 722 ICML 2025 papers come with a code repository. Each shows an AI one-line summary below β get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every ICML 2025 paper, free trial on arXivSub.
"Why Is There a Tumor?": Tell Me the Reason, Show Me the Evidence
Mengmeng Ma (University of Delaware), Xi Peng (University of Delaware)
CodeObject DetectionSegmentationExplainability and InterpretabilityLarge Language ModelVision Language ModelImageBiomedical DataMagnetic Resonance Imaging
π― What it does: This paper proposes an interpretable medical imaging model that can perform tumor detection and segmentation while generating clinical terminology reasoning and specifying location information.
Abdel-Rahim Mezidi (Universite Jean Monnet), Marc Sebban (Universite Jean Monnet)
CodeOptimizationTabularBenchmark
π― What it does: A framework is proposed that views neural operator layers as a minimization of Bregman regularization optimization problems, and based on this, a new Bregman neural operator is designed, unifying and extending existing neural operator models.
π― What it does: A chaotic dynamics framework inspired by the dorsal visual pathway is proposed, which encodes the event streams generated by event cameras using Continuous Coupled Neural Networks (CCNN) and extracts higher-order mappings through continuous wavelet transform, thereby generating stable and general event representations, which are combined with traditional CNNs for object classification.
Mirco Mutti (Technion Israel Institute of Technology), Aviv Tamar (Technion Israel Institute of Technology)
CodeExplainability and InterpretabilityMeta LearningTabular
π― What it does: This paper proposes a classification perspective to study the bandwidth problem in meta-learning, aiming to design interpretable and fast exploration plans for a fixed set of bandwidths.
π― What it does: A first-order gradient-based generative bi-level optimization framework is proposed to adjust the hyperparameters of diffusion models, addressing two types of bi-level problems: fine-tuning (reward alignment) and noise scheduling (noise timetable during training).
A Generic Family of Graphical Models: Diversity, Efficiency, and Heterogeneity
Yufei Huang (Peking University), Ruibin Xi (Peking University)
CodeTabularBiomedical Data
π― What it does: A 'marginally recoverable' family of distributions is proposed to construct graphical models compatible with various data types (continuous, count, binary, etc.), and to avoid high-dimensional integration through maximum marginal likelihood estimation (MMLE) in high dimensions, further extending to mixed models for heterogeneous structure inference.
A Geometric Approach to Personalized Recommendation with Set-Theoretic Constraints Using Box Embeddings
Shib Sankar Dasgupta (University of Massachusetts Amherst), Andrew McCallum (University of Massachusetts Amherst)
CodeRecommendation SystemTabular
π― What it does: This paper proposes a geometric method based on box embeddings to address personalized recommendation problems with set-theoretic constraints.
π― What it does: A Large Recursive Action Model (LRAM) is proposed, centered around xLSTM, to achieve rapid inference in robotic reinforcement learning tasks.
A Machine Learning Approach to Duality in Statistical Physics
Prateek Gupta (Max Planck Institute for Human Development), Nabil Iqbal (Amsterdam Machine Learning Laboratory)
CodeOptimizationRepresentation LearningData-Centric LearningTabularPhysics Related
π― What it does: Using neural networks and loss functions to automatically discover dual relationships in statistical physics models, we reproduced the Kramers-Wannier duality of the two-dimensional Ising model and explored equivalent forms with four-body interactions.
A Market for Accuracy: Classification Under Competition
Ohad Einav (Technion Israel Institute of Technology), Nir Rosenfeld (Technion Israel Institute of Technology)
CodeClassificationOptimizationTabular
π― What it does: This paper studies classification tasks in a competitive environment, proposing to transform the maximization of market share into a weighted accuracy objective, and analyzes the best response dynamics and their convergence;
A Reductions Approach to Risk-Sensitive Reinforcement Learning with Optimized Certainty Equivalents
Kaiwen Wang (Cornell Tech), Wen Sun (Cornell Tech)
CodeReinforcement LearningTabular
π― What it does: Two types of meta-algorithms (optimistic and gradient) are proposed to transform static OCE risk reinforcement learning (RL) into conventional risk-neutral RL, solved in augmented MDPs, thus obtaining theoretical guarantees for global convergence, PAC, and local improvement, and experimentally validating the necessity of non-Markovian policies.
