ICLR 2026 Papers — Page 23
International Conference on Learning Representations · 5356 papers
ICaRus: Identical Cache Reuse for Efficient Multi-Model Inference
Sunghyeon Woo (NAVER Cloud), Dongsoo Lee (NAVER Cloud)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the ICaRus architecture, which allows multi-task dedicated models to share the same KV cache, significantly reducing memory consumption and recomputation costs in multi-model inference.
ICDiffAD: Implicit Conditioning Diffusion Model for Time Series Anomaly Detection
Fan Zhang (Tsinghua University), Wenming Yang (Tsinghua University)
Anomaly DetectionDiffusion modelScore-based ModelTime Series
🎯 What it does: An implicit conditional diffusion model called ICDiffAD for time series anomaly detection was studied, addressing the false alarm issues caused by reconstruction inconsistency in traditional diffusion models.
Ice Cream Doesn’t Cause Drowning: Benchmarking LLMs Against Statistical Pitfalls in Causal Inference
Jin Du (University of Minnesota), Jie Ding (University of Minnesota)
TransformerLarge Language ModelPrompt EngineeringTabularBenchmark
🎯 What it does: Constructed the CausalPitfalls benchmark to evaluate the reliability of LLMs in statistical causal inference, covering six causal pitfalls and fifteen challenges.
IceCache: Memory-Efficient KV-cache Management for Long-Sequence LLMs
Yuzhen Mao (Simon Fraser University), Ke Li (Simon Fraser University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose an IceCache KV-cache strategy based on semantic clustering and page management for memory optimization during long-sequence LLM inference;
ICYM$^2$I: The illusion of multimodal informativeness under missingness
Young Sang Choi (Columbia University), Shalmali Joshi (Columbia University)
Data-Centric LearningImageTextMultimodalityBiomedical DataComputed TomographyElectrocardiogram
🎯 What it does: Investigates the bias in information assessment caused by missing modes in multimodal learning and proposes the ICYM 2 I framework to correct training and evaluation.
IDEAL: Data Equilibrium Adaptation for Multi-Capability Language Model Alignment
Chenlin Ming (Shanghai Jiao Tong University), Conghui He (Shanghai Artificial Intelligence Laboratory)
Domain AdaptationOptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the IDEAL framework, which automatically balances data and enhances the performance of large language models (LLMs) across multiple capabilities by iteratively adjusting the proportions of multi-domain SFT training data.
Identifiability Challenges in Sparse Linear Ordinary Differential Equations
Cecilia Casolo (Technical University of Munich), Niki Kilbertus (Munich Center for Machine Learning)
Physics RelatedOrdinary Differential Equation
🎯 What it does: This paper studies the identifiability problem of sparse linear ordinary differential equations (ODEs), particularly in data-driven dynamic system modeling, exploring the differences between identifiability of sparse systems and dense systems.
Identifying and Evaluating Inactive Heads in Pretrained LLMs
Pedro Sandoval-Segura (University of Maryland), David Jacobs (University of Maryland)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigate and evaluate the presence and extent of passive (inactive) attention heads in large language models
Identifying Robust Neural Pathways: Few-Shot Adversarial Mask Tuning for Vision-Language Models
Wonjeong Choi (Korea Advanced Institute of Science and Technology), Jaekyun Moon (Korea Advanced Institute of Science and Technology)
ClassificationAdversarial AttackSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes AdvMask, a few-shot adversarial robustness enhancement framework for vision-language models (VLMs), which screens robust subnetworks by applying binary masks on frozen pre-trained weights;
Identity-Free Deferral For Unseen Experts
Joshua Strong (University of Oxford), Alison Noble
ClassificationMixture of ExpertsImageBiomedical Data
🎯 What it does: Propose an identity-free deferral framework called Identity-Free Deferral (IFD) that achieves adaptive deferral for unseen experts.
IDER: IDempotent Experience Replay for Reliable Continual Learning
Zhanwang Liu (Shanghai Jiao Tong University), Weiran Huang (Shanghai Jiao Tong University)
Computational EfficiencyKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Proposes the Idempotent Experience Replay (IDER) method, leveraging function idempotency to suppress catastrophic forgetting and enhance prediction reliability in continual learning;
IF-VidCap: Can Video Caption Models Follow Instructions?
Shihao Li (Nanjing University), Jiaheng Liu (Nanjing University)
GenerationTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Propose the IF-VidCap benchmark to evaluate the instruction-following capability of multimodal large language models in video captioning tasks.
iFusion: Integrating Dynamic Interest Streams via Diffusion Model for Click-Through Rate Prediction
Ziheng Ni (JD.com), Jingping Shao (JD.com)
Recommendation SystemDiffusion modelTabularSequential
🎯 What it does: For CTR prediction, we propose a diffusion model-based interest fusion framework called IFUSION, which treats interest fusion as a conditional generation process, gradually merging long-term and short-term interests during the reverse diffusion process.
