NeurIPS 2025 Papers — Page 23
Conference on Neural Information Processing Systems · 5275 papers
Improving Bilinear RNN with Closed-loop Control
Jiaxi Hu (Hong Kong University of Science and Technology), Weigao Sun (Shanghai AI Laboratory)
ClassificationObject TrackingRecurrent Neural NetworkReinforcement LearningImageText
🎯 What it does: A closed-loop control-based variant of Bilinear RNN called Comba is proposed and implemented, combining state feedback and output correction, using scalar + low-rank (SPLR) state transformation;
Improving Data Efficiency for LLM Reinforcement Fine-tuning Through Difficulty-targeted Online Data Selection and Rollout Replay
Yifan Sun (University of Illinois Urbana-Champaign), Huan Zhang (University of Illinois Urbana-Champaign)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed two techniques, Difficulty-targeted Online Data Selection (DOTS) and Rollout Replay (RR), which significantly improve the data efficiency of LLM reinforcement learning fine-tuning.
Improving Decision Trees through the Lens of Parameterized Local Search
Juha Harviainen (University of Helsinki), Manuel Sorge (TU Wien)
ClassificationOptimizationTabularBenchmark
🎯 What it does: This paper studies the execution of two local search operations, threshold adjustment and switching, on an already generated decision tree, aiming to reduce classification errors within a fixed number of operations, and conducts a systematic analysis of its parameterized complexity;
Improving Diffusion-based Inverse Algorithms under Few-Step Constraint via Linear Extrapolation
Jiawei Zhang (Tsinghua University), Yuantao Gu (Tsinghua University)
RestorationGenerationDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a general three-module (sampler, corrector, noise generator) canonical form that unifies existing diffusion inverse problem algorithms, and based on this framework, designs a learnable linear extrapolation method (LLE) to enhance the performance of inverse problem solving within a limited number of steps.
Improving Energy Natural Gradient Descent through Woodbury, Momentum, and Randomization
Andres Guzman-Cordero, Marius Zeinhofer (ETH Zurich)
OptimizationComputational EfficiencyTabularPhysics Related
🎯 What it does: This paper proposes an improved Energy Natural Gradient Descent (ENGD) method for training Physics-Informed Neural Networks (PINNs), significantly enhancing convergence speed and accuracy.
Improving Formal Reasoning of Transformer with State Stack
Kechi Zhang (Peking University), Zhi Jin (Peking University)
TransformerText
🎯 What it does: STACKTRANS is proposed, which adds a differentiable hidden state stack between Transformer layers to enhance the inductive and reasoning capabilities for formal and natural languages.
Improving Generalization of Neural Combinatorial Optimization for Vehicle Routing Problems via Test-Time Projection Learning
Yuanyao Chen (Southern University of Science and Technology), Zhenkun Wang (Southern University of Science and Technology)
OptimizationTransformerLarge Language ModelGraph
🎯 What it does: This paper proposes a reasoning-time projection learning framework (TTPL) and multi-view decision fusion (MVDF) to enhance the generalization ability of neural combinatorial optimization models for large-scale vehicle routing problems.
Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking
Junhyuk So (POSTECH), Eunhyeok Park (POSTECH)
Robotic IntelligenceReinforcement LearningDiffusion modelSequential
🎯 What it does: By incorporating past state-based negative guidance (Self-Guidance) and adaptive chunking that dynamically switches between open-loop and closed-loop control based on action similarity during the denoising process of the Diffusion Policy, the action quality and real-time responsiveness of robot behavior cloning are improved.
Improving LLM General Preference Alignment via Optimistic Online Mirror Descent
Yuheng Zhang (University of Illinois), Dong Yu
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a self-play framework based on Optimistic Online Mirror Descent (ONPO) for aligning large language models (LLMs) to general preferences without relying on BT model assumptions.
Improving Model Representation and Reducing KV Cache via Skip Connections with First Value Heads
Zhoutong Wu (Peking University), Zhouchen Lin (Peking University)
GenerationOptimizationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes SkipV1Former, a variant of cross-layer skip connections that uses the first layer's Value head in the deep layers of the Transformer, significantly reducing KV cache and enhancing representational capacity.
Improving Model-Based Reinforcement Learning by Converging to Flatter Minima
Shrinivas Ramasubramanian (Carnegie Mellon University), Jeff Schneider (Carnegie Mellon University)
Robotic IntelligenceReinforcement LearningWorld ModelSequential
🎯 What it does: This paper studies the improvement of model-based reinforcement learning (MBRL) by encouraging the flatness of model training loss and proposes a method to integrate Sharpness-Aware Minimization (SAM) into world model training to enhance the performance of control policies.
Improving Monte Carlo Tree Search for Symbolic Regression
Zhengyao Huang (Peking University), Yuanhui Wen (Huawei Technologies)
OptimizationReinforcement LearningTabularBenchmark
🎯 What it does: This paper studies and improves Monte Carlo Tree Search (MCTS) for symbolic regression, proposing an extreme multi-armed bandit allocation strategy and evolution-inspired state transition actions.