CodeClassificationMeta LearningMixture of ExpertsImage
π― What it does: This study focuses on long-tail semi-supervised learning and proposes the Meta-Expert algorithm, which achieves high-quality pseudo-labels and predictions through dynamic expert allocation and multi-depth feature fusion.
A Sub-Problem Quantum Alternating Operator Ansatz for Correlation Clustering
Lucas Fabian Naumann (TU Dresden), Bjoern Andres (TU Dresden)
CodeOptimizationGraph Neural NetworkGraph
π― What it does: Proposed and experimentally validated a variant of the Subproblem Quantum Alternating Operator (SQAOA) for the related clustering problem, improving solution quality and resource utilization efficiency through kernel sampling and subproblem decomposition.
A Unified Framework for Entropy Search and Expected Improvement in Bayesian Optimization
Nuojin Cheng (University of Colorado Boulder), Luigi Nardi (Lund University)
CodeOptimizationTabular
π― What it does: Proposes the Variational Entropy Search (VES) framework, unifying Expected Improvement (EI) with information-theoretic sampling functions (such as Max-Value Entropy Search, MES), and designs the VES-Gamma sampling function based on this framework;
π― What it does: A unified framework for Discrete Weak Features Learning (WFL) is proposed, along with its theoretical form for risk minimization; under this framework, the LAC-dWFL learning algorithm class is defined, and a generalization error analysis is conducted, revealing the interaction between the feature estimation model g and the label prediction model f.
A Variational Framework for Improving Naturalness in Generative Spoken Language Models
Li-Wei Chen (Carnegie Mellon University), Alexander Rudnicky (Carnegie Mellon University)
CodeGenerationDiffusion modelAuto EncoderAudio
π― What it does: This paper proposes an end-to-end framework that combines variational autoencoders and speech language models, using learned continuous features to supplement semantic discrete words, thereby enhancing the naturalness and coherence of speech generation.
A Variational Perspective on Generative Protein Fitness Optimization
Lea Bogensperger (University of Zurich), Michael Krauthammer (University of Zurich)
CodeOptimizationDrug DiscoveryFlow-based ModelAuto EncoderBiomedical Data
π― What it does: This paper proposes a variational inference-based protein adaptability optimization framework called VLGPO, which can perform posterior sampling of protein sequences in a continuous latent space to generate highly adaptive sequences.
Ab Initio Nonparametric Variable Selection for Scalable Symbolic Regression with Large $p$
Shengbin Ye (Rice University), Meng Li (Northwestern University)
CodeTabularPhysics Related
π― What it does: A PAN+SR framework is proposed, combining BART-based non-parametric variable selection with symbolic regression to achieve scalable symbolic regression on large feature sets.
π― What it does: This study investigates the use of unstructured sparsification and fixed-point quantization to accelerate linear recurrent neural networks (S5) on edge devices, and implements real-time audio denoising inference on the Intel Loihi 2 neuromorphic chip.
Accelerating PDE-Constrained Optimization by the Derivative of Neural Operators
Ze Cheng (Bosch), Hang Su (Tsinghua University)
CodeOptimizationTabularTime SeriesPhysics Related
π― What it does: This paper proposes a PDE-constrained optimization acceleration framework that combines neural operators with gradient optimization. The core components include a reference neural operator (RNO) that utilizes optimization trajectory data for optimization-guided training, the introduction of a Virtual-Fourier layer to enhance derivative learning, and an iterative strategy that combines mixed solvers with neural operators.
Active Evaluation Acquisition for Efficient LLM Benchmarking
Yang Li (Amazon Web Services), Graham Horwood
CodeComputational EfficiencyLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
π― What it does: This paper proposes an Active Evaluation Acquisition (AEA) strategy and a neural process model to significantly reduce the number of prompts required during benchmark evaluations of large language models, thereby improving evaluation efficiency.