IGC-Net for conditional average potential outcome estimation over time
Konstantin Hess (LMU Munich), Stefan Feuerriegel (LMU Munich)
Recurrent Neural NetworkTransformerTabularTime SeriesElectronic Health Records
🎯 What it does: Proposed an end-to-end neural network named IGC-Net for estimating time-varying conditional average potential outcomes (CAPO) in the presence of time-varying confounders;
IGGT: Instance-Grounded Geometry Transformer for Semantic 3D Reconstruction
Hao Li (Northwestern Polytechnical University), Ziwei Liu (Singapore Lab, Nanyang Technological University)
SegmentationGenerationDepth EstimationTransformerVision Language ModelContrastive LearningGaussian SplattingImagePoint Cloud
🎯 What it does: Propose the IGGT framework, a unified end-to-end Transformer combining 3D geometric reconstruction and instance-level semantic understanding; build the InsScene-15K large-scale instance consistency dataset; propose an instance-driven scene understanding strategy;
IGU-LoRA: Adaptive Rank Allocation via Integrated Gradients and Uncertainty-Aware Scoring
Xuan Cui (Chongqing Technology and Business University), Zhanpeng Zhou (Shanghai Jiao Tong University)
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: To address parameter-efficient fine-tuning for large language models, this paper proposes an adaptive rank allocation LoRA framework called IGU-LoRA, which calculates the importance of each parameter in the low-rank matrix using integrated gradients in the parameter space, and dynamically adjusts the LoRA rank per layer by combining uncertainty-awareness (EMA + variance) to select the most important singular values.
iLLaVA: An Image is Worth Fewer Than 1/3 Input Tokens in Large Multimodal Models
Lianyu Hu (Tianjin University), Wei Feng (Tianjin University)
Computational EfficiencyLarge Language ModelVision Language ModelImageVideo
🎯 What it does: Proposed and implemented iLLaVA, which performs two-phase token merging on both the image encoder and LLM to significantly accelerate large vision-language models.
Image Can Bring Your Memory Back: A Novel Multi-Modal Guided Attack against Image Generation Model Unlearning
Renyang Liu (National University of Singapore), See-Kiong Ng (National University of Singapore)
GenerationAdversarial AttackPrompt EngineeringDiffusion modelMultimodality
🎯 What it does: Proposed a multi-modal adversarial attack framework called RECALL, which uses adversarial image prompts combined with original text prompts to jointly attack machine-unlearned image generation models, thereby evaluating and weakening their unlearning effects.
Image Quality Assessment for Embodied AI
Chunyi Li (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
Robotic IntelligenceSupervised Fine-TuningVision Language ModelVision-Language-Action ModelImageBenchmark
🎯 What it does: Established an image quality assessment (IQA) task applicable to embodied AI, constructed a large-scale database named Embodied-IQA, and conducted a systematic evaluation of the performance of existing IQA methods on this database.
ImageDoctor: Diagnosing Text-to-Image Generation via Grounded Image Reasoning
Yuxiang Guo (Johns Hopkins University), Emad Barsoum (AMD)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: Propose ImageDoctor, which combines multi-dimensional scores (interpretability, semantic alignment, feasibility, overall quality) with pixel-level defect heatmaps for diagnosing and evaluating text-to-image generation.
ImagenWorld: Stress-Testing Image Generation Models with Explainable Human Evaluation on Open-ended Real-World Tasks
Samin Mahdizadeh Sani (University of Waterloo), Wenhu Chen (University of Waterloo)
GenerationExplainability and InterpretabilityVision Language ModelImageTextBenchmark
🎯 What it does: Constructed a comprehensive human evaluation benchmark named ImagenWorld, which includes 3.6K conditional sets, six tasks (generation and editing), and six themes (art, photos, infographics, text diagrams, computer diagrams, screenshots), along with 20K fine-grained human scores and interpretable error labels.
ImageRAG: Dynamic Image Retrieval for Reference-Guided Image Generation
Rotem Shalev Arkushin, Ohad Fried (Reichman University)
GenerationRetrievalVision Language ModelDiffusion modelImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose ImageRAG, a zero-shot retrieval-augmented generation method that uses dynamically retrieved reference images to assist pre-trained text-image models in generating rare or fine-grained concepts.
Imagine How To Change: Explicit Procedure Modeling for Change Captioning
Jiayang Sun (Soochow University), Jorma Laaksonen (Aalto University)
GenerationTransformerVision Language ModelAuto EncoderImageMultimodality
🎯 What it does: Propose a two-stage ProCap framework that addresses the limitations of traditional methods which only compare static images, by explicitly modeling the image change process and implicitly describing it through queries.
Imitating the Truth: Attention-aware Truth-Guided Enhancement for Hallucination Mitigation in Large Vision-Language Models
Hairui Ren (Jilin University), Yi Chang (Jilin University)
GenerationExplainability and InterpretabilityTransformerVision Language ModelMultimodality
🎯 What it does: Propose the AGE framework, which performs attention intervention during decoding without training, reducing hallucinations in large vision-language models by mimicking the attention patterns of real responses.
