These 2283 NeurIPS 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 NeurIPS 2025 paper, free trial on arXivSub.
$\boldsymbol{\lambda}$-Orthogonality Regularization for Compatible Representation Learning
Simone Ricci (University of Florence), Alberto Del Bimbo (University of Florence)
π― What it does: Achieve compatible representations between independently trained retrieval models by proposing Ξ»-orthogonal regularization that combines affine/orthogonal transformations with supervised contrastive loss, and provide an efficient partial filling strategy.
$\epsilon$-Seg: Sparsely Supervised Semantic Segmentation of Microscopy Data
Sheida RahnamaiKordasiabi, Florian Jug (Human Technopole)
CodeSegmentationAuto EncoderContrastive LearningImageBiomedical Data
π― What it does: A sparse supervised semantic segmentation method, Ο΅βSeg, was designed and evaluated. It utilizes a hierarchical variational autoencoder combined with center region mask reconstruction, GMM prior, contrastive learning, and an MLP segmentation head, achieving high-quality segmentation with only a minimal amount of pixel-level annotations.
$\mathcal{X}^2$-DFD: A framework for e$\mathcal{X}$plainable and e$\mathcal{X}$tendable Deepfake Detection
Yize Chen (Chinese University of Hong Kong), Baoyuan Wu (University at Buffalo)
CodeAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningImageVideoMultimodality
π― What it does: A three-stage multimodal large language model (MLLM) deepfake detection framework called X2-DFD has been constructed, which can automatically evaluate, enhance, and supplement the model's understanding of deepfake features and generate interpretable detection results.
π― What it does: This paper proposes a new parameterization method Β΅ PC, which enables the stable training of predictive coding networks (PCNs) at depths exceeding 100 layers, thereby breaking the scale limitations of traditional PC.
$\text{G}^2\text{M}$: A Generalized Gaussian Mirror Method to Boost Feature Selection Power
Hongyu Shen (University of Illinois), Zhizhen Zhao (University of Illinois)
CodeBiomedical Data
π― What it does: A general Gaussian mirror (G2M) method is proposed, improving the variance assumption of traditional data splitting and Gaussian mirror statistics, achieving stronger feature selection efficacy.
$\text{S}^2$Q-VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation
Weilun Feng (Institute of Computing Technology Chinese Academy of Sciences), Yongjun Xu (Institute of Computing Technology Chinese Academy of Sciences)
π― What it does: A post-training quantization framework for video diffusion transformers, named S2Q-VDiT, is proposed, achieving low-bit quantization with almost no performance loss through significant data selection and sparse label distillation techniques.
$\textit{HiMaCon:}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data
Ruizhe Liu (Hong Kong University), Yanchao Yang (Hong Kong University)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerContrastive LearningMultimodalityBenchmark
π― What it does: A self-supervised framework HiMaCon is proposed to learn hierarchical mechanical operation concepts from unlabeled multimodal demonstrations and to enhance the generalization ability of the policy.
π― What it does: The Hyper-GoalNet framework is proposed, which dynamically generates policy network parameters based on target images through a hypernetwork, enabling goal-conditioned manipulation policy learning.
$\texttt{G1}$: Teaching LLMs to Reason on Graphs with Reinforcement Learning
Xiaojun Guo (Peking University), Yisen Wang (Peking University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningGraph
π― What it does: By applying reinforcement learning to a large number of synthetic graph theory tasks, the ability of large language models in graph structure reasoning is enhanced;
CodeOptimizationAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraph
π― What it does: This paper proposes the STRCMP framework, which automatically generates algorithm code that conforms to solver syntax and problem topology by embedding combinatorial optimization structures extracted by graph neural networks into large language models, and iteratively improves it using evolutionary algorithms.
$Q\sharp$: Provably Optimal Distributional RL for LLM Post-Training
Jin Peng Zhou (Cornell University), Wen Sun (Databricks)
CodeLarge Language ModelReinforcement LearningText
π― What it does: A KL regularized value function Qβ― method based on distributed RL is proposed to improve the alignment and inference of LLM after training.