Active Learning for Efficient Discovery of Optimal Combinatorial Perturbations
Jason Qin (Neptune Bio), Yuhan Hao (Neptune Bio)
CodeOptimizationDrug DiscoveryBiomedical Data
π― What it does: The NAIAD framework is proposed, which uses active learning to predict and discover the most effective gene or drug combinations based on single-gene effects, significantly reducing experimental cycles.
Active Learning of Deep Neural Networks via Gradient-Free Cutting Planes
Erica Zhang (Stanford University), Mert Pilanci (Stanford University)
CodeClassificationOptimizationText
π― What it does: This paper proposes a deep neural network training and active learning framework based on gradient-free cutting-plane methods, which can achieve geometric shrinkage in the parameter space of non-convex ReLU networks and converge to the optimal solution.
π― What it does: A framework for active sampling in PDE surrogate model learning is proposed, which can selectively choose only the most important time steps for high-cost numerical solutions, while approximating the remaining time steps with a surrogate model, thereby significantly reducing data collection costs.
π― What it does: The Ad-Hoc Human-AI Coordination Challenge (AH2AC2) is proposed, which evaluates the real-time collaboration ability of AI and humans by using human proxy agents in the cooperative card game Hanabi.
π― What it does: This study investigates the sample complexity of multi-reward and multi-policy evaluation in online discount MDPs, providing a theoretical lower bound for (Ξ΅, Ξ΄)-PAC and corresponding exploration strategies.
π― What it does: This paper proposes an Adaptive Median Smoothing defense for a text-image diffusion model with removed concepts during inference, dynamically injecting anisotropic noise based on the correlation of tokens to enhance robustness against adversarial attacks.
Adaptive Message Passing: A General Framework to Mitigate Oversmoothing, Oversquashing, and Underreaching
Federico Errica (NEC Laboratories Europe), Francesco Alesiani (NEC Laboratories Europe)
CodeGraph Neural NetworkGraph
π― What it does: An Adaptive Message Passing (AMP) framework has been developed, enabling graph neural networks to autonomously determine the depth and filtering strategy of message passing during training, thereby alleviating long-range dependency issues such as over-smoothing, over-compression, and the inability to reach remote nodes.
Adaptive Sensitivity Analysis for Robust Augmentation against Natural Corruptions in Image Segmentation
Laura Yu Zheng, Ming Lin
CodeSegmentationTransformerImage
π― What it does: Proposes an online data augmentation method based on adaptive sensitivity analysis to enhance the robustness of semantic segmentation models against natural noise.
AdaptiveStep: Automatically Dividing Reasoning Step through Model Confidence
Yuliang Liu (Nanjing University), Zhouhan Lin (Shanghai Jiaotong University)
CodeAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: Proposes the AdaptiveStep method, which uses the model's confidence in the next word to automatically divide inference steps, and trains a Process Reward Model (ASPRM) based on this for mathematical reasoning and code generation.
CodeAuto EncoderGenerative Adversarial NetworkTime Series
π― What it does: The AdaPTS framework is proposed, which utilizes feature space adapters to transfer pre-trained univariate time series foundational models (FM) to multivariate probabilistic prediction tasks.
Nuno Gonçalves (Instituto Superior Tecnico), Andre Martins
CodeRetrievalOptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: This paper presents ADASPLASH, a hardware-friendly implementation of Ξ±-entmax that combines Hybrid Halley-bisection iteration and a custom Triton kernel to achieve efficient forward/backward computation of sparse attention.
Leif DΓΆring (University of Mannheim), Martin Slowik (University of Mannheim)
CodeReinforcement LearningTabular
π― What it does: An Adaptive Distributed Double Q-Learning (ADDQ) algorithm is proposed, which dynamically decides whether to use standard Q-learning or double Q-learning updates based on the sample variance estimated by distributed RL, in order to alleviate the overestimation problem.
Addressing Concept Mislabeling in Concept Bottleneck Models Through Preference Optimization
Emiliano Penaloza (Mila Quebec AI Institute), Mateo Espinosa Zarlenga (University of Cambridge)
CodeOptimizationReinforcement LearningImage
π― What it does: A Concept Preference Optimization (CPO) loss based on Direct Preference Optimization is proposed, which directly optimizes the concept posterior of the Concept Bottleneck Model (CBM) to alleviate the impact of concept label noise on model performance.