Imitation Learning as Return Distribution Matching
Filippo Lazzati (Politecnico di Milano), Alberto Maria Metelli (Politecnico di Milano)
Reinforcement LearningTabular
🎯 What it does: This paper proposes a new framework for risk-sensitive imitation learning called Return Distribution Matching (RDM), aiming to make the learned policy's return distribution consistent with the expert's under Wasserstein distance, thereby replicating both the expert's average performance and risk attitude;
Implicit 4D Gaussian Splatting for Fast Motion with Large Inter-Frame Displacements
Seung-gyeom Kim (Hanyang University), Sukmin Yun (Hanyang University)
GenerationGaussian SplattingVideo
🎯 What it does: Proposed the SPIN-4DGS framework, which addresses the problem of attribute collapse caused by large displacement under fast motion by explicitly extracting spatiotemporal Gaussian positions and using a lightweight implicit network to predict Gaussian attributes.
Implicit Bias and Loss of Plasticity in Matrix Completion: Depth Promotes Low-Rankness
Baekrok Shin (KAIST), Chulhee Yun (KAIST)
OptimizationConvolutional Neural NetworkImageTabular
🎯 What it does: Investigates the implicit low-rank bias in deep matrix decomposition (deep linear networks) for matrix completion tasks, proving that deeper networks lead to more coupled training dynamics, resulting in stronger low-rank convergence; and theoretically analyzes the plasticity loss caused by pre-training.
Implicit Bias of Per-sample Adam on Separable Data: Departure from the Full-batch Regime
Beomhan Baek (Seoul National University), Chulhee Yun (Korea Advanced Institute Of Science And Technology)
ClassificationOptimization
🎯 What it does: Investigated the implicit bias of Adam and Signum in logistic regression under small-batch (batch size 1) settings, revealing that Adam loses the ℓ∞-max-margin property in mini-batch scenarios and may converge to ℓ2- or data-dependent directions, whereas Signum maintains ℓ∞-max-margin across all batch sizes.
Implicit bias produces neural scaling laws in learning curves, from perceptrons to deep networks
Francesco D'Amico (University of Rome Sapienza), Matteo Negri (University of Rome Sapienza)
ClassificationOptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper investigates the implicit bias driven by the growth of model weight norm during gradient descent training, and proposes a new neural network scaling law that describes the power-law relationship between learning curves and norm during the period from initial training to convergence.
Implicit Inversion turns CLIP into a Decoder
Antonio D'Orazio, Iacopo Masi (Sapienza University of Rome)
GenerationTransformerNeural Radiance FieldImageTextMultimodality
🎯 What it does: Achieved text-to-image generation by optimizing a frequency-aware implicit neural representation (INR) on a frozen CLIP image encoder, without requiring a pre-trained decoder or CLIP fine-tuning.
Implicit Regularisation in Diffusion Models: An Algorithm-Dependent Generalisation Analysis
Tyler Farghly (University of Oxford), Arnaud Doucet (University of Oxford)
GenerationDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: Propose a theoretical framework based on algorithm stability (score stability) to analyze the sensitivity of diffusion models during the learning and sampling processes to training data, and provide upper bounds on the generalization error for various common training and sampling algorithms.
Implicit Regularization of SGD Reduces Shortcut Learning
Nahal Mirzaie (Sharif University of Technology), Mohammad Hossein Rohban (Sharif University of Technology)
OptimizationExplainability and InterpretabilityData-Centric LearningConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: Investigated the impact of implicit regularization in stochastic gradient descent (SGD) under large learning rates and small batch sizes on models relying on spurious features, and validated the mechanism through theoretical proofs and experiments.
Importance Sampling for Multi-Negative Multimodal Direct Preference Optimization
Xintong Li (University of California San Diego), Jingbo Shang (University of California San Diego)
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningVision Language ModelAuto EncoderMultimodality
🎯 What it does: Proposed the MISP-DPO framework, utilizing multi-semantic multi-negative samples and Plackett-Luce multi-negative optimization to enhance the alignment effectiveness of multi-modal Direct Preference Optimization.
ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Cases
Ziqian Zhong (Carnegie Mellon University), Nicholas Carlini (Anthropic)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the IMPOSSIBLEBENCH framework to quantify the cheating tendency of LLMs when facing test cases.
Improved $\ell_{p}$ Regression via Iteratively Reweighted Least Squares
Alina Ene (Boston University), Adrian Vladu (CNRS and IRIF Université Paris Cité)
OptimizationComputational EfficiencyTabular
🎯 What it does: Proposed a fast ℓp regression algorithm based on iterative reweighted least squares (IRLS), with an implementation achieving iteration complexity that matches the theoretical optimum.
Improved Adversarial Diffusion Compression for Real-World Video Super-Resolution
Bin Chen (Peking University), Jian Zhang (ByteDance Inc.)
Super ResolutionKnowledge DistillationTransformerDiffusion modelVideo
🎯 What it does: Design an improved adversarial diffusion compression method (AdcVSR), compressing the large 3D Diffusion Transformer (DOVE) into a 2D+1D structure to achieve efficient Real-VSR.
Improved high-dimensional estimation with Langevin dynamics and stochastic weight averaging
Stanley Wei (Princeton University), Jason D. Lee (UC Berkeley)
OptimizationStochastic Differential Equation
🎯 What it does: Under high-dimensional hypothesis models (Tensor PCA and single-index models), we propose a learning algorithm that combines Langevin dynamics with iterative averaging, and prove that the hidden direction θ⋆ can be recovered when the sample size n ≈ d⌈k⋆/2⌉, achieving nearly optimal sample complexity.