3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs
Mehdi Makni (Massachusetts Institute of Technology), Rahul Mazumder (Massachusetts Institute of Technology)
CodeCompressionOptimizationTransformerLarge Language ModelText
π― What it does: An efficient post-training method 3BASiL-TM is proposed for the sparse plus low-rank (S + LR) decomposition of large language models, aimed at reducing performance degradation compared to dense models.
3D Gaussian Flats: Hybrid 2D/3D Photometric Scene Reconstruction
Maria Taktasheva (Simon Fraser University), Andrea Tagliasacchi (Simon Fraser University)
CodeSegmentationData SynthesisOptimizationGaussian SplattingSimultaneous Localization and MappingPoint CloudMesh
π― What it does: A hybrid 2D/3D Gaussian plane representation is proposed, which jointly optimizes the in-plane constraints of 2D Gaussians and free-form 3D Gaussians for realistic view synthesis and mesh extraction in indoor scenes.
π― What it does: A molecular relationship learning framework (3DMRL) utilizing a virtual interactive geometric environment for 3D pre-training is proposed, allowing a 2D MRL model to acquire 3D interaction information through global contrast and local relative geometric predictions.
π― What it does: A 3D molecular graph autoencoder, 3D-GSRD, was developed, which combines selective re-masking decoding and structure-agnostic decoding to perform self-supervised pre-training on molecular 3D structures, followed by fine-tuning on downstream tasks using MD17 and QM9.
3D-Prover: Diversity Driven Theorem Proving With Determinantal Point Processes
Sean Lamont (Australian National University), Michael Norrish (Australian National University)
CodeTransformerTabular
π― What it does: This paper proposes a diversity-driven proof strategy filter, 3D-Prover, based on Determinantal Point Process (DPP), which can filter out semantically diverse and high-quality proof strategies before the proof search.
π― What it does: Proposes a 4D-VLA framework that utilizes RGB-D sequences and 3D coordinate embeddings for cross-scene pre-training, constructing a multi-view evaluation benchmark MV-Bench to achieve a more efficient visual language action model.
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A lossless compression framework DFloat11 based on entropy coding is proposed, which compresses LLM and DM weights to about 30% of their size while maintaining 100% accuracy and achieving GPU-efficient inference.
π― What it does: The Bayesian Fast-Slow Framework (BFSF) is proposed to address the interference problem in non-stationary MDPs by dynamically selecting a 'fast' strategy based on recent data and a 'slow' strategy obtained through meta-reinforcement learning;
A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective
Lianghe Shi (University of Michigan), Qing Qu (University of Michigan)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: The study investigates the transition of diffusion models from generalization to memorization in recursive training loops, proposing an entropy-based data subset selection method to alleviate model collapse.
π― What it does: A self-supervised communication framework CELEBI based on Compression-Expression-Iterative Learning is proposed, utilizing three mechanisms: advanced decoding, final state imitation, and message diversity maximization to promote the compressibility, expressiveness, and efficiency of language.
π― What it does: A counterfactual semantics for hybrid systems is proposed, which can perform causal inference on dynamically triggered instantaneous interventions.
A data and task-constrained mechanistic model of the mouse outer retina shows robustness to contrast variations
Kyra L. Kadhim (University of TΓΌbingen), Philipp Berens (University of TΓΌbingen)
CodeClassificationOptical FlowImage
π― What it does: A fully differentiable biophysical model containing 200 cones, horizontal cells, and bipolar cells was constructed and trained to classify MNIST images with varying contrast and brightness.
A Data-Driven Prism: Multi-View Source Separation with Diffusion Model Priors
Sebastian Wagner-Carena (Flatiron Institute), Sydney Erickson (Stanford University)
CodeDiffusion modelScore-based ModelImage
π― What it does: This paper proposes a multi-view source separation framework (DDPRISM) that utilizes diffusion models as priors, capable of learning and inferring the prior and posterior distributions of independent sources from multi-view observations based solely on a known linear mixing matrix and noise covariance.