Addressing Imbalanced Domain-Incremental Learning through Dual-Balance Collaborative Experts
Lan Li (Nanjing University), De-Chuan Zhan (Nanjing University)
CodeDomain AdaptationTransformerMixture of ExpertsImage
π― What it does: In Domain Incremental Learning (DIL), this paper proposes the Dual-Balance Collaborative Experts (DCE) framework to address the issues of internal class imbalance and cross-domain class distribution drift caused by sample imbalance.
π― What it does: This paper studies the subtle differences in floating-point operations among different linear algebra backends (such as Intel MKL, Nvidia cuBLAS, Apple Accelerate, etc.) and utilizes these differences to construct 'Chimera' adversarial inputs, causing the same model to produce contradictory predictions across different backends.
π― What it does: This paper proposes a white-box ordered Top-K adversarial attack method called RisingAttacK based on Sequential Quadratic Programming (SQP), which directly learns perturbations in the image space that satisfy a specified ranking order.
π― What it does: A framework for adversarial image attacks on image-to-image diffusion models (AdvI2I) has been designed and implemented, which generates adversarial images to induce the model to produce NSFW content and can bypass existing safety checks.
Aequa: Fair Model Rewards in Collaborative Learning via Slimmable Networks
Nurbek Tastan (Mohamed bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Michigan State University)
CodeFederated LearningSafty and PrivacyImage
π― What it does: The AEQUA framework is proposed, which achieves model rewards based on participant contributions through slimmable neural networks in collaborative learning, addressing the issue of fair reward distribution.
AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models
Zheng Lian (Institute of Automation Chinese Academy of Sciences), Jianhua Tao (Tsinghua University)
CodeClassificationGenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodalityBenchmark
π― What it does: A large-scale emotional description dataset MER-Caption has been established, a pre-fusion multimodal large language model AffectGPT has been proposed outside of LLMs, and a unified evaluation benchmark MER-UniBench has been created.
CodeRecommendation SystemTransformerLarge Language ModelVision Language ModelVideoMultimodalityAudio
π― What it does: A dataset called AGAVQA-3k was constructed, and the AGAV-Rater model was proposed to evaluate the audio quality, audio-video consistency, and overall quality of AI-generated audio and video, as well as to select the best output.
Zora Zhiruo Wang (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
CodeLarge Language ModelAgentic AITextBenchmark
π― What it does: Proposes Agent Workflow Memory (AWM), which guides language model-driven web navigation agents by extracting reusable workflows from completed task trajectories.
π― What it does: A method considering the dependency between expert distributions, CoDE, is proposed, and based on this method, a CoDE-VAE multimodal variational autoencoder is constructed.
π― What it does: Proposes AKRMap, an adaptive kernel regression supervised dimensionality reduction method for trustworthy visualization of cross-modal embedding metrics.
π― What it does: A resource augmentation algorithm for the active learning label selection (GLS) problem on general weighted graphs is proposed, which can achieve optimal Psi(L) with O(log n) resources; it is also proven that the problem is NP-hard for threshold determination of 2 and 3 on unweighted graphs.
Aligning with Logic: Measuring, Evaluating and Improving Logical Preference Consistency in Large Language Models
Yinhong Liu (University of Cambridge), Nigel Collier (University of Cambridge)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A general framework is proposed to evaluate the logical preference consistency of LLMs (transitivity, symmetry, negation invariance), and the REPAIR method is developed to estimate and enhance consistency through ranking.
π― What it does: A unified All-atom Diffusion Transformer (ADiT) is proposed, which simultaneously generates molecules and materials through a variational autoencoder and latent diffusion model.
All-atom inverse protein folding through discrete flow matching
Kai Yi (MRC Laboratory of Molecular Biology), Sjors HW Scheres
CodeProtein Structure PredictionGraph Neural NetworkFlow-based ModelBiomedical Data
π― What it does: A discrete flow matching-based all-atom inverse folding model, ADFLIP, is proposed, which can generate structurally consistent amino acid sequences for protein complexes containing small molecules, nucleic acids, or metal ions.