Improved Object-Centric Diffusion Learning with Registers and Contrastive Alignment
Bac Nguyen (Sony AI), Yuki Mitsufuji (Stanford University)
Object DetectionSegmentationGenerationRepresentation LearningTransformerSupervised Fine-TuningDiffusion modelContrastive LearningImage
🎯 What it does: Propose the CODA framework, combining input-independent registration slots and contrastive alignment loss to improve diffusion-based object-centric learning
Improving 2D Diffusion Models for 3D Medical Imaging with Inter‑Slice Consistent Stochasticity
Chenhe Du (ShanghaiTech University), Yuyao Zhang (ShanghaiTech University)
RestorationGenerationSuper ResolutionDiffusion modelBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Proposes the Inter-Slice Consistent Stochasticity (ISCS) method, which utilizes spherical linear interpolation (Slerp) to generate correlated noise between adjacent slices during the resampling step of 2D diffusion models, thereby enhancing cross-slice consistency in 3D medical image reconstruction.
Improving and Accelerating Offline RL in Large Discrete Action Spaces with Structured Policy Initialization
Matthew Landers (University of Virginia), Afsaneh Doryab (University of Virginia)
TransformerReinforcement LearningContrastive LearningTabular
🎯 What it does: This paper proposes a two-stage framework called Structured Policy Initialization (SPIN), which first obtains an Action Structure Model (ASM) to capture the structure of discrete composite actions through self-supervised pre-training, and then freezes this representation and uses a lightweight policy head for control;
Improving Attributed Long-form Question Answering with Intent Awareness
Xinran Zhao (Allen Institute for AI), Varsha Kishore (Allen Institute for AI)
GenerationLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and validated an intent-aware writing framework, embedding fine-grained intent labels at the paragraph and citation levels into long-form question-answering generation, significantly improving model output quality.
Improving Autoregressive Video Modeling with History Understanding
Wenyang Luo (Institute of Automation, Chinese Academy of Sciences), Kun Gai (Kuaishou Technology)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: This paper proposes the MiMo (Masked History Modeling) framework, which integrates autoregressive video generation (VideoAR) with masked history modeling to enhance the quality of internal representations of historical frames, thereby improving video prediction and generation performance.
Improving Black-Box Generative Attacks via Generator Semantic Consistency
Jongoh Jeong, Kuk-Jin Yoon (Korea Advanced Institute Of Science And Technology)
Adversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: Studied black-box generative adversarial attacks, proposing to enhance the transfer performance across different models, domains, and tasks by applying Mean Teacher's semantic consistency constraints on early residual blocks of the generator.
Improving Block-Wise LLM Quantization by 4-bit Block-Wise Optimal Float (BOF4): Analysis and Variations
Patrick Blumenberg (Technische Universität Braunschweig), Tim Fingscheidt (Technische Universität Braunschweig)
Computational EfficiencyLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Studied and improved LLM weight quantization methods, proposing BOF4/BOF4-S and OPQ, and evaluated their effectiveness in inference and fine-tuning.
Improving Classifier-Free Guidance in Masked Diffusion: Low-Dim Theoretical Insights with High-Dim Impact
Kevin Rojas (Georgia Institute of Technology), Molei Tao (Georgia Institute of Technology)
GenerationDiffusion modelImageText
🎯 What it does: Conducts a low-dimensional theoretical analysis of unconditional guidance (classifier-free guidance, CFG) in discrete masked diffusion, and proposes an improved mechanism through column normalization; meanwhile, investigates scheduling strategies for guidance strength over time.
Improving Code Localization with Repository Memory
Boshi Wang (Ohio State University), Dongdong Chen (Ohio State University)
AI Code AssistantAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a memory mechanism leveraging repository submission history (event memory and semantic memory), integrated into the existing code localization framework LocAgent to enhance code localization performance.
Improving Diffusion Models for Class-imbalanced Training Data via Capacity Manipulation
Feng Hong (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
GenerationData-Centric LearningDiffusion modelImage
🎯 What it does: Propose and implement a capacity manipulation method that reserves and allocates model capacity specifically for the minority class, thereby improving the generation quality of diffusion models under imbalanced data scenarios.
Improving Discrete Diffusion Unmasking Policies Beyond Explicit Reference Policies
Chunsan Hong (Graduate School of AI, KAIST), Jong Chul Ye (Graduate School of AI, KAIST)
GenerationTransformerDiffusion modelText
🎯 What it does: This paper proposes and trains a learnable decoding strategy to replace traditional heuristics (e.g., max-confidence) used in Masked Diffusion Models for selecting unmasked positions.
Improving Extreme Wind Prediction with Frequency-Informed Learning
Chenrui Xu (Chinese University of Hong Kong), Jianwei Huang (Chinese University of Hong Kong)
Time SeriesPhysics Related
🎯 What it does: Propose an extreme wind speed prediction method based on frequency information, mainly through gradient penalty loss, physics-embedded Navier-Stokes network structure, and frequency domain separation and reweighting to improve the prediction accuracy of extreme wind speeds.