Improving Perturbation-based Explanations by Understanding the Role of Uncertainty Calibration
Thomas Decker (Siemens AG), Florian Buettner (German Cancer Research Center)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerImageTabular
🎯 What it does: This study investigates the relationship between uncertainty calibration and perturbation-based explanations, proposing the ReCalX method to enhance model calibration under explanation-related perturbations, thereby improving explanation quality.
Improving planning and MBRL with temporally-extended actions
Palash Chatterjee (Indiana University), Roni Khardon (Indiana University)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: By treating the duration of actions as a planning variable, this paper introduces temporally extended actions and learns the corresponding dynamics model within a model-based reinforcement learning framework, thereby improving planning and training efficiency.
Improving Progressive Generation with Decomposable Flow Matching
Moayed Haji-Ali (Rice University), Aliaksandr Siarohin (Snap Inc)
GenerationData SynthesisFlow-based ModelImageVideo
🎯 What it does: The Decomposable Flow Matching (DFM) framework is proposed to achieve progressive visual generation under multi-scale representations, avoiding the need for multiple models or complex diffusion processes.
Improving Regret Approximation for Unsupervised Dynamic Environment Generation
Harry Mead, Nick Hawes
Robotic IntelligenceReinforcement Learning
🎯 What it does: This paper proposes a Dynamic Environment Generation (DEGen) method and a new regret approximation called Maximised Negative Advantage (MNA) for automatically generating reinforcement learning training curricula, significantly enhancing zero-shot performance.
Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning
Yiqun Chen (Renmin University of China), Jiaxin Mao (Renmin University of China)
RetrievalOptimizationLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: This study proposes a joint optimization method for a multi-module retrieval-augmented generation (RAG) system, called MMOA-RAG. By treating modules such as query rewriting, document selection, and answer generation as multi-agents that share rewards and employing multi-agent PPO (MAPPO) for reinforcement learning, the method achieves collaborative optimization of the entire RAG pipeline.
Improving Reward Models with Proximal Policy Exploration for Preference-Based Reinforcement Learning
Yiwen Zhu (Zhejiang University), Bo An (Nanyang Technological University)
Reinforcement LearningSequentialBenchmark
🎯 What it does: A Proximal Policy Exploration (PPE) method is proposed to actively expand the coverage of the Preference Buffer to enhance the reliability of the reward model and accelerate preference-driven reinforcement learning.
Improving Target Sound Extraction via Disentangled Codec Representations with Privileged Knowledge Distillation
Dail Kim (Hanyang University), Joon-Hyuk Chang (Hanyang University)
Knowledge DistillationRepresentation LearningAudio
🎯 What it does: This paper proposes a knowledge distillation framework for target sound extraction using disentangled features from a decoder, called Disentangled Codec Knowledge Distillation (DCKD).
Improving Task-Specific Multimodal Sentiment Analysis with General MLLMs via Prompting
Haoyu Zhang (Chinese University of Hong Kong), Tianshu Yu (Chinese University of Hong Kong)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextMultimodalityAudio
🎯 What it does: This paper proposes a teacher-student framework guided by MLLM (MMSLF), which trains the teacher model using context prompts generated by a general MLLM, and then distills knowledge to the student model, enabling it to perform multimodal sentiment analysis without using prompts.
Improving the Euclidean Diffusion Generation of Manifold Data by Mitigating Score Function Singularity
Zichen Liu (Peking University), Tiejun Li (Peking University)
GenerationData SynthesisDiffusion modelMeshStochastic Differential Equation
🎯 What it does: This paper studies the issue of multiscale singularity in the score function when generating data using the Euclidean diffusion model under a known manifold structure, and proposes two methods to alleviate this problem.
Improving the Generation and Evaluation of Synthetic Data for Downstream Medical Causal Inference
Harry Amad (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
GenerationData SynthesisFlow-based ModelGenerative Adversarial NetworkTabularBiomedical Data
🎯 What it does: A theoretical framework for generating and evaluating synthetic data containing treatment variables is proposed, and based on this, the STEAM method is developed to generate synthetic medical data that meets the needs of causal inference.
Improving the Straight-Through Estimator with Zeroth-Order Information
Ningfeng Yang (University of British Columbia), Tor M. Aamodt (University of British Columbia)
OptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: A training method combining first-order and zero-order gradients, called FOGZO, is proposed to improve Quantization-Aware Training (QAT), particularly for low-precision weight quantization.
Improving Time Series Forecasting via Instance-aware Post-hoc Revision
Zhiding Liu (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Anomaly DetectionTransformerTime SeriesBenchmark
🎯 What it does: A model-agnostic post-processing framework called PIR is proposed, which improves the reliability of time series forecasting by identifying instance-level prediction errors and correcting them using local context (covariates, external variables) and global historical retrieval.
Improving Video Generation with Human Feedback
Jie Liu (MMLab, Chinese University of Hong Kong), Wanli Ouyang (Shanghai AI Laboratory)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelFlow-based ModelRectified FlowVideoTextBenchmark
🎯 What it does: A dataset of 182k three-dimensional human preference data was constructed, a multi-dimensional reward model VideoReward was proposed, and three alignment algorithms based on flow matching (Flow-DPO, Flow-RWR, Flow-NRG) were designed to enhance the visual quality, motion quality, and text alignment of modern flow-based video generation.
