ICLR 2026 Papers — Page 2
International Conference on Learning Representations · 5356 papers
A^2TG: Adaptive Anisotropic Textured Gaussians for Efficient 3D Scene Representation
Sheng-Chi Hsu (National Tsing Hua University), Hung-Kuo Chu (National Tsing Hua University)
Computational EfficiencyRepresentation LearningGaussian SplattingPoint CloudBenchmark
🎯 What it does: Propose Adaptive Anisotropic Textured Gaussian Splatting (A2TG), assigning each 2D Gaussian a variable size and orientation texture, improving texture allocation and significantly reducing memory usage.
A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning
Qianben Chen (OPPO AI Agent Team), Wangchunshu Zhou (OPPO AI Agent Team)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextChain-of-Thought
🎯 What it does: A unified adaptive agent foundation model A2FM was constructed, integrating three execution modes: instant answering, chain-of-thought reasoning, and tool interaction, and achieving efficient reasoning through routing-alignment training and adaptive policy optimization.
A$^2$Search: Ambiguity-Aware Question Answering with Reinforcement Learning
Fengji Zhang (City University of Hong Kong), Junyang Lin (Alibaba Group)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Developed an annotation-free reinforcement learning framework called A2SEARCH, which utilizes an automated pipeline to detect multi-answer questions and generate alternative answers, enabling open-domain QA models to provide all valid answers in a single inference.
A2ASecBench: A Protocol-Aware Security Benchmark for Agent-to-Agent Multi-Agent Systems
Tianhao Li (Duke University), Chaowei Xiao (Johns Hopkins University)
Safty and PrivacyLarge Language ModelAgentic AITextBenchmarkFinance Related
🎯 What it does: Proposed and implemented a security benchmark framework named A2ASECBENCH for Agent-to-Agent (A2A) multi-agent systems, and conducted a system-level red team evaluation
A2D: Any-Order, Any-Step Safety Alignment for Diffusion Language Models
Wonje Jeung (Yonsei University), Albert No
Safty and PrivacyLarge Language ModelDiffusion modelTextBenchmark
🎯 What it does: A token-level alignment-based security defense method, A2D, was designed and implemented in Diffusion Large Language Models (dLLMs), which can forcibly make the model output the [EOS] rejection signal at any generation order and any step, thereby preventing the generation of harmful content.
ABBA-Adapters: Efficient and Expressive Fine-Tuning of Foundation Models
Raghav Singhal (Mohamed bin Zayed University of Artificial Intelligence), Praneeth Vepakomma (Mohamed bin Zayed University of Artificial Intelligence)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a new parameter-efficient fine-tuning (PEFT) architecture ABBA, which decomposes weight updates into two independent low-rank matrices via the Hadamard product, significantly enhancing expressiveness while maintaining low parameter counts.
AbdCTBench: Learning Clinical Biomarker Representations from Abdominal Surface Geometry
Muhammad Ahmed Chaudhry (Stanford University), Sanmi Koyejo (Stanford University)
ClassificationRepresentation LearningData-Centric LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningMeshComputed TomographyBenchmark
🎯 What it does: Constructed the AbdCTBench dataset, associating abdominal surface meshes extracted from CT scans with internal CT-derived biomarkers, diagnostic codes, and comorbidity labels, and conducted single-target biomarker prediction benchmark experiments on various computer vision architectures using this dataset;
AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features
Xudong Zhu (Ohio State University), Zhihui Zhu (Ohio State University)
Explainability and InterpretabilityRepresentation LearningLarge Language ModelAuto EncoderText
🎯 What it does: This paper proposes AbsTopK SAE by expanding the proximal gradient on sparse autoencoders (SAE), which can retain both positive and negative activations while maintaining sparsity to achieve bidirectional semantic representation for single features.
Abstracting Robot Manipulation Skills via Mixture-of-Experts Diffusion Policies
Ce Hao (National University of Singapore), Harold Soh (National University of Singapore)
Robotic IntelligenceMixture of ExpertsDiffusion model
🎯 What it does: Propose Skill Mixture-of-Experts Policy (SMP), generating actions in multi-task robotic manipulation through state-adaptive orthogonal skill bases and sticky gating.
AbstRaL: Augmenting LLMs' Reasoning by Reinforcing Abstract Thinking
Silin Gao (Apple), Emmanuel Abbe (Apple)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Learn abstract reasoning through reinforcement learning, constructing the AbstRaL framework to enhance the robustness of LLMs in GSM reasoning.
AC-Foley: Reference-Audio-Guided Video-to-Audio Synthesis with Acoustic Transfer
Pengjun Fang (Hong Kong University of Science and Technology), Harry Yang (Hong Kong University of Science and Technology)
Data SynthesisTransformerFlow-based ModelAuto EncoderVideoMultimodalityAudio
🎯 What it does: AC-Foley proposes a video-to-audio synthesis framework conditioned on a reference audio, achieving precise control over timbre and details while maintaining temporal synchronization.
AC-Sampler: Accelerate and Correct Diffusion Sampling with Metropolis-Hastings Algorithm
Minsang Park (KAIST), Il-chul Moon
GenerationDiffusion modelScore-based ModelImageTextMultimodalityStochastic Differential Equation
🎯 What it does: Proposed AC-Sampler, which accelerates and corrects the sampling process of diffusion models by directly sampling at intermediate time steps and using Metropolis-Hastings correction.