Improving Feasibility via Fast Autoencoder-Based Projections
Maria Chzhen (University of Toronto), Priya L. Donti (Massachusetts Institute of Technology)
OptimizationReinforcement LearningAuto EncoderGenerative Adversarial Network
🎯 What it does: Proposes a fast projection method based on autoencoders (FAB), which constructs a convex mapping in the latent space to distinguish feasible and infeasible regions, enabling one-time rapid approximation of outputs that satisfy non-convex constraints.
Improving Human-AI Coordination through Online Adversarial Training and Generative Models
Paresh R. Chaudhary (University of Washington), Natasha Jaques (University of Washington)
Reinforcement Learning from Human FeedbackReinforcement LearningAuto EncoderGenerative Adversarial NetworkBenchmark
🎯 What it does: A new zero-shot collaborative learning method called GOAT is proposed by combining a pre-trained generative variational autoencoder with online adversarial training, which can generate challenging partners in various collaborative environments to enhance AI collaboration performance with different human partners.
Improving LLM-based Global Optimization with Search Space Partitioning
Andrej Schwanke (University of Freiburg), Arber Zela (University of Freiburg)
OptimizationLarge Language ModelPrompt EngineeringTabular
🎯 What it does: Propose a global optimization algorithm called HOLLM that integrates KD-tree partitioning, Bandit heuristic scoring, and candidate point generation by large language models (LLMs);
Improving Long-Range Interactions in Graph Neural Simulators via Hamiltonian Dynamics
Tai Hoang (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
Graph Neural NetworkGraphBenchmarkPhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposed the Information-Preserving Graph Neural Simulator (IGNS), which learns continuous dynamics through port-Hamiltonian dynamics to achieve efficient long-distance physical system simulation;
Improving Online-to-Nonconvex Conversion for Smooth Optimization via Double Optimism
Francisco Patitucci (University of Texas at Austin), Aryan Mokhtari (University of Texas at Austin)
Optimization
🎯 What it does: Propose a dual optimistic online gradient method to improve the online-to-nonconvex transformation framework for solving smooth nonconvex optimization problems.
Improving Reasoning for Diffusion Language Models via Group Diffusion Policy Optimization
Kevin Rojas (Georgia Institute of Technology), Wei Deng (Morgan Stanley)
OptimizationLarge Language ModelReinforcement LearningDiffusion modelTextBenchmark
🎯 What it does: Proposed a reinforcement learning algorithm called Group Diffusion Policy Optimization (GDPO), specifically designed to enhance the inference capabilities of diffusion language models (DLM).
Improving Semantic Proximity in Information Retrieval through Cross-Lingual Alignment
Seongtae Hong (Korea University), Heuiseok Lim (Korea University)
RetrievalContrastive LearningText
🎯 What it does: Proposed a multi-language coexistence retrieval evaluation scenario and the Max@R metric, and improved cross-lingual retrieval performance through a joint training strategy combining Jensen-Shannon Divergence and InfoNCE.
Improving Set Function Approximation with Quasi-Arithmetic Neural Networks
Tomas Tokar (University of Toronto), Scott Sanner (University of Toronto)
Representation LearningFlow-based ModelImagePoint Cloud
🎯 What it does: This paper proposes a framework for approximating set functions by utilizing a learnable Kolmogorov mean (NKM) as an aggregation function, constructing learnable quantum arithmetic neural networks (QUANNs).
Improving the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models
Weiqing He (University of Pennsylvania), Qi Long (University of Pennsylvania)
GenerationOptimizationTransformerLarge Language ModelText
🎯 What it does: This paper investigates the trade-off between watermark strength and speculative sampling efficiency in large language model outputs, proposes a quantifiable watermark strength metric, converts the trade-off into a constrained optimization problem using this metric, derives the Pareto curve, and designs a pseudo-random acceptance mechanism to achieve maximum watermark strength while maintaining the highest sampling efficiency. The effectiveness of this approach is validated through experiments.
IMSE: Intrinsic Mixture of Spectral Experts Fine-tuning for Test-Time Adaptation
Sunghyun Baek (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
Domain AdaptationComputational EfficiencyTransformerMixture of ExpertsImage
🎯 What it does: The IMSE framework is proposed for Test-Time Adaptation (TTA) and Continuous Test-Time Adaptation (CTTA). It leverages SVD decomposition of the linear layer in ViT, updating only the singular values to achieve parameter-efficient adaptation, and suppresses feature collapse through a diversity maximization loss. In CTTA, domain-aware spectral code retrieval is introduced to rapidly reuse previously adapted parameters from other domains.
In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations
Mohammad Aflah Khan (Max Planck Institute for Software Systems), Abhilasha Ravichander (Max Planck Institute for Software Systems)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Investigate and quantify the potential source preferences of large language models (LLMs) across different domains (news, academic papers, e-commerce) through direct and indirect comparative experiments validated on real-world data;
In Context Semi-Supervised Learning
Jiashuo Fan (Duke University), Xiang Cheng (Duke University)
ClassificationTransformerImageGraph
🎯 What it does: Proposed and implemented a two-stage Transformer architecture for semi-supervised context learning (IC-SSL), which adaptively learns representations of unlabeled data and performs classification prediction within the same network.