In Search of Adam’s Secret Sauce
Antonio Orvieto (ELLIS Institute Tübingen), Robert M. Gower (Flatiron Institute)
OptimizationHyperparameter SearchTransformerLarge Language ModelText
🎯 What it does: A comprehensive evaluation of the training of Transformer language models in large-scale experiments, comparing the performance of Adam with various simplified optimizers (such as SignSGD, Signum, SGD, RMSprop), and revealing the key factors behind Adam's effectiveness through fine-grained hyperparameter search;
In Silico Mapping of Visual Categorical Selectivity Across the Whole Brain
Ethan Hwang (Zuckerman Mind Brain Behavior Institute Columbia University), Nikolaus Kriegeskorte (Zuckerman Mind Brain Behavior Institute Columbia University)
TransformerDiffusion modelImageMagnetic Resonance Imaging
🎯 What it does: This paper proposes a Transformer-based brain encoder that achieves a nonlinear mapping from image features to brain regions through cross-attention. Using this encoder, millions of images are mapped in silico, revealing the selectivity of cortical blocks outside the visual cortex for diverse semantic concepts, and generating 'superstimuli' that maximally activate these blocks. Additionally, experimental hypotheses are validated and generated on the same model, forming a whole-brain category selectivity map that can be directly used for subsequent fMRI experiments.
In-Context Compositional Learning vis Sparse Coding Transformer
Wei Chen (Purdue University), Qiang Qiu (Purdue University)
TransformerImage
🎯 What it does: A Transformer framework based on sparse coding is proposed, transforming the attention mechanism into a sparse projection of the input onto a coding and decoding dictionary, and estimating target coefficients through a linear combination of contextual coefficients, thereby explicitly capturing and transferring combinatorial rules.
In-Context Fully Decentralized Cooperative Multi-Agent Reinforcement Learning
Chao Li (Nanjing University of Posts and Telecommunications), Yang Gao (Nanjing University)
Reinforcement LearningSequential
🎯 What it does: The paper proposes a framework for learning collaborative strategies in a fully decentralized environment, named Return-Aware Context (RAC). It models each agent's local task as a contextual MDP and utilizes training rewards to construct context, achieving a unified solution to non-stationarity and relative overgeneralization issues.
In-context Learning of Linear Dynamical Systems with Transformers: Approximation Bounds and Depth-separation
Frank Cole (University of Minnesota), Tianhao Zhang (University of Minnesota)
TransformerTime SeriesOrdinary Differential Equation
🎯 What it does: This paper studies the approximation capability of Transformers in autoregressive learning of linear dynamical systems, providing upper bounds on the error for deep Transformers and lower bounds for single-layer Transformers, revealing the phenomenon of deep separation.
In-Context Learning of Stochastic Differential Equations with Foundation Inference Models
Patrick Seifner (University of Bonn), Ramses J Sanchez
TransformerTime SeriesFinance RelatedStochastic Differential Equation
🎯 What it does: This paper trains a pre-trained base model FIM-SDE, utilizing the Transformer and neural operator framework, which can estimate the drift and diffusion functions of SDEs from noisy time series in a contextual learning manner.
In-Context Learning Strategies Emerge Rationally
Daniel Wurgaft (Stanford University), Noah Goodman
TransformerText
🎯 What it does: By constructing a hierarchical Bayesian framework, various ICL strategies exhibited by the Transformer under different training conditions are unified and explained, and the framework is used to predict the model's next token prediction behavior.
INC: An Indirect Neural Corrector for Auto-Regressive Hybrid PDE Solvers
Hao Wei (Technical University of Munich), Nils Thuerey (Technical University of Munich)
Recurrent Neural NetworkTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: An Indirect Neural Corrector (INC) is proposed, embedding the learned correction term into the right side of the PDE, thereby achieving long-term accuracy improvement and stability enhancement in mixed operator solvers;
Incentive-Aware Dynamic Resource Allocation under Long-Term Cost Constraints
Yan Dai, Patrick Jaillet
OptimizationReinforcement LearningTabular
🎯 What it does: The study investigates how to dynamically allocate reusable resources to strategic agents under multi-dimensional long-term cost constraints, designing an incentive-friendly primal-dual mechanism to address the vulnerability of traditional methods to strategic manipulation.
Incentivizing Desirable Effort Profiles in Strategic Classification: The Role of Causality and Uncertainty
Valia Efthymiou (Massachusetts Institute of Technology), Juba Ziani (Georgia Institute of Technology)
ClassificationOptimizationBiomedical Data
🎯 What it does: This paper studies the strategic classification problem under the presence of causal structures and information uncertainty, analyzing how agents exert effort to achieve positive classification, and exploring how to design classifiers to incentivize agents to focus their efforts on desirable features as perceived by the organizer.
Incentivizing Dual Process Thinking for Efficient Large Language Model Reasoning
Xiaoxue Cheng (Renmin University of China), Zhiqiang Zhang (Ant Group)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a reinforcement learning framework called ACPO, which combines dual process theory, allowing large reasoning models to dynamically switch between fast and slow thinking modes based on task difficulty, thereby reducing redundant reasoning.