A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning
Daniel Tschernutter (Infermedica), Maciej KasiΕski (Infermedica)
CodeOptimizationComputational EfficiencyText
π― What it does: This paper proposes an energy model inference algorithm called DCAReasoner based on Differential Convex (DC) functions, aimed at efficiently solving energy minimization problems in continuous inference tasks.
A Differential and Pointwise Control Approach to Reinforcement Learning
Minh Phuong Nguyen (University of Texas at Austin), Chandrajit L. Bajaj (University of Texas at Austin)
CodeOptimizationReinforcement LearningPhysics Related
π― What it does: A Differential Reinforcement Learning (Differential RL) framework is proposed, along with the corresponding Differential Policy Optimization (dfPO) algorithm, for reinforcement learning in continuous state-action spaces. Experiments are conducted on three types of scientific computing tasks: surface modeling, multiscale grid control, and molecular dynamics.
A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking
Gal Fadlon (Ben-Gurion University of Negev), Omri Azencot (Ben-Gurion University of Negev)
CodeGenerationData SynthesisTransformerDiffusion modelTime SeriesSequentialFinance Related
π― What it does: A two-step framework is proposed, first using a time series Transformer to complete irregularly sampled sequences, and then using a visual diffusion model (ImagenTime) combined with masking to generate regular time series.
π― What it does: MomentNet is proposed to directly estimate graph structures through subgraph counting (moments) matching and implicit neural representations, and based on this, MomentMixup is introduced for graph data augmentation.
A Gradient Guidance Perspective on Stepwise Preference Optimization for Diffusion Models
Joshua Tian Jin Tee (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
CodeGenerationOptimizationDiffusion modelImage
π― What it does: This paper re-theorizes Stepwise Preference Optimization (SPO) from the perspective of gradient guidance and proposes GradSPO to more efficiently achieve human preference alignment in diffusion models.
π― What it does: A framework called GGDOpt based on gradient-guided diffusion models is proposed to solve opportunity-constrained programming problems with unknown distributions.
π― What it does: This paper addresses the scenario of multi-source transfer learning, investigating and solving how many samples need to be sampled from each source task to achieve optimal generalization performance on the target task. Based on a theoretical framework, the OTQMS algorithm is proposed to implement adaptive sample size selection.
A Implies B: Circuit Analysis in LLMs for Propositional Logical Reasoning
Guan Zhe Hong (Purdue University), Rina Panigrahy (Google Research)
CodeTransformerLarge Language ModelText
π― What it does: This study investigates the internal reasoning circuits of large pre-trained language models in propositional logic reasoning tasks.
A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders
David Chanin (University College London), Joseph Isaac Bloom
CodeExplainability and InterpretabilityRepresentation LearningLarge Language ModelAuto EncoderText
π― What it does: This study investigates and quantifies the phenomenon of 'feature absorption' in Sparse Autoencoders (SAE), exploring the mechanisms caused by hierarchical features and sparsity loss, and proposes a metric for detecting this phenomenon.
π― What it does: A learning-enhanced dynamic programming algorithm DP-NG-ML is proposed, which uses deep learning to predict ng-sets to accelerate the solution of OPTW.
A Little Depth Goes a Long Way: The Expressive Power of Log-Depth Transformers
William Merrill (Allen Institute for AI), Ashish Sabharwal (Allen Institute for AI)
CodeTransformerGraphSequential
π― What it does: This paper studies the expressive power of Transformers when the depth grows logarithmically with the input length (log-depth), and proves that even a fully unified Transformer (parameter sharing) requires only O(log n) layers to recognize regular languages and solve graph connectivity problems.
