These 2207 ICLR 2026 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 ICLR 2026 paper, free trial on arXivSub.
``Noisier'β Noise Contrastive Estimation is (Almost) Maximum Likelihood
Peiyu Yu (University of California, Los Angeles), Ying Nian Wu (University of California, Los Angeles)
π― What it does: The paper proposes an improved noise contrastive estimation method called 'Noisier NCE,' which amplifies the scale of the noise distribution to make its gradient approach maximum likelihood estimation, thereby addressing the convergence difficulties of traditional NCE when there is a large distribution gap;
<SO$G_k$>: One LLM Token for Explicit Graph Structural Understanding
Jingyao Wu (Shanghai Jiao Tong University), Chenghu Zhou (IGSNRR, Chinese Academy of Sciences)
CodeClassificationRepresentation LearningDrug DiscoveryGraph Neural NetworkLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningTextGraph
π― What it does: This paper proposes using a special structural marker <SO G k> to directly map graph structures into the LLM's vocabulary, achieving precise and concise representation of graph structures, and aligning markers through hybrid structural QA.
$\alpha$-DPO: Robust Preference Alignment for Diffusion Models via $\alpha$ Divergence
Yang Li (New Laboratory of Pattern Recognition, MAIS, CASIA), Jing Dong (New Laboratory of Pattern Recognition, MAIS, CASIA)
CodeGenerationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageTextMultimodality
π― What it does: This paper proposes Ξ±-DPO, a direct preference optimization method based on Ξ±-divergence, for robustly aligning diffusion models with human preferences under noisy labels.
π― What it does: Propose a continuous-time graph representation model based on β1 distance, named β1 LD-CTGR, which replaces the squared β2 distance with β1 implicit space distance to address the violation of triangle inequality, and provides a closed-form piecewise exponential integral for the hazard function.
$\mathbf{Li_2}$: A Framework on Dynamics of Feature Emergence and Delayed Generalization
Yuandong Tian (Meta Superintelligence Labs)
CodeOptimizationRepresentation Learning
π― What it does: Studied the gradient dynamics of the 'grokking' phenomenon in two-layer networks and proposed the LiΒ² framework, dividing the learning process into three stages: lazy learning, independent feature learning, and interactive feature learning, elucidating how features emerge from gradient signals and lead to sudden generalization.
π― What it does: This paper investigates the generalization ability of learning optimizers (LO) when facing widths, depths, and training times that exceed the meta-training range, and proposes migrating LO to maximum update parameterization (ΞΈ P), significantly enhancing its meta-generalization performance.
$\nabla$-Reasoner: LLM Reasoning via Test-Time Gradient Descent in Latent Space
Peihao Wang (University of Texas at Austin), Zhangyang Wang (University of Texas at Austin)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningText
π― What it does: Propose a framework called β-Reasoner, which optimizes the generation strategy during inference by performing gradient descent on the token logits of the LLM output, with the core being Differentiable Textual Optimization (DTO);
π― What it does: Propose a visual geometry reconstruction model Ο3 that does not require a reference view, utilizing a fully permutation-equivariant network to predict camera poses and local point clouds.
$AutoDrive\text{-}P^3$: Unified Chain of PerceptionβPredictionβPlanning Thought via Reinforcement Fine-Tuning
Yuqi Ye (Peking University), Wei Gao (Peking University)
CodeAutonomous DrivingSupervised Fine-TuningReinforcement LearningVision Language ModelChain-of-Thought
π― What it does: Proposing the AutoDrive P-3 framework to achieve end-to-end driving decision-making through chain-of-thought reasoning that integrates perception, prediction, and planning.
π― What it does: This work proposes the 3DCS benchmark, systematically evaluating the sensitivity of molecular representations to conformational changes within molecules, including geometry, chirality, and energy.
π― What it does: Propose a brain benchmark model called BrainGFM based on graph neural networks, which is pre-trained on a large-scale multi-modal dataset consisting of 27 fMRI datasets, 25 common brain diseases, and 8 brain parcellation templates. Subsequently, few-shot and zero-shot transfer is achieved through graph prompts and language prompts.
π― What it does: Proposed a novel fMRI-to-video decoding framework called VCFLOW, which does not require additional training for new subjects. It utilizes a dual-stream structure of the brain's visual cortex (early, dorsal, and ventral streams) to extract multi-level cognitive features and align and reconstruct them.