Incentivizing LLMs to Self-Verify Their Answers
Fuxiang Zhang (Nanyang Technological University), Bo An (Nanyang Technological University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a self-verification framework that unifies answer generation and self-verification training through reinforcement learning, enhancing the performance of large language models in mathematical reasoning tasks and achieving scalability during testing.
Incentivizing Reasoning for Advanced Instruction-Following of Large Language Models
Yulei Qin (Tencent Youtu Lab), Xing Sun (Tencent Youtu Lab)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: A diverse and complex instruction set is generated through self-evolution, and reinforcement learning is used to encourage LLMs to produce deep reasoning, thereby enhancing their instruction-following performance under various combinations of constraints.
Incentivizing Time-Aware Fairness in Data Sharing
Jiangwei Chen (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
Tabular
🎯 What it does: A time-sensitive fair data sharing incentive mechanism is proposed to encourage early participation and improve data quality.
Incentivizing Truthful Language Models via Peer Elicitation Games
Baiting Chen (University of California Los Angeles), Xiaowu Dai (University of California Los Angeles)
Large Language ModelReinforcement LearningText
🎯 What it does: This paper proposes an unsupervised LLM alignment framework PEG based on multi-agent peer assessment.
Incomplete Multi-view Clustering via Hierarchical Semantic Alignment and Cooperative Completion
Xiaojian Ding (Nanjing University of Finance and Economics), Xiaoying Zhu (Nanjing University of Finance and Economics)
Auto EncoderImage
🎯 What it does: A missing multi-view clustering framework based on hierarchical semantic alignment and collaborative completion, called HSACC, is proposed.
Incomplete Multi-view Deep Clustering with Data Imputation and Alignment
Jiyuan Liu (National University of Defense Technology), Kehua Guo (Central South University)
Representation LearningAuto EncoderImage
🎯 What it does: The IMDC-DIA method is proposed, which utilizes a unique latent representation to reconstruct, complete, and cluster missing multi-view data;
Increasing the Utility of Synthetic Images through Chamfer Guidance
Nicola Dall'Asen (University of Trento), Michal Drozdzal (FAIR at Meta)
ClassificationGenerationData SynthesisDiffusion modelImage
🎯 What it does: A training-free method guided by Chamfer distance is used in conditional image generation models to enhance the diversity and quality of synthetic images with a small number of real samples, and the generated data is utilized to train downstream classifiers.
Incremental Sequence Classification with Temporal Consistency
Lucas Maystre, David Barber
ClassificationTransformerReinforcement LearningTextSequential
🎯 What it does: This paper studies incremental sequence classification and proposes a loss function TCλ that utilizes temporal consistency to train sequence classifiers, enabling them to make accurate predictions based on progressively observed prefixes.
Individual Fairness In Strategic Classification
Zhiqun Zuo (Ohio State University), Mohammad Mahdi Khalili (Ohio State University)
ClassificationOptimizationTabularFinance Related
🎯 What it does: This study investigates the use of random thresholds in strategic classification to achieve individual fairness and group fairness.
Individual Regret in Cooperative Stochastic Multi-Armed Bandits
Idan Barnea (Tel Aviv University), Yishay Mansour (Tel Aviv University)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This paper studies the cooperative stochastic multi-armed bandit (MAB) problem on arbitrary connected communication graphs, proposing the Coop-SE algorithm and proving that its individual regret at each agent does not depend on the diameter of the graph and nearly achieves the theoretical lower bound. Further variants are provided under the constraints of limited message size (CONGEST model) and limited communication rounds, which also maintain a favorable individual regret.
Individually Fair Diversity Maximization
Ruien Li (East China Normal University), Yanhao Wang (East China Normal University)
Recommendation SystemOptimizationTabular
🎯 What it does: This paper studies the diversity maximization problem under individual fairness constraints and proposes the α-fair k-selection algorithm, providing approximate solutions for three objectives: max-min, max-sum, and sum-min.
Inductive Domain Transfer In Misspecified Simulation-Based Inference
Ortal Senouf (École Polytechnique Fédérale de Lausanne), Pascal Frossard (École Polytechnique Fédérale de Lausanne)
Domain AdaptationOptimizationFlow-based ModelBiomedical Data
🎯 What it does: A fully inductive and amortizable simulation-based inference method called FRISBI is proposed, which achieves accurate posterior estimation through joint distribution and point-to-point domain transfer when the model is underfitting.
IneqSearch: Hybrid Reasoning for Olympiad Inequality Proofs
Zhaoqun Li (Zhejiang University), Qiwei Ye (Beijing Academy of Artificial Intelligence)
Large Language ModelTextBenchmark
🎯 What it does: A hybrid system called IneqSearch, which combines symbolic reasoning and large language models, is proposed for the automatic proof of Olympic-level inequalities.
Inexact Column Generation for Bayesian Network Structure Learning via Difference-of-Submodular Optimization
Yiran Yang (Chinese University of Hong Kong), Rui Chen (Chinese University of Hong Kong)
OptimizationTabular
🎯 What it does: A CG-DCA method based on column generation and differential submodular (DS) minimization is proposed for solving the pricing problem in Bayesian network structure learning.