ACADREASON: Exploring the Limits of Reasoning Models with Academic Research Problems
Xin Gui (Beijing University Of Posts And Telecommunications), Wangchunshu Zhou (OPPO)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the ACADREASON benchmark, aiming to evaluate the capabilities of LLMs and Agents in high-level academic reasoning tasks.
Accelerated co-design of robots through morphological pretraining
Luke Strgar (Northwestern University), Sam Kriegman (Northwestern University)
OptimizationRobotic Intelligence
🎯 What it does: Pre-train a morphology-agnostic general controller using differentiable simulation on tens of millions of soft spring robots with diverse morphologies; subsequently achieve zero-shot evolution and few-shot evolution with this controller to rapidly search for diverse and high-performance robot morphologies.
Accelerated Learning with Linear Temporal Logic using Differentiable Simulation
Alper Kamil Bozkurt (Virginia Commonwealth University), Ming Lin (University of Maryland)
Reinforcement LearningWorld ModelBenchmark
🎯 What it does: This paper proposes an end-to-end reinforcement learning framework that combines linear temporal logic (LTL) specifications with differentiable simulation. By converting discrete automata into differentiable rewards through soft-labeled discretization, it enables direct training of controllers from formal specifications, significantly accelerating learning and improving performance.
Accelerated Parallel Tempering via Neural Transports
Leo Zhang (University of Oxford), Saifuddin Syed (University of British Columbia)
OptimizationComputational EfficiencyDiffusion modelFlow-based ModelBiomedical Data
🎯 What it does: This paper proposes the Accelerated Parallel Tempering (APT) framework, which integrates neural samplers (such as normalizing flows, controlled Monte Carlo diffusion, and diffusion models) into traditional parallel tempering to enhance the overlap between adjacent distributions, thereby significantly improving sample quality and free energy estimation;
Accelerating Benchmarking of Functional Connectivity Modeling via Structure-aware Core-set Selection
Ling Zhan (Southwest University), Tao Jia (Southwest University)
Computational EfficiencyRepresentation LearningTransformerContrastive LearningBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: Propose a functional connectivity (FC) modeling method based on core set selection, which retains the relative performance ranking of statistical pairwise interaction (SPI) methods on large-scale fMRI datasets by selecting a small number of samples.
Accelerating Diffusion Large Language Models with SlowFast Sampling: The Three Golden Principles
Qingyan Wei (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
Computational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Propose the SlowFast Sampling dynamic sampling strategy, integrating the three golden principles to achieve parallel decoding acceleration for diffusion-based large language models (dLLMs).
Accelerating Diffusion Planners in Offline RL via Reward-Aware Consistency Trajectory Distillation
Xintong Duan (Carnegie Mellon University), Jeff Schneider (Carnegie Mellon University)
Computational EfficiencyKnowledge DistillationReinforcement LearningDiffusion modelScore-based ModelBenchmark
🎯 What it does: Proposes Reward‑Aware Consistency Trajectory Distillation (RACTD), an offline reinforcement learning method that directly integrates reward optimization into consistency trajectory distillation, enabling the generation of high-reward action trajectories through single-step sampling.
Accelerating Eigenvalue Dataset Generation via Chebyshev Subspace Filter
Hong Wang (University of Science and Technology of China), Zhen huang
Computational EfficiencyPhysics Related
🎯 What it does: Proposed an algorithm called SCSF that leverages operator similarity to accelerate the generation of eigenvalue datasets.
Accelerating Inference for Multilayer Neural Networks with Quantum Computers
Arthur G. Rattew (University of Oxford), Patrick Rebentrost (National University of Singapore)
ClassificationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: Achieved full quantum inference for multi-layer residual networks, quantizing classical architectures such as convolution, activation, and normalization;
Accessible, Realistic, and Fair Evaluation of Positive-Unlabeled Learning Algorithms
Wei Wang (RIKEN), Masashi Sugiyama (RIKEN)
ClassificationImageTabularBenchmark
🎯 What it does: Proposed the first unified positive and unlabeled (PU) learning benchmark, systematically comparing various PU algorithms and evaluating their performance under different validation criteria.
ACCORD: Alleviating Concept Coupling through Dependence Regularization for Text-to-Image Diffusion Personalization
Shizhan Liu (Ant Group), Jianguo Li (Ant Group)
GenerationSupervised Fine-TuningVision Language ModelDiffusion modelImageText
🎯 What it does: Propose the ACCORD framework, which directly suppresses personalization concept coupling issues in text-to-image diffusion models through two regularization losses (DDLoss and PDLoss).
ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall
Jiayu Yang (Hong Kong University of Science and Technology (Guangzhou)), Yutao Yue (Hong Kong University of Science and Technology (Guangzhou))
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes the ACE (Attribution-Controlled Knowledge Editing) framework to efficiently edit LLM knowledge in multi-hop fact recall tasks, addressing the issue of sharply declining performance of traditional knowledge editing methods in multi-hop reasoning involving implicit subjects.
AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy
Zihan Liu (NVIDIA), Wei Ping (NVIDIA)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper systematically studies the synergistic effects of supervised fine-tuning (SFT) and reinforcement learning (RL), constructing and training the AceReason-Nemotron 1.1 7B model, which significantly improves performance on mathematical and code reasoning tasks.