π― What it does: This paper studies a machine learning-based multi-agent path planning method to find the shortest or approximately shortest paths on large Cayley graphs, particularly for Rubik's Cubes of dimensions 4x4x4 and 5x5x5.
π― What it does: This paper proposes a new partition cover form of tokenization optimization model and presents a greedy approximation algorithm called GREEDTOK.
A Pre-training Framework for Relational Data with Information-theoretic Principles
Quang Truong (Michigan State University), Jiliang Tang (Michigan State University)
CodeGraph Neural NetworkTabularBenchmark
π― What it does: A self-supervised pre-training framework for relational databases called Task Vector Estimation (TVE) is proposed, which generates task vectors by combining schema graphs, temporal dynamics, and task heterogeneity to train the model.
π― What it does: A target intervention paradigm based on Multi-Agent Influence Diagrams (MAID) is proposed, and a Pre-Strategy Intervention (PSI) method is implemented, using interventions from a single agent to guide the entire multi-agent system to achieve predetermined composite goals.
A Principled Path to Fitted Distributional Evaluation
Sungee Hong (Texas A&M University), Raymond K. W. Wong (Texas A&M University)
CodeReinforcement LearningTabular
π― What it does: This paper proposes a unified framework that extends Fitted Q Evaluation (FQE) to Distributed Offline Policy Evaluation (FDE), and provides various feasible divergence metrics and convergence analysis.
A Semantic Parsing Framework for End-to-End Time Normalization
Xin Su, Steven Bethard (University of Arizona)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: In this paper, the authors propose an end-to-end temporal normalization method that treats time expression recognition and normalization as a code generation task. They use the executable SCATE Python library to directly convert natural language time vocabulary into executable time expression code, and obtain precise time intervals through code execution.
A Set of Generalized Components to Achieve Effective Poison-only Clean-label Backdoor Attacks with Collaborative Sample Selection and Triggers
Zhixiao Wu (Harbin Institute of Technology), Guangming Lu (Harbin Institute of Technology)
CodeAdversarial AttackImage
π― What it does: This paper proposes a set of general components that collaborate sample selection and triggers to enhance the success rate and stealth of clean-label poisoning attacks.
π― What it does: This study investigates the oversmoothing problem in deep Graph Neural Networks (GNNs) and proposes a method called Structure Balanced Symbolic Graph Propagation (SBP) to alleviate oversmoothing while maintaining the distinguishability of node representations.
A Single-Loop First-Order Algorithm for Linearly Constrained Bilevel Optimization
Wei Shen (University of Virginia), Cong Shen (University of Virginia)
CodeOptimizationHyperparameter SearchTabular
π― What it does: This paper studies a bilevel optimization problem with linear coupling constraints and proposes a single-loop first-order algorithm SFLCB.
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackRecurrent Neural NetworkReinforcement LearningWorld ModelImage
π― What it does: A robust imitation learning method named SAILOR was trained, utilizing a world model, reward model, and MPPI planning learned from expert demonstrations, searching for residual plans during testing to correct errors in the baseline policy.
A Snapshot of Influence: A Local Data Attribution Framework for Online Reinforcement Learning
Yuzheng Hu (University of Illinois Urbana Champaign), Han Zhao (University of Illinois Urbana Champaign)
CodeOptimizationComputational EfficiencyLarge Language ModelReinforcement LearningSequential
π― What it does: This study investigates data attribution in online reinforcement learning, proposing a local attribution framework for PPO and designing an iterative influence filtering algorithm based on it to enhance training efficiency and performance.
π― What it does: A model-free differential temporal difference (dTD) method based on the Hamilton-Jacobi-Bellman equation is proposed for continuous-time reinforcement learning.
π― What it does: This paper proposes the TRIANGLE method, which achieves tri-modal alignment by calculating the area of triangles formed by tri-modal embeddings in high-dimensional space, thereby directly completing multi-modal matching in the embedding space.