In Good GRACES: Principled Teacher Selection for Knowledge Distillation
Abhishek Panigrahi, Surbhi Goel (Microsoft Research)
Knowledge DistillationText
🎯 What it does: Proposes a lightweight teacher selection score named GRACE, which leverages the gradient distribution characteristics of the student on teacher-generated data to evaluate the compatibility between teacher and student, thereby predicting the student's performance after distillation without using a validator, teacher logits, or test data.
In-Context Algebra
Eric Todd (Northeastern University), David Bau (Northeastern University)
Explainability and InterpretabilityRepresentation LearningTransformerSequential
🎯 What it does: Studying the in-context reasoning mechanism of Transformer in algebraic tasks represented solely by variable symbols without fixed meanings, training models to use word symbols as variables within each sequence and analyzing how they learn algebraic operations.
In-Context Algorithm Emulation in Fixed-Weight Transformers
Jerry Yao-Chieh Hu (Northwestern University), Han Liu (Northwestern University)
OptimizationComputational EfficiencyMeta LearningTransformerPrompt EngineeringTabular
🎯 What it does: This paper demonstrates that a frozen-weight Transformer (using only softmax attention) can simulate a class of widely ranging algorithms, including gradient descent, linear regression, ridge regression, etc., through prompts; in two modes: in the task-specific mode, single-head attention can approximate any continuous function f(wxᵀ−y); in the prompt-programmable mode, a single two-layer attention module can achieve all these algorithms by altering prompts;
In-Context Compositional Q-Learning for Offline Reinforcement Learning
Qiushui Xu (Penn State University), Jiang Bian (Microsoft Research)
RetrievalTransformerReinforcement LearningSequential
🎯 What it does: Proposes a framework called ICQL that treats Q-learning as a context reasoning problem in offline reinforcement learning, leveraging linear Transformers to retrieve local transitions for adaptive inference of local Q-functions
In-Context Learning for Pure Exploration
Alessio Russo (Boston University), Aldo Pacchiano (Boston University)
Meta LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningImage
🎯 What it does: Studied the pure exploration problem, proposing the ICPE framework that uses Transformer to actively collect data and infer correct hypotheses through Meta-RL in contextual scenarios.
In-Context Learning of Temporal Point Processes with Foundation Inference Models
David Berghaus (Lamarr Institute), Ramses J Sanchez
Meta LearningTransformerSupervised Fine-TuningTime SeriesSequentialBenchmark
🎯 What it does: Proposes FIM-PP, a foundational inference model for marked time point processes (MTPP), capable of directly estimating the conditional intensity function of event sequences under zero-shot or few-shot settings.
In-Context Multi-Objective Optimization
Xinyu Zhang (Aalto University), Samuel Kaski (Aalto University)
OptimizationTransformerReinforcement LearningTabular
🎯 What it does: Proposes a fully amortized multi-objective black-box optimization framework TAMO, which utilizes transformers to directly generate the next query based on historical data, without requiring modeling or hyperparameter tuning.
In-Context Watermarks for Large Language Models
Yepeng Liu (University Of California Santa Barbara), Yuheng Bu (University Of California Santa Barbara)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an In-Context Watermark (ICW) technique achieved solely through prompt engineering, which can embed detectable watermarks into generated text without accessing the internal decoding process of LLMs;
In-Place Test-Time Training
Guhao Feng (ByteDance Seed), Tianle Cai (ByteDance Seed)
OptimizationComputational EfficiencyMeta LearningTransformerLarge Language ModelText
🎯 What it does: Propose an In-Place training framework during inference, enabling large language models to dynamically update the MLP projection matrix during reasoning, achieving immediate adaptation to long contexts.
In-the-Flow Agentic System Optimization for Effective Planning and Tool Use
Zhuofeng Li (Stanford University), Pan Lu (Stanford University)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Proposes AGENTFLOW, a trainable agent system that collaborates within a process, integrating a planner, executor, validator, and generator, and achieving multi-tool long-term reasoning through evolvable memory.
Incentive-Aligned Multi-Source LLM Summaries
Yanchen Jiang (Harvard University), Aranyak Mehta (Google Research)
GenerationTransformerLarge Language ModelText
🎯 What it does: This paper proposes the Truthful Text Summarization (TTS) framework, which incentivizes truthful behavior in content sources by splitting summaries into atomic statements and using multi-task peer prediction to score sources;
Incentives in Federated Learning with Heterogeneous Agents
Ariel D. Procaccia (Harvard University), Itai Shapira (Harvard University)
OptimizationFederated LearningImage
🎯 What it does: Proposes a game model based on PAC threshold to analyze incentive issues in federated learning under heterogeneous data distributions, proving that pure Nash equilibrium may not exist and optimal equilibrium cost can be infinitely high, and designs a computable linear programming approximation and strategy-independent 'pay-as-you-go' mechanism;
Incentivizing Agentic Reasoning in LLM Judges via Tool-Integrated Reinforcement Learning
Ran Xu (Emory University), Hongkun Yu (Google)
AI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Proposed TIR-Judge, an end-to-end reinforcement learning framework for training LLM judges capable of calling Python executors, enabling reasoning, code generation, and execution to achieve verifiable judgment results in evaluation tasks.