Achieving Approximate Symmetry Is Exponentially Easier than Exact Symmetry
Behrooz Tahmasebi (Harvard University), Melanie Weber (Harvard University)
Computational EfficiencyRepresentation Learning
🎯 What it does: The study investigates the cost difference of enforcing exact symmetry and approximate symmetry in machine learning models, proposing and analyzing a metric called 'averaging complexity'.
Achieving low-bit Muon through subspace preservation and grid quantization
Huaijin Wu (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationComputational EfficiencyText
🎯 What it does: This paper proposes a technique to compress the Muon optimizer to 4-bit precision, called 4-bit-Muon-GRASP, achieving significant memory savings by performing subspace separation and grid quantization on the momentum matrix while maintaining orthogonality.
Achieving Olympia-Level Geometry Large Language Model Agent via Complexity Boosting Reinforcement Learning
Haiteng Zhao (Shanghai AI Laboratory), Kai Chen (Shanghai AI Laboratory)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Proposed and implemented InternGeometry, an LLM-based geometry proof agent that solves IMO-level geometry problems through long-term interaction, dynamic memory, and symbolic engine feedback;
ACPBench Hard: Unrestrained Reasoning about Action, Change, and Planning
Harsha Kokel (IBM Research San Jose), Shirin Sohrabi (IBM Thomas J. Watson Research Center)
Large Language ModelTextBenchmark
🎯 What it does: Constructed ACPBench Hard, an open-ended generative planning reasoning task set encompassing eight tasks: feasibility, progression, reachability, action reachability, verification, simplification, milestones, and next steps.
Action Chunking and Data Augmentation Yield Exponential Improvements in Behavior Cloning for Continuous Spaces
Thomas TCK Zhang, Max Simchowitz (Carnegie Mellon University)
Data-Centric LearningRobotic IntelligenceReinforcement LearningSequentialOrdinary Differential Equation
🎯 What it does: This paper theoretically analyzes and experimentally verifies two key interventions in continuous control behavior cloning—action chunking and expert noise injection—demonstrating their ability to eliminate exponential error accumulation without requiring interactive expert feedback.
Action-aware Dynamic Pruning for Efficient Vision-Language-Action Manipulation
Xiaohuan Pei (University of Sydney), Chang Xu (University of Sydney)
Computational EfficiencyRobotic IntelligenceTransformerVision-Language-Action ModelMultimodality
🎯 What it does: Proposed an Action-Aware Dynamic Pruning (ADP) framework that combines text-driven visual token selection with pruning gates based on execution trajectories to improve the inference efficiency of vision-language-action models.
Action-Free Offline-To-Online RL via Discretised State Policies
Natinael Solomon Neggatu (University of Warwick), Giovanni Montana (University of Warwick)
Reinforcement LearningTabular
🎯 What it does: On datasets without action labels, researchers proposed an offline-to-online action-free reinforcement learning framework that can pretrain state policies from records containing only (s, r, s') and accelerate learning during the online phase.
Action-Guided Attention for Video Action Anticipation
Tsung-Ming Tai (Free University of Bozen-Bolzano), Oswald Lanz (NVIDIA)
RecognitionExplainability and InterpretabilityRepresentation LearningTransformerVideo
🎯 What it does: Propose an attention mechanism based on past action prediction—Action-Guided Attention—which combines adaptive gating to dynamically fuse historical and current visual features for video action anticipation.
Actions as Language: Fine-Tuning VLMs into VLAs Without Catastrophic Forgetting
Asher James Hancock (Princeton University), Anirudha Majumdar (Princeton University)
Robotic IntelligenceTransformerSupervised Fine-TuningVision-Language-Action ModelMultimodality
🎯 What it does: Fine-tune large vision-language models (VLMs) into vision-language-action models (VLAs) by converting low-level actions into natural language descriptions, and avoid catastrophic forgetting using only LoRA fine-tuning.
Actions Speak Louder than Prompts: A Large-Scale Study of LLMs for Graph Inference
Ben Finkelshtein (University of Oxford), Ryen W White
ClassificationTransformerLarge Language ModelPrompt EngineeringGraph
🎯 What it does: This paper systematically evaluates the performance of large language models (LLMs) in node classification tasks through large-scale, controlled experiments. It compares three graph interaction modes (prompting, tool calls, and graph-as-code), analyzes the impacts of dimensions such as different datasets, homogeneity, text length, and model scale, and conducts fine-grained ablation analysis on dependencies of features, structure, and labels.
Activation Function Design Sustains Plasticity in Continual Learning
Lute Lillo (University of Vermont), Nick Cheney (University of Vermont)
Reinforcement LearningImageBenchmark
🎯 What it does: Investigate the impact of activation functions on model plasticity (the ability to learn new tasks) in continual learning environments, and propose two novel smooth leaky activation functions (Smooth-Leaky and Randomized Smooth-Leaky). Their advantages in continual supervised learning and reinforcement learning are theoretically analyzed and experimentally validated.
Activation Steering with a Feedback Controller
Dung Viet Nguyen, Tan Minh Nguyen
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes a control-theory-based activation regulation method called PID Steering, categorizing existing activation regulation techniques as proportional (P) controllers, and further achieving more precise and robust behavior control of large language models (LLMs) and diffusion models by introducing integral (I) and derivative (D) terms.
ActivationReasoning: Logical Reasoning in Latent Activation Spaces
Lukas Helff (TU Darmstadt), Kristian Kersting (TU Darmstadt)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationLarge Language ModelAuto EncoderTextChain-of-Thought
🎯 What it does: Proposed the ACTIVATIONREASONING (AR) framework, embedding logical reasoning into the sparse activation space of large language models. By constructing a concept dictionary, detecting activations, and executing forward chain reasoning, it achieves interpretable, controllable, and safe regulation of internal model reasoning.