A Unified Analysis of Stochastic Gradient Descent with Arbitrary Data Permutations and Beyond
Yipeng Li (Shenzhen International Graduate School Tsinghua University), Zhenyu Liu (Shenzhen International Graduate School Tsinghua University)
CodeOptimizationFederated LearningImageTabular
π― What it does: This paper proposes a unified convergence analysis framework that encompasses all permutation-based stochastic gradient descent (SGD) algorithms (including arbitrary permutations, independent permutations, and dependent permutations such as GraBs) as well as client permutations in federated learning (FL) with regularized client participation;
π― What it does: A unified framework is proposed to translate various first-order causal effect measures (such as risk difference, risk ratio, odds ratio, etc.) from randomized controlled trials to target populations under covariate distribution shift.
π― What it does: A proxy memory system named A-MEM is proposed, allowing large language model agents to dynamically create, link, and evolve memory nodes without predefined storage and retrieval patterns, thereby supporting long-term interactions.
π― What it does: This paper proposes a chain-of-thought compression method based on A* search, called A*-Thought, which transforms lengthy reasoning into concise and efficient inference by first evaluating the bidirectional importance of each step and then searching for thinking paths with high information density and low cost on the search tree.
Hongming Piao (City University of Hong Kong), Ying Wei (Zhejiang University)
CodeLarge Language ModelTextBiomedical Data
π― What it does: The AΒ³E model editing framework is proposed to address the issues of knowledge loss, pre-existing errors, and knowledge drowning in large language models under multiple editing scenarios.
Abstain Mask Retain Core: Time Series Prediction by Adaptive Masking Loss with Representation Consistency
Renzhao Liang (Beihang University), Takahiro Yabe (New York University)
CodeTransformerTime Series
π― What it does: The AMRC method is proposed, which suppresses redundant features in time series prediction through adaptive masking loss and representation consistency constraints, thereby improving prediction accuracy.
π― What it does: An abstract rendering framework is proposed that can generate verifiable image sets for 3D scenes (Gaussian Splats and NeRF) under continuous changes in camera pose or scene parameters, and provides formal safety guarantees for downstream visual models (classification, detection, pose estimation) through this abstract image.
π― What it does: A local personalized PageRank (PPR) computation framework based on the nested evolution set process (AESP) is proposed, utilizing approximate accelerated proximal point iteration for faster solving.
π― What it does: The DecVFAL framework is proposed, which significantly accelerates vertical federated adversarial training (VFAL) through a dual-layer decoupling (lazy sequential backpropagation and decoupled parallel backpropagation) while maintaining model robustness.
Accelerating 3D Molecule Generative Models with Trajectory Diagnosis
Zhilong Zhang (Tsinghua University), Wei-Ying Ma (Tsinghua University)
CodeGenerationComputational EfficiencyDrug DiscoveryGraph Neural NetworkFlow-based ModelGraphBiomedical Data
π― What it does: A method is proposed that divides the 3D molecular generation process into a rearrangement phase and a feature adjustment phase, with acceleration strategies designed for each phase, significantly improving generation efficiency.
π― What it does: A fast feature consistency prediction (FFCP) based on gradients is proposed, which accelerates traditional feature consistency prediction (FCP) through first-order Taylor expansion, achieving faster confidence interval construction.
π― What it does: Designed and implemented CompactFusionβa parallel diffusion model inference acceleration framework based on residual compression, which reduces communication volume while maintaining high generation quality.
π― What it does: A new reinforcement learning framework Aβ-PO is proposed, which directly approximates the optimal advantage function through a two-stage process, eliminating the need for explicit value networks and the overhead of multiple generations;
Acceleration via silver step-size on Riemannian manifolds with applications to Wasserstein space
Jiyoung Park (Texas A&M University), Anirban Bhattacharya (Texas A&M University)
CodeOptimization
π― What it does: This study investigates the use of silver step size accelerated optimization algorithms on Riemannian manifolds, particularly for applications in Wasserstein spaces.