A Comprehensive Information-Decomposition Analysis of Large Vision-Language Models
Lixin Xiu (University of Tokyo), Hideki Nakayama (University of Tokyo)
CodeExplainability and InterpretabilityRepresentation LearningMultimodalityBenchmark
π― What it does: This paper constructs an unsupervised evaluation framework based on Partial Information Decomposition (PID), performing information spectrum decomposition (redundant, visual-unique, language-unique, synergistic information) on the decision-making processes of 26 large vision-language models (LVLM) across four multiple-choice VQA benchmarks, and systematically analyzes the information flow from three perspectives: model hierarchy, task dimensions, and training stages.
A Dense Subset Index for Collective Query Coverage
Kartik Nair (Carnegie Mellon University), Abir De (IIT Bombay)
CodeRetrievalTextTabularBenchmark
π― What it does: This paper proposes a new 'Dense Subset Index' (DISCO), aiming to achieve multi-content collaborative retrieval through set cover approaches, solving the problem that traditional single-document retrieval cannot cover scenarios such as multi-step reasoning and table retrieval.
A Derandomization Framework for Structure Discovery: Applications in Neural Networks and Beyond
Nikos Tsikouras (National and Kapodistrian University of Athens), Christos Tzamos (National and Kapodistrian University of Athens)
CodeOptimizationRepresentation LearningGraph
π― What it does: Proposed a structure discovery framework based on derandomization, proving that under weak assumptions neural networks can automatically learn low-rank first-layer weights via optimization methods such as gradient descent, thereby achieving structural discovery.
A Fair Bayesian Inference through Matched Gibbs Posterior
Jihu Lee (Seoul National University), Yongdai Kim (Seoul National University)
CodeClassificationSafty and PrivacyComputational EfficiencyImageTextTabular
π― What it does: Propose matching Gibbs posterior as a proxy distribution to achieve group fair Bayesian inference, addressing the challenge of posterior computation under demographic balance constraints in deep models.
A Federated Generalized Expectation-Maximization Algorithm for Mixture Models with an Unknown Number of Components
Michael Ibrahim (Georgia Institute of Technology), Weijun Xie (Georgia Institute of Technology)
CodeOptimizationFederated LearningImageTabular
π― What it does: Propose FedGEM: a federated generalized expectation maximization (GEM) algorithm that can train a mixture model without knowing the global number of clusters and when the local cluster sets across clients are inconsistent, utilizing uncertainty sets to detect cross-client cluster overlaps and aggregating them on the server side;
A Function-Centric Graph Neural Network Approach for Predicting Electron Densities
Manuel V. Klockow (Heidelberg University), Fred A. Hamprecht (Heidelberg University)
CodeGraph Neural NetworkGraphPhysics Related
π― What it does: Built an equivariant graph neural network based on the quadratic expansion of the ground-state electron density to predict the electron density under molecular geometry.
A General Spatio-Temporal Backbone with Scalable Contextual Pattern Bank for Urban Continual Forecasting
Aoyu Liu (Tongji University), Yaying Zhang (Tongji University)
CodeGraph Neural NetworkPrompt EngineeringTime Series
π― What it does: Proposed a continuous spatiotemporal prediction framework named STBP, capable of performing efficient prediction in real-world streaming data environments where nodes continuously expand and graph structures and distributions experience persistent drift;
A Genetic Algorithm for Navigating Synthesizable Molecular Spaces
Alston Lo (Massachusetts Institute of Technology), Wojciech Matusik (Massachusetts Institute of Technology)
CodeDrug DiscoveryBiomedical Data
π― What it does: Propose SynGA, a genetic algorithm that directly operates on synthetic routes (synthesis trees), further enhancing search efficiency with a lightweight machine learning filter (MLP or neural additive model), and finally embedding it into Bayesian optimization to obtain SynGBO, achieving the design of synthetically feasible molecules and property optimization.
A Graph Meta-Network for Learning on KolmogorovβArnold Networks
Guy Bar-Shalom (Technion), Haggai Maron (Technion)
CodeClassificationGraph Neural NetworkImage
π― What it does: Proposes WS-KAN, a weight-space graph primitive network for Kolmogorov-Arnold networks (KAN), designed to directly learn and infer KAN parameters.