Active Learning for Decision Trees with Provable Guarantees
Arshia Soltani Moakhar (Univeristy of Maryland), MohammadTaghi Hajiaghayi (Univeristy of Maryland)
OptimizationTabular
🎯 What it does: The paper proposes a complete theoretical framework for analyzing the label complexity of decision trees in active learning, and presents the first general active learning algorithm that achieves a (1+ε) multiplicative error guarantee within polynomial logarithmic levels.
Active Learning of 3D Gaussian Splatting with Consistent Region Partition and Robust Pose Estimation
Ruiqi Li (Hong Kong Baptist University), Yiu-ming Cheung (Hong Kong Baptist University)
Pose EstimationOptimizationGaussian SplattingImage
🎯 What it does: Propose an online active learning 3D Gaussian Splatting algorithm that selects the next optimal shooting view by segmenting consistent regions and evaluating semantic feature variance, while performing robust pose optimization during actual shooting.
ActiveCQ: Active Estimation of Causal Quantities
Erdun Gao (Australian Institute for Machine Learning), Dino Sejdinovic (Australian Institute for Machine Learning)
OptimizationExplainability and InterpretabilityData-Centric LearningTabular
🎯 What it does: This paper proposes a unified active framework for estimating causal quantities (ActiveCQ), designed to efficiently collect experimental data for estimating multiple causal quantities.
ActiveDPO: Active Direct Preference Optimization for Sample-Efficient Alignment
Xiaoqiang Lin (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: To address the alignment issues in large language models, the ActiveDPO method is proposed, which actively selects preferred data based on uncertainty quantification derived from the LLM's own gradients, thereby improving alignment performance under a limited label budget.
Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making
Fan Feng (University of California San Diego), Kun Zhang (MBZUAI)
Robotic IntelligenceConvolutional Neural NetworkRecurrent Neural NetworkTransformerReinforcement LearningDiffusion modelTime SeriesSequentialBenchmark
🎯 What it does: Propose Ada-Diffuser, a causal diffusion model that can automatically identify and utilize time-evolving latent factors while generating trajectories in decision-making tasks.
AdaBlock-dLLM: Semantic-Aware Diffusion LLM Inference via Adaptive Block Size
Guanxi Lu (Imperial College London), Hongxiang Fan (Imperial College London)
Computational EfficiencyLarge Language ModelDiffusion modelText
🎯 What it does: Propose AdaBlock-dLLM, a scheduler that dynamically adjusts block size during the semi-autoregressive diffusion LLM inference process;
AdaCache: Adaptive Caching and Context Augmentation for Efficient LLM Serving
Zeng Zihao (Nanyang Technological University), Wei Yang Bryan Lim (Nanyang Technological University)
RetrievalComputational EfficiencyTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose AdaCache to improve caching and context augmentation in retrieval-augmented generation (RAG) systems, significantly reducing inference latency.
AdAEM: An Adaptively and Automated Extensible Measurement of LLMs' Value Difference
Jing Yao (Renmin University of China), Xing Xie (Microsoft Research Asia)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This work proposes an adaptive and automatically scalable evaluation framework, AdAEM, to measure internal value orientation differences in large language models (LLMs), automatically generating and expanding high-discrimination assessment questions;
Adapt Data to Model: Adaptive Transformation Optimization for Domain-shared Time Series Foundation Models
Yunzhong Qiu (Tsinghua University), Jianmin Wang (Tsinghua University)
Domain AdaptationOptimizationHyperparameter SearchTransformerTime Series
🎯 What it does: Design a data-transformation-based adaptation framework TATO for frozen large time series models (LTMs), enhancing cross-domain prediction performance through automatic search of transformation pipelines.
Adapting Self-Supervised Representations as a Latent Space for Efficient Generation
Ming Gui (CompVis @ LMU Munich), Björn Ommer (CompVis @ LMU Munich)
GenerationRepresentation LearningTransformerFlow-based ModelContrastive LearningImageText
🎯 What it does: Propose the RepTok framework, which fine-tunes the [cls] token from a pre-trained self-supervised vision Transformer into a single continuous token in the latent space, and jointly trains it with a flow matching decoder to achieve high-quality image reconstruction and generation.
Adaptive Acquisition Selection for Bayesian Optimization with Large Language Models
Giang Ngo (Deakin University), Svetha Venkatesh (Deakin University)
OptimizationHyperparameter SearchTransformerLarge Language ModelPrompt EngineeringTabular
🎯 What it does: This paper proposes a new framework called LMABO, which dynamically selects the most suitable sampling function in the Bayesian Optimization (BO) process using pre-trained large language models in a zero-shot manner;
Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols
Mikhail Terekhov (MATS), Jonas Geiping (ELLIS Institute Tübingen & Max Planck Institute for Intelligent Systems)
Adversarial AttackLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Studied how attackers exploit prompt injection attacks on trusted monitors in AI control protocols, leading to the failure of security protocols;
Adaptive Augmentation-Aware Latent Learning for Robust LiDAR Semantic Segmentation
Wangkai Li (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
SegmentationDomain AdaptationAnomaly DetectionKnowledge DistillationAuto EncoderPoint Cloud
🎯 What it does: Propose the A3Point framework, enhancing robustness in LiDAR semantic segmentation under adverse weather conditions through adaptive augmentation and latent space learning.