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A new distributed LLM training algorithm called ACCO is proposed, which achieves overlapping of gradient communication and computation without increasing additional memory overhead, thereby reducing GPU idle time and supporting heterogeneous hardware.
Accurate KV Cache Eviction via Anchor Direction Projection for Efficient LLM Inference
Zijie Geng (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: A new KV cache eviction method called AnDPro is proposed, which introduces a projection-based scoring function to more accurately measure the importance of tokens, thereby optimizing memory usage and inference latency of large language models.
Accurately Predicting Protein Mutational Effects via a Hierarchical Many-Body Attention Network
Dahao Xu (Sun Yat-sen University), Yuedong Yang (Sun Yat-sen University)
CodeDrug DiscoveryProtein Structure PredictionGraph Neural NetworkSupervised Fine-TuningBiomedical Data
π― What it does: H3-DDG is proposed, a hierarchical multi-body attention network based on hypergraphs, for accurate prediction of protein interaction ΞΞG.
AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play
Ran Xu (Emory University), Carl Yang (Emory University)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes AceSearcher, a retrieval-augmented model that simultaneously performs question decomposition and solving through cooperative self-play training of a single LLM.
CodeComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: Designed and implemented GRESO, an online lightweight method for filtering non-informative prompts in RL training, significantly reducing roll-out computation.
Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models
Zekai Zhao (University of California San Diego), Biwei Huang (University of California San Diego)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: By amplifying activation values and inserting the 'wait' trigger word, the long chain reasoning (Long-CoT) capability of large language models is activated without any training;
Activation-Guided Consensus Merging for Large Language Models
Yuxuan Yao (City University of Hong Kong), Linqi Song (City University of Hong Kong)
CodeOptimizationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The Activation-Guided Consensus Merging (ACM) method is proposed, which adaptively determines hierarchical weight coefficients by calculating the mutual information of activations at different layers, achieving a gradient-free and additional training-free merging of pre-trained models and multi-task fine-tuned models.
Activation-Informed Merging of Large Language Models
Amin Heyrani Nobari (Massachusetts Institute of Technology), Navid Azizan (Massachusetts Institute of Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
π― What it does: Proposes the Activation-Informed Merging (AIM) method, which uses activation space information to assist in merging multiple specialized fine-tuned LLMs based on the same pre-trained model, achieving a more robust model merging.
π― What it does: This paper proposes an active measurement framework for human-computer interaction, which obtains unbiased and variance-estimable scientific measurement results through AI model prediction, importance sampling, and incremental label updates.
Heeju Ko (Korea University), Sangpil Kim (Korea University)
CodeDomain AdaptationRobotic IntelligenceTransformerReinforcement LearningVision Language ModelMultimodality
π― What it does: ATENA proposes an online active testing adaptation framework that enhances the robustness of visual language navigation agents in unknown environments by utilizing human turn feedback and self-supervision.
π― What it does: A method for activity pruning based on adaptive thresholds is designed, utilizing AT-LIF neurons to dynamically adjust thresholds during training and output 0/ΞΈ signals, significantly reducing the firing rate of SNNs and computational load.
Actor-Free Continuous Control via Structurally Maximizable Q-Functions
Yigit Korkmaz (University of Southern California), Erdem Biyik
CodeReinforcement LearningSequential
π― What it does: A continuous control reinforcement learning algorithm without an actor, Q3C, is proposed, which utilizes learnable control points to achieve a structure that maximizes the Q function;
AcuRank: Uncertainty-Aware Adaptive Computation for Listwise Reranking
Soyoung Yoon (Seoul National University), seung-won hwang
CodeRetrievalRecommendation SystemComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: This paper presents AcuRank, an uncertainty-driven adaptive computing framework for dynamically selecting documents that require further reasoning in list-based re-ranking assisted by large language models (LLMs), and iteratively updating their relevance estimates.
Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference
Yuan Feng (University of Science and Technology of China), S Kevin Zhou
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper studies and proposes Ada-KV, which allocates budgets adaptively based on attention heads to improve the quality of KV cache eviction and enhance the efficiency of LLM long sequence inference.