CodeGenerationProtein Structure PredictionTransformerFlow-based ModelBiomedical Data
π― What it does: A generative model APM based on atomic-level design has been developed to generate multi-chain protein complexes from scratch, capable of performing folding, unfolding, and specific functional protein (antibody, peptide) design tasks.
An Effective and Secure Federated Multi-View Clustering Method with Information-Theoretic Perspective
Xinyue Chen (University of Electronic Science and Technology of China), Yazhou Ren (University of Electronic Science and Technology of China)
CodeFederated LearningSafty and PrivacyAuto EncoderMultimodality
π― What it does: This paper proposes a method for splitting features in federated multi-view clustering from an information-theoretic perspective, sharing only the features relevant to clustering to enhance performance and reduce the risk of privacy leakage.
An End-to-End Model for Logits-Based Large Language Models Watermarking
KA HIM WONG, Yain-Whar Si (University of Macau)
CodeGenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: An end-to-end logits perturbation watermarking model is proposed, capable of embedding detectable watermarks in LLM-generated text while maintaining text quality.
any4: Learned 4-bit Numeric Representation for LLMs
Mostafa Elhoushi (Meta), Jeff Johnson (Meta)
CodeTransformerLarge Language ModelText
π― What it does: This paper proposes an arbitrary 4-bit (any4) numerical representation method for unprocessed low-bit quantization of large language model (LLM) weights, and provides a corresponding GPU-efficient matrix multiplication library called tinygemm.
Weiwei Li (Nanjing University of Aeronautics and Astronautics), Xiuyi Jia (Nanjing University of Science and Technology)
CodeOptimizationSupervised Fine-TuningTabular
π― What it does: This study proposes DeltaLDLβa performance evaluation metric and learning objective based on 'approximately correct' label distributions to improve the performance assessment and training of label distribution learning.
Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models
Thomas Fel (Harvard University), Talia Konkle (Harvard University)
CodeExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderImage
π― What it does: This paper addresses the instability exhibited by existing Sparse Autoencoders (SAE) in concept extraction, proposing and implementing two new methodsβArchetypal SAE (A-SAE) and its relaxed version RA-SAEβto stabilize and enhance the interpretability of concepts by constraining dictionary atoms within the convex hull of the data.
Arrow: Accelerator for Time Series Causal Discovery with Time Weaving
Yuanyuan Yao (Zhejiang University), TIANYI LI
CodeOptimizationComputational EfficiencyTime Series
π― What it does: Developed the ARROW (Accelerator for Time Series Causal Discovery with Time Weaving) accelerator, aimed at significantly improving the operational efficiency and accuracy of existing time series causal discovery algorithms.
Manu Bhat (University of California San Diego), Rose Yu (University of California San Diego)
CodeConvolutional Neural NetworkGraph Neural NetworkTabularTime SeriesPhysics Related
π― What it does: This paper proposes the AtlasD framework, which can automatically discover local symmetries (i.e., atlas equivariance) from data and express them as Lie group bases and discrete cosets;
Attributes Shape the Embedding Space of Face Recognition Models
Pierrick Leroy (Politecnico di Torino), Francesco Vaccarino (Politecnico di Torino)
CodeRecognitionExplainability and InterpretabilityGenerative Adversarial NetworkImage
π― What it does: This paper studies the geometric structure of the embedding space of facial recognition models, exploring how interpretable attributes (such as hair color, age, pose, etc.) shape this space on both macro scales (relationships between identity clouds) and micro scales (internal structure of a single identity), and proposes corresponding quantification methods.
AutoAdvExBench: Benchmarking Autonomous Exploitation of Adversarial Example Defenses
Nicholas Carlini (Google DeepMind), Florian Tramèr (ETH Zurich)
CodeAdversarial AttackTransformerLarge Language ModelImageBenchmark
π― What it does: A new benchmark called AutoAdvExBench has been constructed, providing 75 real or CTF-style implementations of adversarial sample defenses, allowing large language models (LLMs) to automatically generate attack samples and evaluate their success rates given defense papers and code.