Adaptive Canonicalization with Application to Invariant Anisotropic Geometric Networks
Ya-Wei Eileen Lin (Technical University of Munich), Ron Levie (Technion Israel Institute of Technology)
ClassificationGraph Neural NetworkPoint CloudGraph
🎯 What it does: Propose an adaptive canonicalization framework that leverages prior maximization to make the network's normalized input depend on both the input and the network itself, thereby achieving a continuous, symmetry-preserving model with universal approximation properties.
Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning
Wei Yang (University of Southern California), Yan Liu (University of Southern California)
OptimizationMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Proposed the HILA framework, which endows multi-agent systems with metacognitive strategies to learn when to autonomously solve problems and when to consult human experts, achieving continuous learning through a dual-loop strategy optimization.
Adaptive Concept Discovery for Interpretable Few-Shot Text Classification
ZHENG Lifang (Hong Kong University of Science and Technology), Kani Chen (Hong Kong University of Science and Technology)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a structured concept bottleneck model (StructCBM) that performs two-stage text classification using prototype concepts and discriminative concepts generated by LLM, with inference completely independent of LLM;
Adaptive Conformal Anomaly Detection with Time Series Foundation Models for Signal Monitoring.
Natalia Martinez (IBM Research), Roman Vaculin (IBM Research)
Anomaly DetectionTransformerTime SeriesBenchmark
🎯 What it does: Proposes a post-hoc adaptive conformal anomaly detection framework that leverages predictions from pre-trained time series foundation models for zero-shot monitoring, generating interpretable p-value anomaly scores;
Adaptive Conformal Guidance for Learning under Uncertainty
Rui Liu (University of Maryland), Pratap Tokekar (University of Maryland)
Domain AdaptationAutonomous DrivingKnowledge DistillationReinforcement LearningImage
🎯 What it does: Propose the AdaConG method, which quantifies the uncertainty of guidance signals using split conformal prediction, and dynamically adjusts the weight of guidance signals across three tasks: supervised learning, semi-supervised learning, and reinforcement learning, thereby improving model learning performance in noisy or domain shift scenarios.
Adaptive Conformal Prediction via Mixture-of-Experts Gating Similarity
Jingsen Kong (Jinan University), Bei Jiang (University Of Alberta)
Domain AdaptationMixture of ExpertsTabularTime Series
🎯 What it does: Proposes the Mixture-of-Experts Conformal Prediction (MoE-CP) framework, which utilizes MoE gate vectors to perform similarity-weighted calibration of residuals, thereby generating adaptive prediction intervals for heterogeneous multi-domain data.
Adaptive Data-Knowledge Alignment in Genetic Perturbation Prediction
Yuanfang Xiang (Nanjing University), Lun Ai (EMBL-EBI)
Explainability and InterpretabilityData-Centric LearningReinforcement LearningBiomedical Data
🎯 What it does: Proposed the ALIGNED framework, integrating neural networks with symbolic knowledge for adaptive alignment and continuous improvement of gene regulatory networks to predict transcriptional responses to gene perturbations
Adaptive Debiasing Tsallis Entropy for Test-Time Adaptation
Xiangyu Wu (Nanjing University of Science and Technology), Jianfeng Lu (Nanjing University of Science and Technology)
ClassificationDomain AdaptationVision Language ModelImage
🎯 What it does: Propose an adaptive test-time adaptation method ADTE based on adaptive bias-corrected Tsallis entropy to address the entropy estimation bias in VLM during TTA, thereby improving high-confidence view selection and final prediction performance.
Adaptive Domain Shift in Diffusion Models for Cross-Modality Image Translation
Zihao WANG (University of Tennessee), Shaogang Ren (University of Tennessee)
Image TranslationDomain AdaptationDiffusion modelMultimodalityBiomedical DataMagnetic Resonance ImagingStochastic Differential Equation
🎯 What it does: Propose a cross-modal image translation method that embeds adaptive domain transfer dynamics into diffusion models.
Adaptive Gaussian Expansion for On-the-fly Category Discovery
Chunming Li (Nanjing University of Science and Technology), Haofeng Zhang (Nanjing University of Science and Technology)
RecognitionAnomaly DetectionTransformerContrastive LearningImage
🎯 What it does: Propose the Adaptive Gaussian Expansion (AGE) framework, decomposing the On-the-fly Category Discovery (OCD) task into open-set recognition and real-time novel category discovery. It utilizes soft threshold detection to identify known categories and sends anomalous samples to AGE for incremental Gaussian clustering and category inference.
Adaptive gradient descent on Riemannian manifolds and its applications to Gaussian variational inference
Jiyoung Park (Texas A&M University), Shiqian Ma (Rice University)
OptimizationTabular
🎯 What it does: Proposed RAdaGD, an adaptive gradient descent algorithm on Riemannian geometry, and first proved non-average O(1/k) convergence under conditions of local geodesic smoothness and generalized geodesic convexity; applied it to Gaussian Variational Inference (GVI), providing convergence guarantees when the target log-density does not satisfy L-smoothness.