A Hierarchical Circuit Symbolic Discovery Framework for Efficient Logic Optimization
Yinqi Bai (University of Science and Technology of China), Jianye HAO
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerReinforcement LearningGraph
π― What it does: This paper proposes a hierarchical circuit symbolic discovery framework (HIS), which identifies invalid subgraphs in logic optimization by learning lightweight, interpretable symbolic functions, thereby achieving efficient pruning.
CodeGenerationExplainability and InterpretabilityDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
π― What it does: Propose the Variational Hard Concept Bottleneck (VHCB) layer, which maps internal representations to hard concepts using a binary VAE, achieving interpretable and controllable generation.
A Representer Theorem for Hawkes Processes via Penalized Least Squares Minimization
Hideaki Kim (NTT, Inc.), Tomoharu Iwata (NTT, Inc.)
CodeOptimizationTime SeriesSequentialFinance Related
π― What it does: A penalized least squares estimation method is proposed in RKHS for the triggering kernel function of multivariate Hawkes processes, along with a corresponding representer theorem. It is proven that the optimal solution can be expressed as a linear combination of equivalent kernel functions, with all dual coefficients equal to 1, leading to a non-iterative closed-form solution.
A Resolution-Agnostic Geometric Transformer for Chromosome Modeling Using Inertial Frame
Yize Zhou (Wave Intelligence Lab), Shengchao Liu (Wave Intelligence Lab)
CodeTransformerBiomedical Data
π― What it does: Propose an InertialGenome framework based on Transformer for 3D chromosome reconstruction, which utilizes an inertial frame to normalize the initial coordinates' pose and incorporates geometric-aware position encoding in the Transformer to achieve high-precision 3D chromosomal structure reconstruction across resolutions.
π― What it does: This paper conducts theoretical research on a semi-dual adversarial neural optimal transport solver from a statistical learning perspective, bridging the gap in theoretical analysis of existing methods.
A Statistical Theory of Overfitting for Imbalanced Classification
Jingyang Lyu (University of Wisconsin-Madison), Yiqiao Zhong (University of Wisconsin-Madison)
CodeClassificationImageTextBiomedical Data
π― What it does: Studied the problem of high-dimensional imbalanced classification, constructed the statistical theory of linear classifiers, and revealed the truncation effect of the training set on the numerical (logit) distribution;
π― What it does: This paper proposes decoupling and restructuring the optimization process of 3D Gaussian Splatting (3DGS), designing Sparse Adam, Re-State Regularization (RSR), and Decoupled Attribute Regularization (DAR), and integrating them into a new optimizer called AdamW-GS to improve optimization efficiency, controllable regularization, and automatic redundancy removal.
A Stitch in Time Saves Nine: Proactive Self-Refinement for Language Models
Jinyi Han (East China Normal University), Yanghua Xiao (East China Normal University)
CodeOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: Developed a proactive self-refinement method (PASR) that enables large language models to actively decide whether, when, and how to improve their own outputs during the generation process;
A Structured, Tagged, and Localized Visual Question Answering Dataset with Full Sentence Answers and Scene Graphs for Chest X-ray Images
Philip MΓΌller (Technical University of Munich), Daniel Rueckert (Technical University of Munich)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBiomedical DataElectronic Health RecordsBenchmark
π― What it does: Automatically constructed a chest X-ray visual question answering dataset named CXRβQBA containing 42 million QA pairs, featuring complete sentence answers, precise bounding boxes, and multi-level labels.
CodeRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: Proposes an untrained long video understanding framework that integrates adaptive frame sampling, dynamic resolution allocation, and video-query-option similarity calculation.
A Two-Phase Deep Learning Framework for Adaptive Time-Stepping in High-Speed Flow Modeling
Jacob Helwig (Texas A&M University), Shuiwang Ji (Texas A&M University)
CodeConvolutional Neural NetworkGraph Neural NetworkMixture of ExpertsMeshPhysics Related
π― What it does: Propose the ShockCast two-stage framework: the first stage uses a neural network to predict adaptive time steps, and the second stage drives the flow field evolution with these steps.
A Unification of Discrete, Gaussian, and Simplicial Diffusion
Nuria Alina Chandra (New York University), Andrew Gordon Wilson (New York University)
CodeGenerationData SynthesisDiffusion modelImageTextBiomedical Data
π― What it does: Unify discrete, Gaussian, and simplex diffusion models under the Wright-Fisher evolutionary model framework, and propose a sufficient statistic parameterization, enabling the same network to be simultaneously trained and inferred across three domains.