Incentivizing Consistent, Effective and Scalable Reasoning Capability in Audio LLMs via Reasoning Process Rewards
Jiajun Fan (University of Illinois Urbana-Champaign), Yile Gu (Amazon)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningChain-of-ThoughtAudio
🎯 What it does: Train audio large language models for reasoning, addressing the issue of inverse scale decline during reasoning, and achieving controllable, structured reasoning chains;
Incentivizing LLM Reasoning via Reinforcement Learning with Functional Monte Carlo Tree Search
Kongcheng Zhang (Zhejiang University), Shunyu Liu (Nanyang Technological University)
OptimizationTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a Reinforcement Learning-based Functional Token Tuning (RFTT) framework, which internalizes the reasoning behavior of LLMs through learnable functional tokens (e.g., <analyze>, <verify>), and achieves self-improvement via self-generated tree search samples.
InclusiveVidPose: Bridging the Pose Estimation Gap for Individuals with Limb Deficiencies in Video-Based Motion
Heming Du (University of Queensland), Xin Yu (University of Adelaide)
Pose EstimationVideoBenchmark
🎯 What it does: Proposed the InclusiveVidPose dataset and evaluated the performance of existing multi-person pose estimation methods on individuals with limb loss.
Incomplete Data, Complete Dynamics: A Diffusion Approach
Zihan Zhou (Chinese University of Hong Kong Shenzhen), Tianshu Yu (Chinese University of Hong Kong Shenzhen)
RestorationDiffusion modelImagePhysics Related
🎯 What it does: Propose a diffusion model-based framework that can directly learn and reconstruct complete physical dynamics processes from data with only incomplete observations.
Incomplete Multi-View Multi-Label Classification via Shared Codebook and Fused-Teacher Self-Distillation
Xu Yan (Shanghai Maritime University), Minghua Wan (Shanghai Maritime University)
ClassificationKnowledge DistillationRepresentation LearningMultimodality
🎯 What it does: Studies the multi-view multi-label classification problem with both views and labels incomplete, proposing the SCSD framework to achieve joint learning.
Inconsistency Biases in Dynamic Data Pruning
Qing Zhou (Northwestern Polytechnical University), Qi Wang (Northwestern Polytechnical University)
Computational EfficiencyData-Centric LearningImageVideoTextMultimodalityAudio
🎯 What it does: Propose the RePB framework to address the issues of score context drift and gradient time bias in dynamic data pruning, achieving efficient training through local window pruning, uniform resampling, and cumulative temporal rescaling.
Incorporating Expert Priors into Bayesian Optimization via Dynamic Mean Decay
Chongqi Qu (Xi'an Jiaotong University), Zhunga Liu (Northwestern Polytechnical University)
OptimizationHyperparameter SearchBenchmark
🎯 What it does: Introduce expert priors in Bayesian optimization and achieve gradual abandonment of priors through a dynamically decaying mean function, enabling the algorithm to rapidly leverage priors for accelerated search in the early stages, while gradually restoring the robustness of standard BO in later stages.
IncVGGT: Incremental VGGT for Memory-Bounded Long-Range 3D Reconstruction
Keyu Fang (Duke University), Yiran Chen (Duke University)
Computational EfficiencyTransformerVideoMultimodality
🎯 What it does: Propose IncVGGT, an incremental and memory-friendly visual geometric transformer for long-sequence 3D reconstruction;
Independence Test for Linear Non-Gaussian Data and Applications in Causal Discovery
Yiqing Li (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
TabularBiomedical Data
🎯 What it does: Proposes a specialized independence test method for linear non-Gaussian mixed data, simplifying independence determination to checking the constancy of conditional mean and variance, and constructing a single kernel statistic LiNGIC based on this.
IndicVisionBench: Benchmarking Cultural and Multilingual Understanding in VLMs
Ali Faraz (Krutrim AI), Shubham Agarwal (Krutrim AI)
Vision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Constructed IndicVisionBench, a multilingual multimodal benchmark covering 13 Indian cultural themes, 5K images, and 37K+ question-answer pairs, encompassing three tasks (VQA, OCR, and MMT), and conducted systematic evaluations of 8 mainstream VLMs (including Gemini, GPT-4o, LLaMA-4, Gemma-3, Chitrarth, etc.).
Inducing Dyslexia in Vision Language Models
Melika Honarmand (École Polytechnique Fédérale de Lausanne), Martin Schrimpf (École Polytechnique Fédérale de Lausanne)
Explainability and InterpretabilityTransformerVision Language ModelImageTextBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Leverage large-scale vision-language models to locate and ablate visual morphological units similar to those in the visual word form area (VWFA), simulate reading difficulties resembling dyslexia, and verify that the model's behavior aligns with human dyslexics.