AutoAL: Automated Active Learning with Differentiable Query Strategy Search
Yifeng Wang (Carnegie Mellon University), Siyu Huang (Clemson University)
CodeClassificationOptimizationConvolutional Neural NetworkReinforcement LearningImageBiomedical Data
π― What it does: Proposes AutoAL, a method for automatically searching for optimal active learning query strategies through a bi-level differentiable optimization framework;
π― What it does: Developed and validated AutoCATE, a fully automated, end-to-end causal effect estimation framework that can automatically search for and construct the optimal CATE estimation pipeline in observational data.
π― What it does: This study investigates the impact of Concept-Based Models (CBMs) on concept intervention under distribution shift (OOD) and proposes a new MixCEM model to avoid leakage poisoning.
Balanced Learning for Domain Adaptive Semantic Segmentation
Wangkai Li (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
CodeSegmentationDomain AdaptationImage
π― What it does: A balanced learning method BLDA is proposed for unsupervised domain adaptation semantic segmentation, addressing the prediction bias of different categories in the target domain;
π― What it does: BARNN is proposed, a framework that converts any autoregressive or recurrent model into a Bayesian version, applicable to tasks such as time series, PDE solving, and molecular generation.
Bayesian Basis Function Approximation for Scalable Gaussian Process Priors in Deep Generative Models
Mehmet YiΔit BalΔ±k (Aalto University), Harri LΓ€hdesmΓ€ki (Aalto University)
CodeGenerationData SynthesisComputational EfficiencyTime SeriesSequentialBiomedical Data
π― What it does: A scalable Gaussian Process prior VAE model DGBFGP based on basis function approximation is proposed, which handles high-dimensional time series data with linear time complexity.
Bayesian Inference for Correlated Human Experts and Classifiers
Markelle Kelly (University of California), Padhraic Smyth (University of California)
CodeClassificationImage
π― What it does: Under the premise of given pre-trained classifier outputs, a Bayesian framework is proposed to accurately predict expert voting (consensus) with minimal human expert queries and to perform inference in online sequential queries.
Bayesian Optimization from Human Feedback: Near-Optimal Regret Bounds
Aya Kayal (University College London), Alberto Bernacchia (MediaTek Research)
CodeOptimizationReinforcement Learning from Human FeedbackTabular
π― What it does: A multi-round learning strategy MR-LPF is proposed to find the globally optimal action from Bayesian optimization with only comparative feedback, and a tighter sub-linear regret upper bound is provided.
BECAME: Bayesian Continual Learning with Adaptive Model Merging
Mei Li (Shanghai Jiao Tong University), Hongtao Lu (Shanghai Jiao Tong University)
CodeClassificationOptimizationImage
π― What it does: By linearly mixing the gradient projection model with the unconstrained model in continual learning, a two-stage training framework called BECAME is proposed to achieve a balance between stability and plasticity.
Zihan Guan (University of Virginia), Anil Vullikanti (University of Virginia)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a sample selection method based on Self-Influence Score Normalization (Self-Inf-N) to identify the most destructive outlier samples for LLM safety alignment from seemingly benign datasets, and fine-tunes the model using only these 100 samples, significantly increasing the probability of generating harmful content.
BEST-Route: Adaptive LLM Routing with Test-Time Optimal Compute
Dujian Ding (University of British Columbia), Victor RΓΌhle
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The BEST-Route framework is proposed, which dynamically selects LLMs of different scales and adjusts the best-of-n sampling number based on query difficulty to reduce inference costs while maintaining high quality.
Beyond Bradley-Terry Models: A General Preference Model for Language Model Alignment
Yifan Zhang (Tsinghua University), Quanquan Gu (University of California)
CodeRecommendation SystemOptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: A preference embedding model (GPM) that embeds answers into a multidimensional latent space is proposed, enabling linear queries of complex non-transitive preferences.
Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation
Zixuan Hu (Peking University), LINGYU DUAN
CodeDomain AdaptationGaussian SplattingImage
π― What it does: A method called ReCAP is proposed for Wild Testing Time Adaptation (WTTA), which replaces traditional entropy minimization with region confidence, thereby more effectively updating the model in data-scarce and multi-shift environments.