A universal compression theory for lottery ticket hypothesis and neural scaling laws
Hong-Yi Wang (Princeton University), Liu Ziyin (MIT)
CodeCompressionComputational EfficiencyTabular
π― What it does: This paper proposes and proves a general compression theorem, demonstrating that any smooth function invariant to object permutations can be compressed to a polynomial logarithmic number of objects with negligible error; based on this, it proves the dynamic lottery ticket hypothesis and shows that compression significantly enhances the scaling laws of neural networks and datasets.
A-TPT: Angular Diversity Calibration Properties for Test-Time Prompt Tuning of Vision-Language Models
Shihab Aaqil Ahamed (University of Moratuwa), Muhammad Haris Khan (Mohamed bin Zayed University of Artificial Intelligence)
CodeClassificationRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBiomedical Data
π― What it does: Propose A-TPT in test-time prompt tuning by maximizing the minimal angular distance of text features to enhance the calibration performance of vision-language models (VLM);
A$^2$Search: Ambiguity-Aware Question Answering with Reinforcement Learning
Fengji Zhang (City University of Hong Kong), Junyang Lin (Alibaba Group)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
π― 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.
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)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― 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.
π― 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.
π― 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.
CodeComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
π― 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).
Accessible, Realistic, and Fair Evaluation of Positive-Unlabeled Learning Algorithms
Wei Wang (RIKEN), Masashi Sugiyama (RIKEN)
CodeClassificationImageTabularBenchmark
π― 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)
CodeGenerationSupervised Fine-TuningVision Language ModelDiffusion modelImageText
π― 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).
π― 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.
Action-Guided Attention for Video Action Anticipation
Tsung-Ming Tai (Free University of Bozen-Bolzano), Oswald Lanz (NVIDIA)
CodeRecognitionExplainability and InterpretabilityRepresentation LearningTransformerVideo
π― What it does: Propose an attention mechanism based on past action predictionβAction-Guided Attentionβwhich combines adaptive gating to dynamically fuse historical and current visual features for video action anticipation.
Activation Function Design Sustains Plasticity in Continual Learning
Lute Lillo (University of Vermont), Nick Cheney (University of Vermont)
CodeReinforcement LearningImageBenchmark
π― 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.
CodeSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelDiffusion modelImageText
π― 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)
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationLarge Language ModelAuto EncoderTextChain-of-Thought
π― 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.
π― 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.
AdaBlock-dLLM: Semantic-Aware Diffusion LLM Inference via Adaptive Block Size
Guanxi Lu (Imperial College London), Hongxiang Fan (Imperial College London)
CodeComputational EfficiencyLarge Language ModelDiffusion modelText
π― What it does: Propose AdaBlock-dLLM, a scheduler that dynamically adjusts block size during the semi-autoregressive diffusion LLM inference process;
AdAEM: An Adaptively and Automated Extensible Measurement of LLMs' Value Difference
Jing Yao (Renmin University of China), Xing Xie (Microsoft Research Asia)
CodeTransformerLarge Language ModelTextBenchmark
π― 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)
CodeDomain AdaptationOptimizationHyperparameter SearchTransformerTime Series
π― 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.
π― 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.
CodeOptimizationHyperparameter SearchTransformerLarge Language ModelPrompt EngineeringTabular
π― 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 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)
CodeOptimizationMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
π― 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)
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
π― 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;
π― 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 Debiasing Tsallis Entropy for Test-Time Adaptation
Xiangyu Wu (Nanjing University of Science and Technology), Jianfeng Lu (Nanjing University of Science and Technology)
CodeClassificationDomain AdaptationVision Language ModelImage
π― 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.
π― 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)
CodeOptimizationTabular
π― 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 Methods Are Preferable in High Privacy Settings: An SDE Perspective
Enea Monzio Compagnoni (University of Basel), Anastasia Koloskova
CodeOptimizationSafty and PrivacyTextStochastic Differential Equation
π― 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.
CodeDomain AdaptationRepresentation LearningGraph Neural NetworkMixture of ExpertsGraph
π― What it does: Propose AdaMixβa 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
CodeRestorationGenerationDiffusion modelImage
π― 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)
CodeCompressionTransformerImageText
π― 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)
CodeRecommendation SystemTabular
π― 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.