InfantAgent-Next: A Multimodal Generalist Agent for Automated Computer Interaction
Bin Lei (University of Minnesota), Caiwen Ding (Cisco Research)
Large Language ModelAgentic AIVision Language ModelImageVideoTextMultimodalityBenchmarkAudio
🎯 What it does: A multimodal general agent INFANTAGENT-NEXT is proposed, capable of interacting with computers through text, images, audio, and video;
Inference of Whole Brain Electrophysiological Networks Through Multimodal Integration of Simultaneous Scalp and Intracranial EEG
Shihao Yang (Stevens Institute of Technology), Feng Liu (Stevens Institute of Technology)
MultimodalityBiomedical DataElectrocardiogram
🎯 What it does: This paper proposes a method for fusing simultaneously acquired scalp EEG and intracranial EEG, using a first-order state space model and the EM algorithm to estimate the entire brain's electrophysiological network in a single step.
Inference with correlated priors using sisters cells
Sina Tootoonian (Francis Crick Institute), Andreas T. Schaefer (Francis Crick Institute)
Time Series
🎯 What it does: This paper proposes a perception reasoning framework that constructs structured priors using sister cells, achieving MAP inference on relevant latent factors through indirect connections.
Inference-time Alignment in Continuous Space
Yige Yuan (Institute of Computing Technology Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology Chinese Academy of Sciences)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a reasoning alignment method called Simple Energy Adaptation (SEA), which achieves reasoning alignment for large language models through gradient optimization of the energy function in a continuous latent space.
Inference-Time Hyper-Scaling with KV Cache Compression
Adrian Łańcucki (NVIDIA), Edoardo Ponti
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Achieving inference-time hyper-scaling by compressing the KV cache of Transformer LLMs, allowing the model to generate longer or more parallel inference chains under the same computational budget, thereby improving inference accuracy.
Inference-Time Personalized Alignment with a Few User Preference Queries
Victor-Alexandru Pădurean (Max Planck Institute for Software Systems), Adish Singla (Max Planck Institute for Software Systems)
GenerationRecommendation SystemReinforcement Learning from Human FeedbackLarge Language ModelImageText
🎯 What it does: A method called USERALIGN is proposed to quickly obtain user preferences through a small number of pairwise comparisons during inference and to personalize the output of generative models.
Inference-Time Reward Hacking in Large Language Models
Hadi Khalaf (Harvard University), Flavio Calmon
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies and mitigates the reward hacking problem of large language models during inference, proposing the Best-of-Poisson (BoP) sampling and HedgeTune tuning method.
Inference-Time Scaling for Flow Models via Stochastic Generation and Rollover Budget Forcing
Jaihoon Kim (Korea Advanced Institute of Science and Technology), Minhyuk Sung (Korea Advanced Institute of Science and Technology)
GenerationOptimizationFlow-based ModelTextStochastic Differential Equation
🎯 What it does: This paper proposes a scaling method for inference in pre-trained flow models, combining SDE sampling, VP interpolation transformation, and adaptive budget allocation to achieve particle sampling and enhance reward alignment.
Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search
Yuta Oshima (University of Tokyo), Hiroki Furuta (Google DeepMind)
GenerationData SynthesisComputational EfficiencyVision Language ModelDiffusion modelVideoTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A beam search method for inference during the reverse process of diffusion models (Diffusion Latent Beam Search, DLBS) is proposed, which combines a deterministic DDIM lookahead estimator and utilizes linear weighting of multi-dimensional video quality metrics to calibrate rewards, significantly improving the alignment quality of text-to-video generation without updating model parameters.
Inferring stochastic dynamics with growth from cross-sectional data
Stephen Y. Zhang, Victor Chardès (Flatiron Institute)
Time SeriesSequentialBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This study proposes the Unbalanced Probability Flow Inference (UPFI) method, which utilizes the Lagrangian form of the Fokker-Planck equation along with score matching and neural ODEs to infer stochastic dynamic models with growth (drift, noise, and cell proliferation/death) from data consisting only of cross-sectional snapshots, and trains and evaluates using unbiased Sinkhorn divergence.
InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models
Yanggan Gu (Hong Kong Polytechnic University), Hongxia Yang (Hong Kong Polytechnic University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the InfiFPO method, which achieves implicit model fusion during the preference alignment phase by directly integrating the sequence probabilities of multiple source models into the reference model, thereby enhancing the performance of the Pivot model.
InfiGFusion: Graph-on-Logits Distillation via Efficient Gromov-Wasserstein for Model Fusion
Yuanyi Wang (Hong Kong Polytechnic University), Hongxia Yang (Hong Kong Polytechnic University)
Knowledge DistillationLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Integrate multiple heterogeneous large language models into a unified model and perform knowledge distillation through a structured logits graph.
InfiniPot-V: Memory-Constrained KV Cache Compression for Streaming Video Understanding
Minsoo Kim (Hanyang University), Simyung Chang (Qualcomm)
RecognitionCompressionTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes InfiniPot-V, a training-free, query-agnostic KV cache compression framework for streaming video understanding under a fixed memory budget.