π― What it does: This paper proposes an activation compression method based on single-shot subspace iteration (ASI), which significantly reduces the memory of activations during backpropagation and improves training efficiency.
Beyond The Rainbow: High Performance Deep Reinforcement Learning on a Desktop PC
Tyler Clark (University Of Southampton), Jonathon Hare (University Of Southampton)
CodeReinforcement LearningVideo
π― What it does: BTR constructs an efficient and user-friendly reinforcement learning algorithm by combining six improvements of Rainbow DQN with a new architecture, regularization, distributed learning, adjustable exploration, parallelization, and adaptive pooling.
Ruth Wan Theng Chew (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
CodeOptimization
π― What it does: A novel two-layer Bayesian optimization algorithm BILBO is proposed, which can simultaneously optimize the upper and lower layers in noisy, constrained, and gradient-free black-box two-layer problems.
π― What it does: This paper proposes BiMaCoSR, a first-order binary diffusion model that achieves extreme compression and acceleration for real super-resolution methods through low-rank and sparse matrix branches.
Black-Box Adversarial Attacks on LLM-Based Code Completion
Slobodan Jenko (ETH Zurich), Martin Vechev (ETH Zurich)
CodeAdversarial AttackAI Code AssistantTransformerLarge Language ModelText
π― What it does: A black-box adversarial attack targeting code completion engines based on large language models (LLMs) is proposed, where the attacker injects short comment strings into queries to induce the engine to generate code with security vulnerabilities.
Boosting Virtual Agent Learning and Reasoning: A Step-Wise, Multi-Dimensional, and Generalist Reward Model with Benchmark
Bingchen Miao (Ant Group), Juncheng Li (Zhejiang University)
CodeLarge Language ModelReinforcement LearningAgentic AIMultimodalityBenchmark
π― What it does: A step-by-step, multi-dimensional reward model (Similar) and corresponding SRM benchmark are proposed and implemented. The automated MCTS-P algorithm is used to collect cross-task process-level annotated data across four major platforms, which is then utilized to train the reward model, providing fine-grained feedback during the training and inference phases of the virtual agent (GVA).
π― What it does: A preference optimization training paradigm based on best anchoring and goal guidance (BOPO) is proposed to enhance the sample efficiency of neural combinatorial optimization models when solving NP-hard problems.
BoxLM: Unifying Structures and Semantics of Medical Concepts for Diagnosis Prediction in Healthcare
Yanchao Tan (Fuzhou University), Carl Yang (Emory University)
CodeClassificationRecommendation SystemGraph Neural NetworkTransformerLarge Language ModelTabularBiomedical DataElectronic Health Records
π― What it does: Proposes the BoxLM framework to implement diagnostic prediction in electronic health records (EHR), unifying the structure and semantics of medical concepts.
Breaking the Barrier of Hard Samples: A Data-Centric Approach to Synthetic Data for Medical Tasks
MAYNARA DONATO DE SOUZA, Cleber Zanchettin (Universidade Federal de Pernambuco)
CodeData SynthesisData-Centric LearningDiffusion modelFlow-based ModelGenerative Adversarial NetworkTabularBiomedical Data
π― What it does: The Profile2Gen framework is proposed to enhance the quality of synthetic data for medical regression tasks through data analysis and iterative generation.
Bridging Layout and RTL: Knowledge Distillation based Timing Prediction
Mingjun Wang (Institute of Computing Technology Chinese Academy of Sciences), Huawei Li (Institute of Computing Technology Chinese Academy of Sciences)
π― What it does: The RTLDistil framework is proposed, which transfers layout-level physical features to RTL-level models through cross-stage knowledge distillation, achieving high-accuracy timing prediction.
Bring Reason to Vision: Understanding Perception and Reasoning through Model Merging
Shiqi Chen (City University of Hong Kong), Junxian He (Hong Kong University of Science and Technology)
CodeRecognitionGenerationOptimizationTransformerLarge Language ModelVision Language ModelTextMultimodality
π― What it does: This paper achieves an enhancement in the multimodal reasoning ability of VLM by performing a weighted average combination of the language module of LLM focused on mathematical reasoning and VLM, without the need for additional training.