Adaptive Hopfield Network: Rethinking Similarities in Associative Memory
Shurong Wang (Zhejiang University), Guoqi Li (Institute of Automation Chinese Academy of Sciences)
ClassificationRetrievalRecurrent Neural NetworkImageTabular
🎯 What it does: Propose a 'correct retrieval' theory based on mutation distribution, design a learnable adaptive similarity, and construct the Adaptive Hopfield Network (A-Hop) to achieve more precise pattern retrieval and multi-task learning.
Adaptive Logit Adjustment for Debiasing Multimodal Language Models
Hoin Jung (Purdue University), Xiaoqian Wang (Purdue University)
Explainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Proposes Adaptive Logit Adjustment (ALA), a post-hoc method that dynamically adjusts logits during autoregressive text generation to mitigate bias in multimodal language models.
Adaptive Mamba Neural Operators
Zeyuan Song (Oklahoma State University), Zheyu Jiang (Oklahoma State University)
Finance RelatedPhysics Related
🎯 What it does: Propose the Adaptive Mamba Neural Operator (AMO), a neural operator that integrates Adaptive Fourier Decomposition (AFD) with the Mamba architecture;
Adaptive Methods Are Preferable in High Privacy Settings: An SDE Perspective
Enea Monzio Compagnoni (University of Basel), Anastasia Koloskova
OptimizationSafty and PrivacyTextStochastic Differential Equation
🎯 What it does: Conducts theoretical analysis based on stochastic differential equations (SDE) for differential privacy (DP) optimizers, particularly DP-SGD and DP-SignSGD (along with experimental extensions to DP-Adam), comparing their convergence rates and privacy-utility trade-offs under two protocols: fixed hyperparameters and optimal tuning.
Adaptive Mixture of Disentangled Experts for Dynamic Graph Out-of-Distribution Generalization
Haibo Chen (Tsinghua University), Wenwu Zhu (Tsinghua University)
Domain AdaptationRepresentation LearningGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: Propose AdaMix—a self-adaptive Mixture-of-Experts (MoE) framework for out-of-distribution (OOD) generalization in dynamic graphs, which dynamically routes expert networks and learns invariant patterns based on distribution shifts of nodes across different time points.
Adaptive Moments are Surprisingly Effective for Plug-and-Play Diffusion Sampling
Christian Belardi (Cornell University), Carla P Gomes
RestorationGenerationDiffusion modelImage
🎯 What it does: Investigated using adaptive momentum estimation (Adam) to stabilize noise gradients in plug-and-play diffusion sampling, improving performance in various inverse problems and class-conditional generation tasks.
Adaptive Nonlinear Compression for Large Foundation Models
Liang Xu (Chinese Academy of Sciences), Shuhui Wang (Zhejiang University)
CompressionTransformerImageText
🎯 What it does: Propose a compression framework called NLA based on nonlinear low-rank approximation combined with adaptive budget allocation to compress large foundational models;
Adaptive Regularization for Large-Scale Sparse Feature Embedding Models
Mang Li (Institute of Intelligent Technology Alibaba International Digital Commerce Group), Wei Lyu (Institute of Intelligent Technology Alibaba International Digital Commerce Group)
Recommendation SystemTabular
🎯 What it does: For CTR/CVR estimation models based on large-scale sparse features, this paper proposes an adaptive regularization method and provides its theoretical basis and experimental validation.
Adaptive Rollout Allocation for Online Reinforcement Learning with Verifiable Rewards
Hieu Trung Nguyen (Chinese University of Hong Kong), Viet Anh Nguyen (Chinese University of Hong Kong)
OptimizationComputational EfficiencyReinforcement LearningText
🎯 What it does: Propose the VIP framework, which utilizes Gaussian processes to predict the success probability and dynamically allocates the rollout budget for each prompt through convex optimization that minimizes gradient variance, thereby improving the sampling efficiency of RLVR.
Adaptive Scaling of Policy Constraints for Offline Reinforcement Learning
Tan Jing (University of Science and Technology Beijing), Zhaolin Yuan (University of Science and Technology Beijing)
Reinforcement LearningContrastive Learning
🎯 What it does: Proposes the Adaptive Scaling of Policy Constraints (ASPC) framework, which utilizes second-order differentiation to optimize the adaptive adjustment of the ratio between RL objectives and behavior cloning (BC) constraints in offline reinforcement learning, achieving a unified hyperparameter configuration across different datasets and tasks.
Adaptive Social Learning via Mode Policy Optimization for Language Agents
Minzheng Wang (University of Chinese Academy of Sciences), Wenji Mao (Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the Adaptive Social Learning (ASL) framework, enabling language models to adaptively adjust reasoning depth in dynamic social contexts, and guide dialogue strategies through four reasoning modes designed based on a four-tier cognitive control theory;
Adaptive Test-Time Training for Predicting Need for Invasive Mechanical Ventilation in Multi-Center Cohorts
Xiaolei Lu (University of California San Diego), Shamim Nemati (University of California San Diego)
ClassificationDomain AdaptationAuto EncoderBiomedical DataElectronic Health Records
🎯 What it does: Proposed an adaptive test-time training framework (AdaTTT) for predicting the need for invasive mechanical ventilation (IMV) in ICU patients within 24 hours
Adaptive Thinking: Large Language Models Know When to Think in Latent Space
Pingzhi Li (Apple), Xianzhi Du (Apple)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper studies how large language models adaptively allocate their 'thinking' budget during inference, i.e., deciding whether to generate chain-of-thought (CoT) and how many thinking tokens are needed based on the difficulty of the query, and proposes the Sonata framework to realize real-time budget allocation.