Ada-R1: Hybrid-CoT via Bi-Level Adaptive Reasoning Optimization
Haotian Luo (Shenzhen Campus of Sun Yat-sen University), Li Shen (Shenzhen Campus of Sun Yat-sen University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
π― What it does: A dual-stage adaptive reasoning framework called Ada-R1 is proposed, allowing large language models to dynamically select between long-chain or short-chain reasoning based on input difficulty, significantly reducing reasoning length and cost while maintaining accuracy.
AdaDetectGPT: Adaptive Detection of LLM-Generated Text with Statistical Guarantees
Hongyi Zhou (Tsinghua University), Chengchun Shi (London School of Economics)
CodeClassificationAnomaly DetectionTransformerLarge Language ModelText
π― What it does: An adaptive LLM text detector called AdaDetectGPT is designed, which learns and applies the witness function to enhance the discriminative ability of existing logits-based detectors, providing finite sample guarantees for statistical metrics such as FNR, TNR, FPR, and TPR.
CodeExplainability and InterpretabilityConvolutional Neural NetworkImage
π― What it does: This paper proposes an adaptive noise sampling-based gradient smoothing method called AdaptGrad, aimed at reducing noise in gradient interpretation methods and improving visualization quality.
Adaptive Batch-Wise Sample Scheduling for Direct Preference Optimization
Zixuan Huang (Beihang University), deqing wang
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: The 'sample scheduling' problem is proposed, and the SamS algorithm is designed to dynamically and adaptively select samples from each batch during Direct Preference Optimization (DPO) training to enhance the human preference alignment effect of large language models (LLMs).
Adaptive Classifier-Free Guidance via Dynamic Low-Confidence Masking
Pengxiang Li (Polytechnic University), Xiaowei Gao (Peking University)
CodeGenerationDiffusion modelText
π― What it does: Proposes Adaptive Classifier-Free Guidance (A-CFG), which improves the unsupervised guidance of iterative masked diffusion language models by dynamically identifying low-confidence positions in the model and temporarily re-masking them.
π― What it does: A multi-agent reinforcement learning framework based on a central agent that adaptively optimizes context length and combines Fourier low-frequency truncation to extract global temporal trends (ACL-LFT) is proposed.
Adaptive Defense against Harmful Fine-Tuning for Large Language Models via Bayesian Data Scheduler
Zixuan Hu (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This study investigates adaptive defenses against harmful fine-tuning in large language models, proposing the Bayesian Data Scheduler (BDS) framework, which can safely weight training data without the need for attack simulations, thereby enhancing the model's safety and performance.
Adaptive Distraction: Probing LLM Contextual Robustness with Automated Tree Search
Yanbo Wang (Mohamed bin Zayed University of Artificial Intelligence), Xiangliang Zhang (University of Notre Dame)
CodeOptimizationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: An adaptive interference generation framework based on tree search is proposed, which automatically generates semantically coherent and task-independent contextual interference to test the contextual robustness of LLMs.
π― What it does: Proposes Adaptive Fission, a post-training multi-neuron group encoding method that can convert ANNs to SNNs while significantly reducing latency and energy consumption.
Adaptive Frontier Exploration on Graphs with Applications to Network-Based Disease Testing
Davin Choo (Harvard University), Cheryl Johnson (World Health Organization)
CodeOptimizationGraph Neural NetworkReinforcement LearningGraphBiomedical Data
π― What it does: An adaptive decision framework AFEG suitable for frontier exploration on graphs is proposed, and a Gittins index-based strategy GITTINS is designed, proving its optimality on tree-structured graphs.
Adaptive Kernel Design for Bayesian Optimization Is a Piece of CAKE with LLMs
Richard Cornelius Suwandi (Chinese University of Hong Kong), Sergios Theodoridis (University of Athens)
CodeOptimizationHyperparameter SearchLarge Language ModelPrompt EngineeringTabular
π― What it does: Proposes a method for adaptive Gaussian process kernel evolution based on large language models (CAKE), which generates and optimizes kernel structures in real-time during the Bayesian optimization process.