π― 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 Social Learning via Mode Policy Optimization for Language Agents
Minzheng Wang (University of Chinese Academy of Sciences), Wenji Mao (Chinese Academy of Sciences)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextMultimodalityBenchmarkChain-of-Thought
π― 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)
CodeClassificationDomain AdaptationAuto EncoderBiomedical DataElectronic Health Records
π― 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
AdaReasoner: Dynamic Tool Orchestration for Iterative Visual Reasoning
Mingyang Song (Fudan University), Yu Cheng (Chinese University of Hong Kong)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelImageTextMultimodality
π― What it does: Constructed the AdaReasoner framework, enabling multi-modal large language models to achieve iterative visual reasoning through dynamic tool orchestration.
π― What it does: Study the node distinguishability of spectral graph neural networks and propose the AdaSpec adaptive spectral matrix module to enhance node distinguishability.
Addressing divergent representations from causal interventions on neural networks
Satchel Grant (Stanford University), Christopher Potts (Stanford University)
CodeExplainability and InterpretabilityRepresentation LearningTransformerText
π― 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.
CodeDomain AdaptationTransformerLarge Language ModelTextBiomedical DataBenchmark
π― 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.
π― 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.
π― 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)
CodeTransformerGraphPhysics Related
π― 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.
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)
CodeAdversarial AttackVision Language ModelMultimodalityBenchmark
π― 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.
π― 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.
π― 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.
AFTER: Mitigating the Object Hallucination of LVLM via Adaptive Factual-Guided Activation Editing
Tianbo Wang (Beihang University), Xianglong Liu (Beihang University)
CodeExplainability and InterpretabilityLarge Language ModelVision Language ModelMultimodality
π― What it does: The Adaptive Factual Guidance for Visual-Text Editing (AFTER) method edits the internal activations of large vision-language models (LVLMs) using factual text to alleviate object hallucinations caused by language bias.
Agent-X: Evaluating Deep Multimodal Reasoning in Vision-Centric Agentic Tasks
Tajamul Ashraf (Mohamed bin Zayed University of Artificial Intelligence), Salman Khan (Mohamed bin Zayed University of Artificial Intelligence)
CodeLarge Language ModelAgentic AIVision Language ModelImageVideoTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Proposed the Agent-X benchmark to evaluate the ability of vision-centric agents to perform multi-step deep reasoning and tool calling in real multi-modal environments.
AgentGym-RL: An Open-Source Framework to Train LLM Agents for Long-Horizon Decision Making via Multi-Turn RL
Zhiheng Xi (Fudan University), Yu-Gang Jiang (Fudan University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AITextChain-of-Thought
π― What it does: This paper develops a unified reinforcement learning framework AgentGym-RL for training large language model (LLM) agents to complete multi-round decision tasks, and proposes a phased interaction expansion method called ScalingInter-RL;
Agentic Jigsaw Interaction Learning for Enhancing Visual Perception and Reasoning in Vision-Language Models
Yu Zeng (University of Science and TechnolZengogy of China), Feng Zhao (University of Science and TechnolZengogy of China)
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision-Language-Action ModelImageTextMultimodality
π― What it does: Propose the AGILE framework, treating the puzzle-solving process as an interactive multi-round dialogue with the environment, leveraging VLM to generate executable Python code to perform actions such as Swap, Observe, Crop, Zoom, progressively enhancing visual perception and reasoning capabilities.
AgentPO: Enhancing Multi-Agent Collaboration via Reinforcement Learning
Lin Sun (MatrixRobotics), Ning Wu (UAES AI Lab)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmark
π― What it does: Propose the AgentPO framework, which trains lightweight collaborators (Collaborator) using reinforcement learning within a fixed multi-agent topology to optimize collaboration methods, thereby enhancing the overall reasoning performance of the primary executor (Actor).
AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems?
Guibin Zhang, Shuicheng YAN
CodeData SynthesisAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AITextSequential
π― What it does: Proposes the AgenTracer framework, which automatically generates annotated multi-agent failure trajectories and trains a lightweight failure localization model, AgenTracer-8B, to help quickly identify root causes in LLM agent systems.