Infinite Neural Operators: Gaussian processes on functions
Daniel Augusto de Souza (University College London), Marc Peter Deisenroth (University College London)
Time SeriesPhysics Related
🎯 What it does: The study investigates infinite-width neural operators (NO), proving that their limit is a function-valued Gaussian process and providing the corresponding covariance function.
Infinite-Width Limit of a Single Attention Layer: Analysis via Tensor Programs
Mana Sakai (University of Tokyo), Masaaki Imaizumi (University of Tokyo)
Transformer
🎯 What it does: This paper derives the infinite-width limit distribution of a single-layer multi-head attention under the standard scaling of 1/√n, proving that it exhibits a hierarchical non-Gaussian distribution.
InfinityStar: Unified Spacetime AutoRegressive Modeling for Visual Generation
Jinlai Liu (ByteDance), Zehuan Yuan (ByteDance)
GenerationData SynthesisTransformerImageVideoTextMultimodalityBenchmark
🎯 What it does: A unified spatiotemporal autoregressive framework called InfinityStar is proposed, capable of generating high-resolution images and dynamic videos, supporting various tasks such as text-to-image, text-to-video, image-to-video, and video extrapolation.
Influence Functions for Edge Edits in Non-Convex Graph Neural Networks
Jaeseung Heo (POSTECH), Dongwoo Kim (POSTECH)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes an edge editing influence function suitable for non-convex graph neural networks, which can simultaneously estimate parameter shifts and changes in information propagation, predicting the impact of edge deletion/insertion on model evaluation metrics (over-compression, over-smoothing, validation loss);
Influence Guided Context Selection for Effective Retrieval-Augmented Generation
Jiale Deng (Shanghai Jiao Tong University), Linpeng Huang (Shanghai Jiao Tong University)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a context selection method for retrieval-augmented generation (RAG) based on Context Influence Value (CI Value), aimed at filtering noisy retrieval results and improving generation quality.
InfMasking: Unleashing Synergistic Information by Contrastive Multimodal Interactions
Liangjian Wen (Southwestern University of Finance and Economics), Jiang Duan (Southwestern University of Finance and Economics)
Representation LearningTransformerContrastive LearningMultimodality
🎯 What it does: The InfMasking method is proposed, which uses infinite masks to randomly obscure multimodal features and align unmasked representations to enhance collaborative information, forming an unsupervised multimodal representation learning framework.
Information Retrieval Induced Safety Degradation in AI Agents
Cheng Yu (Technical University of Munich), Orestis Papakyriakopoulos (Princeton University)
RetrievalSafty and PrivacyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: The study investigates the phenomenon of security degradation in retrieval-augmented AI agents, assessing the impact of retrieval on rejection rates, bias, and harmfulness.
Information Theoretic Learning for Diffusion Models with Warm Start
Yirong Shen, Cong Ling
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: A diffusion model information-theoretic learning framework suitable for arbitrary homoscedastic noise is proposed, improving the log-likelihood upper bound and bridging the gap between warm-start noise elimination training and evaluation.
Information-Computation Tradeoffs for Noiseless Linear Regression with Oblivious Contamination
Ilias Diakonikolas (University of Wisconsin Madison), Ankit Pensia (Carnegie Mellon University)
Tabular
🎯 What it does: This study investigates the noise-free linear regression problem based on Gaussian covariates in the presence of unobservable contamination, with the goal of accurately recovering the regression coefficients β.
Information-Driven Design of Imaging Systems
Henry Pinkard (University of California), Laura Waller (University of California)
OptimizationComputational EfficiencyImage
🎯 What it does: This paper proposes a method for directly evaluating and designing imaging systems using mutual information, avoiding the traditional evaluation methods that rely on decoders, and is suitable for measuring unknown objects in real systems.
Information-Theoretic Discrete Diffusion
Moongyu Jeon (Yonsei University), Albert No (Yonsei University)
Diffusion modelTextSequential
🎯 What it does: This paper proposes an information-theoretic framework for discrete diffusion models and proves two types of information-minimization relationships (I-MDSE and I-MDCE), thereby providing an exact likelihood interpretation of DSE and DCE losses.
Information-theoretic Generalization Analysis for VQ-VAEs: A Role of Latent Variables
Futoshi Futami (Osaka University), Masahiro Fujisawa (Osaka University)
GenerationAuto EncoderImage
🎯 What it does: This paper conducts a theoretical analysis of the generalization error of Vector Quantized Variational Autoencoders (VQ-VAE) using information theory methods, proposing a generalization upper bound that is independent of the decoder and explaining the role of latent variables in generation quality.
Information-Theoretic Reward Decomposition for Generalizable RLHF
Liyuan Mao (Shanghai Jiao Tong University), Chenjia Bai
Reinforcement Learning from Human FeedbackReinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes a method that uses information theory to split rewards into 'non-prompt rewards' and 'prompt-related rewards', and utilizes the difference in non-prompt rewards to prioritize the training of the reward model, thereby enhancing the generalization ability of the reward model.