Adaptive Latent-Space Constraints in Personalized Federated Learning
Sana Ayromlou (Vector Institute), D. B. Emerson (Vector Institute)
CodeFederated LearningImage
π― What it does: In federated learning, adaptive MMD metrics (MK-MMD and MMD-D) are introduced into personalized FL frameworks such as Ditto and MR-MTL to constrain the potential space distribution differences between local models and global models, thereby enhancing model performance.
Adaptive Neighborhood-Constrained Q Learning for Offline Reinforcement Learning
Yixiu Mao (Tsinghua University), Xiangyang Ji (Tsinghua University)
CodeReinforcement LearningTabular
π― What it does: An offline reinforcement learning method based on adaptive neighborhood constraints, ANQ, is proposed, utilizing bi-level optimization to implement Q-learning and avoid OOD errors.
Adaptive Prediction-Powered AutoEval with Reliability and Efficiency Guarantees
Sangwoo Park (King's College London), Osvaldo Simeone (King's College London)
CodeLarge Language ModelPrompt EngineeringText
π― What it does: A self-assessment method R-AutoEval+ is proposed, which can provide reliable performance estimates with limited samples and automatically reduce reliance on synthetic data when needed, ensuring improved sample efficiency without compromising reliability.
Adaptive Preference Arithmetic: A Personalized Agent with Adaptive Preference Arithmetic for Dynamic Preference Modeling
Hongyi Nie (Northwestern Polytechnical University), Zhen Wang (Northwestern Polytechnical University)
CodeRecommendation SystemTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: This paper proposes the AdaPA-Agent framework for dynamically modeling user preference intensity in LLM agents, combining alignment estimation and preference arithmetic to achieve personalized generation.
π― What it does: An adaptive recalibration learning framework (ARL) for multimodal intent recognition is proposed, which alleviates the modality imbalance problem by dynamically recalibrating modality importance.
AdaSPEC: Selective Knowledge Distillation for Efficient Speculative Decoders
Yuezhou Hu (University of California), Tuo Zhao (Georgia Institute of Technology)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes AdaSPEC, a selective knowledge distillation method under the Speculative Decoding framework, aimed at training small draft models and enhancing their alignment with the target model.
AdaVideoRAG: Omni-Contextual Adaptive Retrieval-Augmented Efficient Long Video Understanding
Zhucun Xue (Zhejiang University), Dacheng Tao (Nanyang Technological University)
CodeRetrievalComputational EfficiencyLarge Language ModelVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes the AdaVideoRAG framework, which utilizes query intent classification to dynamically select different retrieval strategies (no retrieval, naive retrieval, graph retrieval) for efficient deep understanding of multimodal question answering in long videos.
Addressing Mark Imbalance in Integration-free Marked Temporal Point Processes
Sishun Liu (RMIT University), Xiuzhen Zhang (RMIT University)
CodeRecurrent Neural NetworkTime SeriesSequential
π― What it does: This study investigates the impact of label imbalance on the prediction of marked temporal point processes (MTPP), proposes a method for adjusting label probabilities based on thresholds, and develops an integral-free neural MTPP (IFNMTPP) that implements a strategy of predicting labels first and then predicting time.
ADG: Ambient Diffusion-Guided Dataset Recovery for Corruption-Robust Offline Reinforcement Learning
Zeyuan Liu (Tsinghua University), Xiu Li (Tsinghua University)
CodeReinforcement LearningDiffusion modelTabular
π― What it does: This study proposes the Ambient Diffusion-Guided Dataset Recovery (ADG) method, which utilizes diffusion models to detect, recover, and denoise trajectories with noise or damage in offline reinforcement learning, thereby enhancing the robustness of policy learning.