Adaptive Width Neural Networks
Federico Errica (NEC Laboratories Europe), Francesco Alesiani (NEC Laboratories Europe)
Computational EfficiencyNeural Architecture SearchImageTextGraphTabularSequential
🎯 What it does: Proposed a new method that dynamically adjusts the width of each layer by learning the unbounded width of neural network layers during training.
AdaRank: Adaptive Rank Pruning for Enhanced Model Merging
Chanhyuk Lee (Korea Advanced Institute Of Science And Technology), Seunghoon Hong (Korea Advanced Institute Of Science And Technology)
TransformerImageText
🎯 What it does: Propose AdaRank, which introduces a learnable binary mask into the SVD-based model fusion, adaptively trimming the singular components of the task vector to achieve better multi-task model fusion;
AdaReasoner: Dynamic Tool Orchestration for Iterative Visual Reasoning
Mingyang Song (Fudan University), Yu Cheng (Chinese University of Hong Kong)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelImageTextMultimodality
🎯 What it does: Constructed the AdaReasoner framework, enabling multi-modal large language models to achieve iterative visual reasoning through dynamic tool orchestration.
AdaSpec: Adaptive Spectrum for Enhanced Node Distinguishability
Fangbing Liu (Australian National University), Qing Wang (Australian National University)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Study the node distinguishability of spectral graph neural networks and propose the AdaSpec adaptive spectral matrix module to enhance node distinguishability.
AdaViewPlanner: Adapting Video Diffusion Models for Viewpoint Planning in 4D Scenes
Yu Li (Tsinghua University), Yujiu Yang (Tsinghua University)
GenerationTransformerVision Language ModelDiffusion modelFlow-based ModelVideoText
🎯 What it does: Repurpose a pre-trained text-to-video model into a view planner to generate camera trajectories in 4D scenes that match human motion.
Addressing divergent representations from causal interventions on neural networks
Satchel Grant (Stanford University), Christopher Potts (Stanford University)
Explainability and InterpretabilityRepresentation LearningTransformerText
🎯 What it does: This paper demonstrates that in various causal intervention methods, post-intervention internal representations often deviate from the model's natural distribution, and provides theoretical and experimental analysis of this bias; subsequently, an improved strategy is proposed by applying a Counterfactual Latent (CL) loss to the intervention results, aiming to reduce bias and minimize the activation of hidden paths and dormant behaviors.
Addressing Pitfalls in the Evaluation of Uncertainty Estimation Methods for Natural Language Generation
Mykyta Ielanskyi (Johannes Kepler University Linz), Sepp Hochreiter (Johannes Kepler University Linz)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper systematically analyzes and improves the evaluation protocol for uncertainty estimation methods in natural language generation tasks, pointing out the biases and noise present in current approximate correctness assessments based on QA datasets.
ADEPT: Continual Pretraining via Adaptive Expansion and Dynamic Decoupled Tuning
Jinyang Zhang (Peking University), Xu Chu (Peking University)
Domain AdaptationTransformerLarge Language ModelTextBiomedical DataBenchmark
🎯 What it does: Proposed and implemented a continuous pre-training framework called ADEPT for large language models, focusing on reducing catastrophic forgetting and improving domain performance during domain adaptation.
Adjusting Prediction Model Through Wasserstein Geodesic for Causal Inference
Yuguang Yan (Guangdong University of Technology), Zhifeng Hao (Shantou University)
Domain AdaptationTabular
🎯 What it does: Propose generating a continuous intermediate distribution between the control group and the treatment group through Wasserstein geodesic, and gradually adjust the result prediction model by adopting self-training and confidence filtering on these intermediate groups to enhance the generalization ability of causal inference.
ADM-v2: Pursuing Full-Horizon Roll-out in Dynamics Models for Offline Policy Learning and Evaluation
Haoxin Lin (Nanjing University), Yang Yu (University of Montréal)
Recurrent Neural NetworkReinforcement LearningWorld ModelSequential
🎯 What it does: Propose ADM-v2 and PARoll methods that support reliable full-length roll-out, applied to policy evaluation and optimization in offline reinforcement learning.
AdPO: Enhancing the Adversarial Robustness of Large Vision-Language Models with Preference Optimization
Chaohu Liu (University of Science and Technology of China), Linli Xu (University of Science and Technology of China)
OptimizationAdversarial AttackTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes AdPO, an adversarial defense method based on preference optimization, specifically enhancing the robustness of large vision-language models (such as LLaVA, Qwen-2.5-VL) against visual adversarial examples while maintaining high performance on clean inputs.
Ads that Stick: Near-Optimal Ad Optimization through Psychological Behavior Models
Kailash Gopal Darmasubramanian (Indian Institute of Technology Madras), Arpit Agarwal (Indian Institute of Technology Bombay)
Recommendation SystemOptimization
🎯 What it does: This paper proposes an approximately optimal algorithm for generating ad scheduling timelines based on principles of human psychology (mere exposure effect, hedonic adaptation, operant conditioning), and verifies its effectiveness through theoretical analysis and simulation experiments.
AdS-GNN - a Conformally Equivariant Graph Neural Network
Maksim Zhdanov (University of Amsterdam), Patrick Forré (University of Amsterdam)
Graph Neural NetworkImagePoint CloudPhysics Related
🎯 What it does: Designed and implemented a graph neural network called ADS-GNN and its Clifford algebra extension AdS-CEGNN, which achieves global conformal equivariance by elevating Euclidean point clouds to Anti-de Sitter space and leveraging its isometric transformations.