Inductive Reasoning for Temporal Knowledge Graphs with Emerging Entities
Ze Zhao (Shanghai Jiao Tong University), Chenghu Zhou (Chinese Academy Of Sciences)
Representation LearningTransformerLarge Language ModelTextGraph
🎯 What it does: Proposed the TRANSFIR framework for inductive reasoning in temporal knowledge graphs when dealing with new entities lacking historical interactions;
InfBaGel: Human-Object-Scene Interaction Generation with Dynamic Perception and Iterative Refinement
Yude Zou (Shanghai Jiao Tong University), Guanjie Zheng (Sichuan University)
GenerationData SynthesisRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelMultimodality
🎯 What it does: Generating complete human-object-scene interaction action sequences in dynamic scenes
Inference-Cost-Aware Dynamic Tree Construction for Efficient Inference in Large Language Models
Yinrong Hong (Beihang University), Kai Hu (Beihang University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes CAST (Cost-Aware Speculative Tree) — a method for accelerating LLM inference by dynamically constructing tree drafts and adaptively pruning based on system costs such as GPU devices, batch sizes, etc.;
Inference-Time Dynamic Modality Selection for Incomplete Multimodal Classification
Siyi Du (Imperial College London), Chen Qin (Imperial College London)
ClassificationTransformerContrastive LearningImageTextMultimodalityTabularBiomedical Data
🎯 What it does: Designed a dynamic modal selection framework DyMo for inference, addressing incomplete multimodal classification by adaptively fusing restored reliable modalities to maximize task-related information.
Inference-Time Personalized Safety Control via Paired Difference-in-Means Intervention
Tran Huynh (Virginia Tech), Ruoxi Jia (Virginia Tech)
Safty and PrivacyPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: The paper proposes a training-agnostic, activation intervention method during inference to achieve personalized safe control, utilizing three estimation strategies (ILCS, UMS, PCMS) to compute intervention directions and evaluating on multi-model, multi-safety surfaces.
Inference-time scaling of diffusion models through classical search
XiangCheng Zhang, Yilun Du (Harvard University)
GenerationOptimizationComputational EfficiencyDiffusion modelImageTextPoint Cloud
🎯 What it does: Propose a unified inference-time diffusion model scaling framework that combines classical search (BFS/DFS) with local Annealed Langevin MCMC, performing global and local searches during the inference phase to achieve target control and performance improvement across multiple tasks (image generation, planning, offline reinforcement learning).
Inference-Time Scaling of Discrete Diffusion Models via Importance Weighting and Optimal Proposal Design
Zijing Ou (Imperial College London), Yingzhen Li (Imperial College London)
GenerationDiffusion modelImageTextBiomedical Data
🎯 What it does: Propose a Sequential Monte Carlo (SMC) framework tailored for discrete diffusion models, enabling scalable control and alignment during testing
Inferring brain plasticity rule under long-term stimulation with structured recurrent dynamics
Zhichao Liang (Southern University of Science and Technology), Quanying Liu (Southern University of Science and Technology)
Recurrent Neural NetworkTime SeriesBiomedical Data
🎯 What it does: This study proposes the STEER framework to learn brain plasticity rules under long-term stimulation, formalizing long-term plasticity as a potential dynamic system under stimulation conditions.
Inferring the Invisible: Neuro-Symbolic Rule Discovery for Missing Value Imputation
Wendi Ren (Chinese University of Hong Kong), Shuang Li (Chinese University of Hong Kong)
Data SynthesisExplainability and InterpretabilityTabular
🎯 What it does: Propose a neural-symbolic framework that achieves reasoning and imputation of missing entries in incomplete tables through bidirectional feedback loops between rule learning and missing value imputation.
Infinite Horizon Markov Economies
Denizalp Goktas (Cornell Tech), Amy Greenwald (Brown University)
OptimizationGenerative Adversarial NetworkFinance Related
🎯 What it does: This paper studies the Markov pseudo-game (MPG) framework and proves the existence of recursive Radner equilibria in infinite-horizon Markov exchange economies, proposing a polynomial-time approximation solution method.
InfLLM-V2: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation
Weilin Zhao (Tsinghua University), Zhiyuan Liu (Tsinghua University)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose an InfLLM-V2 framework that can switch to sparse attention without additional parameters or significant distribution changes after short-sequence pretraining, achieving long-sequence processing.
Influence Dynamics and Stagewise Data Attribution
Jin Hwa Lee (University College London), Jesse Hoogland (Timaeus)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed and validated a training data attribution framework based on stage-wise development (Singular Learning Theory), demonstrating that the impact of training samples dynamically changes during the learning process;
Influence without Confounding: Causal Discovery from Temporal Data with Long-term Carry-over Effects
Fan Li (Tsinghua University), TAN Kun
Explainability and InterpretabilityComputational EfficiencyReinforcement LearningTime SeriesPhysics Related
🎯 What it does: Propose a time series causal structure learning method called LEVER for long-term deferred effects
Influence-Preserving Proxies for Gradient-Based Data Selection in LLM FineTuning
Sirui Chen (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
OptimizationComputational EfficiencyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Designed the IPROX two-stage framework to generate an influence-preserving proxy model for gradient data selection in large language models.
InfoBridge: Mutual Information estimation via Bridge Matching
Sergei Kholkin (Applied AI Institute), Alexander Korotin (Applied AI Institute)
Diffusion modelImageTextStochastic Differential Equation
🎯 What it does: By constructing an unbiased mutual information estimator, Info Bridge, the efficient estimation of mutual information for any random variables is achieved through the diffusion bridge matching method.