Yu Liang (University of Electronic Science and Technology of China), Yang Yang (University of Electronic Science and Technology of China)
CodeOptimizationSpiking Neural NetworkImageAudio
π― What it does: This paper proposes an online training algorithm for Binary Spiking Neural Networks (BSNN) - BSO and its time-aware variant T-BSO, significantly reducing training memory requirements.
π― What it does: C-3PO is proposed, a proxy center framework based on a lightweight multi-agent system, achieving unmodified alignment between the retriever and large language models (LLM), and simulating the interactive process of human search behavior.
Cache Me If You Must: Adaptive Key-Value Quantization for Large Language Models
Alina Shutova (HSE University), Dan Alistarh (ISTA)
CodeCompressionTransformerLarge Language ModelText
π― What it does: An adaptive quantization method AQUA-KV is proposed for the Key-Value cache of large language models, which predicts and quantizes the residuals by utilizing the linear correlation between keys and values in adjacent and same layers, significantly compressing the KV cache while maintaining inference accuracy.
π― What it does: A new multi-task learning model fusion method called CALM is proposed, which can merge multiple fine-tuned models into a unified model without retraining, while maintaining the performance of each task.
Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence
Yuankai Luo (Beihang University), Xiao-Ming Wu (Hong Kong Polytechnic University)
CodeGraph Neural NetworkGraphBenchmark
π― What it does: This study investigates the performance of classic GNNs in graph-level tasks and proposes a unified GNN+ framework to enhance their capabilities.
Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression
Peijie Dong (Hong Kong University of Science and Technology), Bo Li (Hong Kong University of Science and Technology)
CodeCompressionOptimizationTransformerLarge Language ModelAgentic AITextBenchmark
π― What it does: A benchmark for evaluating the capabilities of compressed LLMs in agent tasks, called ACBench, is proposed, and a systematic assessment of the impact of quantization and pruning on workflow generation, tool usage, long context understanding, and real applications is conducted.
Can Large Language Models Understand Intermediate Representations in Compilers?
Hailong Jiang (Kent State University), Qiang Guan (Kent State University)
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: Systematically evaluate the understanding ability of six major LLMs on compiler intermediate representation (IR) through four types of tasks (CFG reconstruction, IR decompilation, code summarization, execution reasoning).
π― What it does: A new parameter-efficient fine-tuning method called Canonical Rank Adaptation (CaRA) is proposed for fine-tuning a small number of parameters on Vision Transformers.
π― What it does: This paper studies the robustness of using protective perturbations in latent diffusion models to prevent unauthorized model customization and proposes a Contrastive Adversarial Training (CAT) method based on lightweight adapters, which reduces the effect of protective perturbations by reshaping latent representations.
Causal Abstraction Learning based on the Semantic Embedding Principle
Gabriele D'Acunto (Sapienza University), Paolo Di Lorenzo (Sapienza University)
CodeBiomedical DataMagnetic Resonance Imaging
π― What it does: A causal abstraction learning framework based on the principle of semantic embedding is proposed to address the problem of linear causal abstraction learning in the absence of observable aligned data.
Causal Effect Identification in lvLiNGAM from Higher-Order Cumulants
Daniele Tramontano (Technical University of Munich), Mathias Drton (Munich Center for Machine Learning)
CodeTabular
π― What it does: This paper utilizes higher-order cumulants in the latent variable linear non-Gaussian model (lvLiNGAM) to identify and estimate causal effects, particularly addressing the challenges of single proxy variables and underdetermined instrumental variables.
Causality-Aware Contrastive Learning for Robust Multivariate Time-Series Anomaly Detection
HyunGi Kim (Seoul National University), Sungroh Yoon (Seoul National University)
CodeAnomaly DetectionRecurrent Neural NetworkTransformerContrastive LearningTime Series
π― What it does: A multivariate time series anomaly detection framework called CAROTS is proposed, which utilizes causal-aware contrastive learning to train a contrastive encoder with causal preservation and perturbation-enhanced samples, combined with prediction error for anomaly scoring.