Informed Correctors for Discrete Diffusion Models
Yixiu Zhao (Stanford University), Scott Linderman (Stanford University)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: A prediction-correction sampling scheme is proposed for discrete diffusion models, where the corrector is driven by the model's own information to more reliably correct cumulative errors; a hollow Transformer is constructed to efficiently implement the correction, and a new ELBO objective is proposed for training.
Informed Initialization for Bayesian Optimization and Active Learning
Carl Hvarfner (Meta), Maximilian Balandat
OptimizationHyperparameter SearchTabular
🎯 What it does: A new initialization method called HIPE is proposed for Bayesian optimization and active learning, balancing predictive uncertainty with hyperparameter learning.
Infrequent Exploration in Linear Bandits
Harin Lee (University of Washington), Min-hwan Oh (Seoul National University)
Reinforcement Learning
🎯 What it does: Proposes an INFEX framework for sparse exploration through a predetermined schedule in linear bandits, using greedy decisions in the remaining phases, and provides theoretical and experimental validation.
Injecting Frame-Event Complementary Fusion into Diffusion for Optical Flow in Challenging Scenes
Haonan Wang (Huazhong University of Science and Technology), Luxin Yan (Huazhong University of Science and Technology)
TransformerDiffusion modelOptical FlowImageVideoMultimodality
🎯 What it does: This paper proposes Diff-ABFlow, a dual-modal optical flow estimation framework that utilizes event-frame dual modalities, employing a diffusion model and integrating appearance-boundary complementary information from frames and events.
Inner Speech as Behavior Guides: Steerable Imitation of Diverse Behaviors for Human-AI coordination
Rakshit Trivedi, David C. Parkes (Harvard University)
Robotic IntelligenceTransformerLarge Language ModelDiffusion modelAuto EncoderTextMultimodality
🎯 What it does: Proposes the MIMIC framework, using 'inner speech' as a mediator to guide behavior, achieving steerable imitation learning that can be controlled by text instructions during inference.
Inpainting the Neural Picture: Inferring Unrecorded Brain Area Dynamics from Multi-Animal Datasets
Ji Xia (Columbia University), Kenneth D. Miller (Columbia University)
TransformerTime SeriesBiomedical Data
🎯 What it does: A model called NeuroPaint based on a masked Transformer has been developed to infer the neural dynamics of unrecorded brain regions in Neuropixels data from multiple animals and multiple regions.
INST-IT: Boosting Instance Understanding via Explicit Visual Prompt Instruction Tuning
Wujian Peng (Fudan University), Yu-Gang Jiang (Fudan University)
RecognitionSegmentationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageVideoMultimodalityBenchmark
🎯 What it does: In response to the shortcomings of large-scale multimodal models (LMM) in instance-level understanding, the INST-IT approach is proposed, which includes an instance-level evaluation benchmark, a large-scale instance-level instruction tuning dataset containing images and videos, and a continuous instruction fine-tuning training paradigm based on explicit visual prompts.
InstaInpaint: Instant 3D-Scene Inpainting with Masked Large Reconstruction Model
Junqi You (Shanghai Jiao Tong University), Ming-Hsuan Yang
RestorationGenerationTransformerDiffusion modelVideoPoint Cloud
🎯 What it does: The InstaInpaint framework is proposed, which completes 3D scene inpainting in 0.4 seconds using a single 2D inpainting result, and supports object deletion, insertion, and multi-region inpainting.
Instance-Dependent Regret Bounds for Nonstochastic Linear Partial Monitoring
Federico Di Gennaro (École Polytechnique Fédérale de Lausanne), Nicolò Cesa-Bianchi (Università degli Studi di Milano)
Optimization
🎯 What it does: This paper studies the non-random (adversarial) linear partial monitoring problem, proposing an algorithm based on the EXO (Exploration-Optimization) method, and provides theoretical convergence rates for two types of games: locally observable and globally observable.
Instance-Level Composed Image Retrieval
Bill Psomas (Czech Technical University in Prague), Giorgos Tolias
RetrievalVision Language ModelImageTextMultimodality
🎯 What it does: This paper first proposes an instance-level combined image retrieval (i-CIR) evaluation dataset, which includes explicit visual, textual, and combined hard negative samples; subsequently, a training-free baseline method called BASIC is designed, utilizing a pre-trained visual-language model (such as CLIP) to achieve combined retrieval of images and text queries through centralization, semantic projection, text contextualization, min-max normalization, and Harris cross-fusion.
Instance-Optimality for Private KL Distribution Estimation
Jiayuan Ye (National University of Singapore), Kunal Talwar (Apple)
Safty and PrivacyText
🎯 What it does: This study investigates the problem of estimating discrete distributions under differential privacy (DP) constraints, using KL divergence to measure error, and proposes an instance-optimal estimator based on the 'two-sampling' idea.
InstanceAssemble: Layout-Aware Image Generation via Instance Assembling Attention
Qiang Xiang (Fudan University), Junping Zhang (Fudan University)
GenerationData SynthesisTransformerDiffusion modelImageMultimodality
🎯 What it does: Proposes InstanceAssemble, a method for layout-aware image generation that achieves this through instance assembly attention, allowing precise control over bounding box positions and supporting multimodal content.