Advancing Complex Video Object Segmentation via Progressive Concept Construction
Zhixiong Zhang (Shanghai Jiao Tong University), Jiaqi Wang (Shanghai Innovation Institute)
SegmentationTransformerVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: Propose the Segment Concept (SeC) framework, which leverages a large audio-visual language model (LVLM) to progressively build concept-level object representations in video sequences and injects them into the low-level pixel-matching VOS network; simultaneously creates a new SeCVOS benchmark specifically to evaluate models' concept reasoning capabilities under multi-camera and frequent scene change scenarios.
Advancing Multi-agent Traffic Simulation via R1-Style Reinforcement Fine-Tuning
Muleilan Pei (Hong Kong University of Science and Technology Voyager Research Didi Chuxing), Shaojie Shen (Hong Kong University of Science and Technology Voyager Research Didi Chuxing)
Autonomous DrivingOptimizationTransformerSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Developed the SMART-R1 framework, which significantly improves multi-agent traffic simulation quality through R1-style SFT-RFT-SFT iterative fine-tuning.
Advancing Spatiotemporal Representations in Spiking Neural Networks via Parametric Invertible Transformation
Yinsong Yan (Hong Kong Polytechnic University), Jibin Wu (Hong Kong Polytechnic University)
ClassificationRepresentation LearningSpiking Neural NetworkFlow-based ModelImage
🎯 What it does: Proposes integrating Parameter Invertible Transformation (PIT) into the synaptic dynamics of spiking neural networks to enhance their spatiotemporal representation capabilities.
Advancing Universal Deep Learning for Electronic-Structure Hamiltonian Prediction of Materials
Shi Yin (Hefei Comprehensive National Science Center), Lixin He (University of Science and Technology of China)
TransformerGraphPhysics Related
🎯 What it does: Propose the NextHAM framework for efficiently and generally predicting material electronic structure Hamiltonians, and release a dataset covering over 60 elements with 17,000 structures including spin-orbit coupling.
AdvChain: Adversarial Chain-of-Thought Tuning for Robust Safety Alignment of Large Reasoning Models
Zihao Zhu (Chinese University of Hong Kong Shenzhen), Baoyuan Wu (Chinese University of Hong Kong Shenzhen)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposes AdvChain, a secure alignment method that trains large-scale reasoning models to dynamically self-correct through adversarial chain-of-thought (CoT) fine-tuning.
Adversarial Attacks Already Tell the Answer: Directional Bias-Guided Test-time Defense for Vision-Language Models
Liangsheng Liu (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
Adversarial AttackVision Language ModelMultimodalityBenchmark
🎯 What it does: For test-time defense of Vision-Language models (e.g., CLIP), reconstruct more robust features to resist adversarial attacks by leveraging feature direction bias generated through multiple input transformations.
Adversarial Déjà Vu: Jailbreak Dictionary Learning for Stronger Generalization to Unseen Attacks
Mahavir Dabas (Virginia Tech), Ruoxi Jia (Virginia Tech)
Safty and PrivacyAdversarial AttackLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By constructing an interpretable 'jailbreak dictionary' and performing skill-combination training (ASCoT) based on it, the robustness of large language models against unseen jailbreak attacks is enhanced.
Adversarial Encoding Perturbation and Synthesis for Set Representation Auxiliary Learning
Yankai Chen (MBZUAI), Xue Liu (MBZUAI)
RetrievalRecommendation SystemRepresentation LearningContrastive LearningPoint Cloud
🎯 What it does: Proposes SRAL, an auxiliary learning framework that captures inter-set correlations through 2-slice Wasserstein distance and adversarial encoding perturbation.
Adversarially Pretrained Transformers May Be Universally Robust In-Context Learners
Soichiro Kumano (University of Tokyo), Toshihiko Yamasaki (University of Tokyo)
ClassificationAdversarial AttackTransformerImage
🎯 What it does: This paper theoretically proves that a single-layer linear Transformer, after adversarial pre-training on multi-class classification tasks, can achieve robustness on previously unseen classification tasks through in-context learning without additional adversarial training.
AEGIS: Adversarial Target-Guided Retention-Data-Free Robust Concept Erasure from Diffusion Models
Fengpeng Li (University of Macau), Jiantao Zhou (University of Macau)
GenerationAdversarial AttackSupervised Fine-TuningPrompt EngineeringDiffusion modelImageText
🎯 What it does: This paper proposes a framework named AEGIS, aiming to achieve concept erasure in diffusion models while simultaneously enhancing robustness against adversarial prompting attacks and maintaining the model's ability to retain irrelevant concepts.
Aegis: Automated Error Generation and Attribution for Multi-Agent Systems
Fanqi Kong (Peking University), Xue Feng (Beijing Institute of General Artificial Intelligence)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextSequential
🎯 What it does: Constructed a multi-agent system error dataset (AEGIS) capable of automatically injecting errors and generating verifiable labels, and conducted error attribution research based on this dataset.
AetherCode: Evaluating LLMs’ Ability to Win In Premier Programming Competitions
Zihan Wang (Peking University), Ming Ding (ByteDance)
AI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Propose the AetherCode benchmark, collecting top-tier competition problems from IOI, ICPC, etc., and equipping them with high-quality test cases verified by experts to more rigorously evaluate the code reasoning ability of LLMs.