π― What it does: Introduces a post-training structured pruning method called AIRE-Prune, which performs layer-adaptive pruning on deep state-space models based on the infinite impulse response energy of each state.
Align to Misalign: Automatic LLM Jailbreak with Meta-Optimized LLM Judges
Hamin Koo (Yonsei University), Jaehyung Kim (Yonsei University)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Through the meta-optimization framework AMIS, simultaneously evolve jailbreak prompts and scoring templates, significantly improving the jailbreak success rate of large language models.
Aligned Agents, Biased Swarm: Measuring Bias Amplification in Multi-Agent Systems
Keyu Li (Shanghai Jiao Tong University), Dequan Wang (Shanghai Jiao Tong University)
CodeLarge Language ModelTextBenchmark
π― What it does: Systematically study the amplification of bias in multi-agent systems (MAS), constructing and utilizing the Discrim-Eval-Open benchmark to conduct comprehensive evaluations across different roles, topologies, depths, and model heterogeneity.
AlignFlow: Improving Flow-based Generative Models with Semi-Discrete Optimal Transport
Lingkai Kong (Georgia Institute of Technology), Huidong Liu (Amazon.com Inc)
CodeGenerationFlow-based ModelImage
π― What it does: Propose the AlignFlow method, which achieves explicit alignment between noise and data through semi-discrete optimal transport (SDOT), improving the training of flow-based generative models.
Aligning Collaborative View Recovery and Tensorial Subspace Learning via Latent Representation for Incomplete Multi-View Clustering
Youqing Wang (Beijing University of Chemical Technology), Jipeng Guo (Beijing University of Chemical Technology)
CodeOptimizationRepresentation LearningBenchmark
π― What it does: Propose a missing multi-view clustering method (ARSL-IMVC) that unifies view recovery with tensor subspace learning through shared latent representations
Alignment through Meta-Weighted Online Sampling: Bridging the Gap between Data Generation and Preference Optimization
Junming Yang (Southeast University), Xin Geng (Southeast University)
CodeOptimizationComputational EfficiencyData-Centric LearningMeta LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningTextBenchmark
π― What it does: Propose the MetaAPO framework, which uses a lightweight meta-learner to adaptively assess the alignment gap between offline preference data and online-generated samples during training, dynamically performing online sampling and weighted optimization to enhance LLM alignment efficiency.
All Code, No Thought: Language Models Struggle to Reason in Ciphered Language
Shiyuan Guo (Anthropic Fellows Program), Fabien Roger (Anthropic)
CodeExplainability and InterpretabilityTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
π― What it does: Investigate the reasoning ability of large language models on encrypted text and assess its potential threat to Chain-of-Thought monitoring.
All That Glitters Is Not Gold: Key-Secured 3D Secrets within 3D Gaussian Splatting
Yan Ren (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)
CodeSafty and PrivacyVision Language ModelGaussian SplattingPoint CloudMesh
π― What it does: Propose the KeySS framework to achieve end-to-end key-locked 3D steganography, jointly optimizing the 3D Gaussian points in the cover scene and the key decoder, maintaining the standard 3D Gaussian Splatting format, supporting multi-secret hiding and anti-erroneous key attacks.
ALM-MTA: Front-Door Causal Multi-Touch Attribution Method for Creator-Ecosystem Optimization
Yuguang Liu (Beijing Dajia Internet Information Technology Co., Ltd), Kun Gai (Beijing Dajia Internet Information Technology Co., Ltd)
CodeRecommendation SystemOptimizationExplainability and InterpretabilityContrastive LearningTabularSequential
π― What it does: Developed a causal multi-touch attribution framework ALM-MTA based on front-door identification, addressing the challenge of reliable attribution in recommendation systems caused by unlabelled data, unobserved confounders, and large-scale touchpoint spaces.
AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning
Yi Zhang (University Of Science And Technology Of China), Xiang Wang
CodeSafty and PrivacyLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: Proposes AlphaAlign, a pure reinforcement learning framework that activates the intrinsic safety awareness of LLMs through verifiable safety rewards, supplemented by normalized helpfulness rewards to achieve a balance between safety and practicality;
π― What it does: By deeply analyzing the training objective of MeanFlow and gradient conflicts, this paper proposes the Ξ±-Flow framework, achieving progressive training from trajectory flow matching to MeanFlow, thereby improving the quality of few-step generative models trained from scratch.