Instant Video Models: Universal Adapters for Stabilizing Image-Based Networks
Matthew Dutson (University of Wisconsin Madison), Mohit Gupta (University of Wisconsin Madison)
RestorationSegmentationDepth EstimationSupervised Fine-TuningVideo
🎯 What it does: To address the temporal inconsistency of frame-level networks and insufficient robustness to time-varying noise in video reasoning, this paper proposes a universal 'stability adapter' that achieves adaptive smoothing of features and output space through a lightweight controller without modifying the original model weights, significantly enhancing the temporal consistency and robustness of multi-task videos.
Instant4D: 4D Gaussian Splatting in Minutes
Zhanpeng Luo, Li Lu
Depth EstimationOptimizationComputational EfficiencyGaussian SplattingSimultaneous Localization and MappingVideo
🎯 What it does: A fast 4D Gaussian Splating framework named INSTANT4D is proposed, capable of reconstructing dynamic 3D scenes from uncalibrated monocular videos and achieving real-time rendering within minutes.
InstructFlow: Adaptive Symbolic Constraint-Guided Code Generation for Long-Horizon Planning
Haotian Chi (Jilin University), Haiyan Yin (Agency for Science, Technology and Research)
Robotic IntelligenceAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This paper studies a multi-agent, symbolic, feedback-driven robot code generation framework called InstructFlow, which can decompose natural language instructions into a hierarchical instruction graph, generate executable code, and perform local corrections by abstracting symbolic constraints when execution fails.
InstructHOI: Context-Aware Instruction for Multi-Modal Reasoning in Human-Object Interaction Detection
Jinguo Luo (Harbin Institute of Technology), Honghai LIU
Object DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes the InstructHOI method, which guides large models for multimodal reasoning through context-aware instructions to achieve human-object interaction detection.
InstructRestore: Region-Customized Image Restoration with Human Instructions
Shuaizheng Liu (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
RestorationDiffusion modelImage
🎯 What it does: A region-customized image restoration framework called InstructRestore is proposed, which can achieve local restoration and bokeh control based on user-specified areas and enhancement levels.
InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition
Yijie Zheng (Aerospace Information Research Institute, Chinese Academy of Sciences), Xue Yang (Shanghai Jiao Tong University)
RecognitionObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: The InstructCDS task and the EarthInstruct benchmark are proposed, along with the development of a training-free framework, InstructSAM, for instruction-driven remote sensing object counting, detection, and segmentation.
Integral Imprecise Probability Metrics
Siu Lun Chau (Nanyang Technological University), Krikamol Muandet (RI Lab CISPA Saarbrücken)
ClassificationOptimizationImageTabular
🎯 What it does: The Integral Imprecise Probability Metric (IIPM) framework is proposed to measure the differences between imprecise probabilities (such as lower bound probabilities, Belief functions, etc.), and based on this, Maximum Mean Imprecision (MMI) is defined as a new measure of epistemic uncertainty (EU).
Integrating Drug Substructures and Longitudinal Electronic Health Records for Personalized Drug Recommendation
Wenjie Du (University of Science and Technology of China), Yang Wang (Huazhong Agricultural University)
Recommendation SystemDrug DiscoveryGraph Neural NetworkTransformerBiomedical DataElectronic Health Records
🎯 What it does: The SubRec framework is proposed, which combines drug molecular substructures with patients' long-term electronic health records to achieve personalized drug recommendations.
Integration Matters for Learning PDEs with Backwards SDEs
Sungje Park (University of Southern California), Stephen Tu (University of Southern California)
BenchmarkPhysics RelatedStochastic Differential Equation
🎯 What it does: This paper studies the discretization problem of solving high-dimensional partial differential equations (PDEs) using the backward stochastic differential equation (BSDE) method. It theoretically and experimentally verifies the irreducible bias generated by the Euler-Maruyama (EM) integration scheme in first-order self-consistency loss, and subsequently proposes converting forward-backward SDEs to Stratonovich form and using stochastic Heun integration, successfully eliminating this bias.
Interaction-Centric Knowledge Infusion and Transfer for Open Vocabulary Scene Graph Generation
Lin Li (Hong Kong University of Science and Technology), Long Chen (Tencent)
Object DetectionKnowledge DistillationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: A framework for open vocabulary scene graph generation based on interactive perception (ACC) is proposed, achieving more accurate object and relationship detection through interaction-driven knowledge injection and transfer.
Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning
Yichen Li (University of Arizona), Chicheng Zhang (University of Arizona)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: This paper proposes a state-level interaction annotation-based imitation learning algorithm called STAGGER, and further designs Warm-STAGGER, which combines mixed offline trajectories and interactive state annotations. It demonstrates that under the cost measurement of state-level annotations, it can significantly outperform traditional behavior cloning.
Interactive Anomaly Detection for Articulated Objects via Motion Anticipation
Ankan Kumar Bhunia (University of Edinburgh), Hakan Bilen (University of Edinburgh)
Anomaly DetectionRobotic IntelligenceTransformerPoint Cloud
🎯 What it does: This paper proposes an interactive anomaly detection framework for movable parts, utilizing active manipulation and motion prediction to identify functional anomalies.