arXivSub Start free trial

NeurIPS 2025 Papers with Code — Page 12

Conference on Neural Information Processing Systems · 2283 papers

Learning to Integrate Diffusion ODEs by Averaging the Derivatives

Wenze Liu (MMLab, Chinese University of Hong Kong), Xiangyu Yue (MMLab, Chinese University of Hong Kong)

CodeGenerationData SynthesisTransformerDiffusion modelImageOrdinary Differential Equation

🎯 What it does: A diffusion ODE integration method based on secant losses is proposed, utilizing the secant function to learn the trajectory from noise to the target image for fast inference.

Learning to Learn with Contrastive Meta-Objective

Shiguang Wu (Tsinghua University), Quanming Yao (Tsinghua University)

CodeMeta LearningContrastive LearningImage

🎯 What it does: This paper proposes a method called Contrastive Meta Learning (ConML) that incorporates contrastive meta objectives into meta-learning, enhancing the alignment and discrimination capabilities of meta-learners through task-level contrastive learning, and can be seamlessly integrated into various meta-learning algorithms and contextual learning frameworks.

Learning to Rank for In-Context Example Retrieval

Yuwen Ji (Zhejiang University), Yue Zhang (Westlake University)

CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: A retrieval method based on learning to rank, SeDPO, is proposed, which uses LLM to generate probabilistic comparisons to train the retriever, thereby improving the example retrieval performance in In-Context Learning (ICL).

Learning to Route: Per-Sample Adaptive Routing for Multimodal Multitask Prediction

Marzieh Ajirak (Weill Cornell Medicine), Logan Grosenick (Weill Cornell Medicine)

CodeOptimizationExplainability and InterpretabilityMixture of ExpertsTextMultimodalityTabularElectronic Health Records

🎯 What it does: This paper proposes a unified adaptive routing framework for multimodal multitask prediction, which can dynamically select the modality processing path and task sharing strategy based on the input features of each sample;

Learning to Steer: Input-dependent Steering for Multimodal LLMs

Jayneel Parekh (Sorbonne Université), Matthieu Cord (Valeo.ai)

CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringImageTextMultimodality

🎯 What it does: This paper proposes a learnable input-dependent model-driven (L2S) approach to guide the behavior of multimodal large language models, enabling the model to automatically adjust its output under different inputs, primarily for safety constraints and reducing hallucinations.

Learning to Watermark: A Selective Watermarking Framework for Large Language Models via Multi-Objective Optimization

Chenrui Wang (Harbin Institute of Technology), Jing Li (Harbin Institute of Technology)

CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: The paper proposes the 'Learning to Watermark' (LTW) framework, which achieves selective watermark insertion in large language model texts through training a lightweight selection network, balancing detection rate and text quality.

Learning Urban Climate Dynamics via Physics-Guided Urban Surface–Atmosphere Interactions

Jiyang Xia (University of Manchester), Zhonghua Zheng (University of Manchester)

CodeTransformerTime SeriesPhysics Related

🎯 What it does: This paper proposes UCformer, a physics-guided Transformer for multi-task prediction of urban 2-meter temperature, specific humidity, and dew point temperature, simulating urban surface-atmosphere interactions.

Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage RL

Songjun Tu (Institute of Automation Chinese Academy of Sciences), Dongbin Zhao (Institute of Automation Chinese Academy of Sciences)

CodeComputational EfficiencyReinforcement LearningPrompt EngineeringText

🎯 What it does: Proposes AutoThink, a multi-stage reinforcement learning framework that learns to adaptively decide whether to perform explicit reasoning based on task difficulty in R1-style large models.

Learning with Calibration: Exploring Test-Time Computing of Spatio-Temporal Forecasting

Wei Chen (Hong Kong University of Science and Technology), Yuxuan Liang (Hong Kong University of Science and Technology)

CodeTime SeriesSequentialBenchmark

🎯 What it does: A testing moment computation framework ST-TTC is proposed to compensate for non-stationary errors in spatiotemporal sequence prediction through calibration during the inference phase.

Learning Without Augmenting: Unsupervised Time Series Representation Learning via Frame Projections

Berken Utku Demirel (ETH Zurich), Christian Holz (ETH Zurich)

CodeRepresentation LearningConvolutional Neural NetworkContrastive LearningTime Series

🎯 What it does: A framework for unsupervised time series representation learning is proposed, which generates different views for instance contrastive learning using orthogonal bases and overcomplete frames without manual data augmentation.

Least squares variational inference

Yvann Le Fay (ENSAE), Simon Barthelmé (GIPSA-Lab)

CodeOptimizationTabular

🎯 What it does: A new gradient-free, least squares-based variational inference method called LSVI is proposed to approximate the target distribution within the exponential family.

LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation

Huanlin Gao (China Unicom), Shiguo Lian (China Unicom)

CodeGenerationOptimizationComputational EfficiencyDiffusion modelVideo

🎯 What it does: A training-free, globally optimized caching framework called LeMiCa is proposed to accelerate video generation with diffusion models. This framework models cache scheduling as a directed acyclic graph (DAG) path search problem and controls the worst-case error caused by caching through global error metrics and lexicographic minimax optimization, significantly improving inference speed while maintaining video quality.

Less Greedy Equivalence Search

Adiba Ejaz (Columbia University), Elias Bareinboim (Columbia University)

CodeGraph

🎯 What it does: Developed Less Greedy Equivalence Search (LGES) and I-ORIENT algorithms to improve the greedy search of traditional GES, enhancing the efficiency and accuracy of causal structure learning under observational and interventional data.

Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior

Yulin Li (Harbin Institute of Technology), Zhuotao Tian (Shenzhen Loop Area Institute)

CodeCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideo

🎯 What it does: A training-agnostic dynamic video token compression method called DyToK is proposed, which utilizes the internal attention mechanism of VLLM to automatically generate keyframe priors and dynamically allocate the number of tokens based on frame importance, thereby significantly improving inference efficiency while maintaining video understanding capabilities.

Less is More: Improving LLM Alignment via Preference Data Selection

Xun Deng (University of Science and Technology of China), Xiangnan He

CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: A Bayesian aggregation strategy based on marginal maximization (BeeS) is proposed to select high-quality preference samples during Direct Preference Optimization (DPO) training, thereby enhancing the alignment effect of large language models and significantly reducing training costs.

Less is More: Local Intrinsic Dimensions of Contextual Language Models

Benjamin Matthias Ruppik (Heinrich Heine University Düsseldorf), Milica Gasic

CodeRepresentation LearningTransformerLarge Language ModelText

🎯 What it does: This paper studies the geometric structure of internal representations in large-scale language models (LLMs), using the local intrinsic dimension (LID) of context embeddings as an unsupervised diagnostic metric to analyze processes such as model training, fine-tuning, grokking, overfitting, and generalization.

Less is More: Unlocking Specialization of Time Series Foundation Models via Structured Pruning

Lifan Zhao (Shanghai Jiao Tong University), Jiaji Deng (Alibaba Group)

CodeTransformerTime Series

🎯 What it does: Proposed the 'prune first, fine-tune later' paradigm, which focuses on the sub-networks important for specific tasks in time series foundational models (TSFMs) through structured pruning, thereby enhancing the overall sample prediction performance.

Lessons Learned: A Multi-Agent Framework for Code LLMs to Learn and Improve

Yuanzhe Liu (Rensselaer Polytechnic Institute), Jie Chen (IBM Research)

CodeOptimizationAI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark

🎯 What it does: The LessonL framework is proposed and implemented, allowing multiple large language models (LLMs) to collaboratively improve code in code optimization and code generation tasks by exchanging 'lessons' with each other.

Let a Neural Network be Your Invariant

Mirco Giacobbe (University of Birmingham), Michael Tautschnig (Amazon Web Services)

CodeOptimizationSafty and PrivacyTabular

🎯 What it does: A neural network-based model checking method is proposed, capable of simultaneously verifying safety and liveness properties, and generating formal proofs through a neural ranking function.

Let LRMs Break Free from Overthinking via Self-Braking Tuning

Haoran Zhao (Tianjin University), Yueting Zhuang (Zhejiang University)

CodeTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: A Self-Braking Tuning (SBT) framework is proposed, which trains large-scale reasoning models to autonomously identify and terminate excessive reasoning steps, significantly reducing token consumption while maintaining accuracy.

Let Me Think! A Long Chain of Thought Can Be Worth Exponentially Many Short Ones

Parsa Mirtaheri (University of California San Diego), Enric Boix-Adserà (University of Pennsylvania)

CodeComputational EfficiencyTransformerLarge Language ModelReinforcement LearningGraphChain-of-Thought

🎯 What it does: This paper systematically studies the performance differences in reasoning computation between sequential expansion (long chain thinking) and parallel expansion (multiple short chain thinking) through the design of a graph connectivity task.

Let's Revise Step-by-Step: A Unified Local Search Framework for Code Generation with LLMs

Zhiyi Lyu (Nanyang Technological University), Bo An (Nanyang Technological University)

CodeGenerationOptimizationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringContrastive LearningText

🎯 What it does: The ReLoc framework is proposed, unifying the code generation problem as a local search, achieving iterative code revision through four core components (initial draft, neighborhood generation, candidate evaluation, and master update), and can be instantiated as Hill-Climbing and Genetic-Algorithm;

Leveraging Conditional Dependence for Efficient World Model Denoising

Shaowei Zhang (Nanjing University), De-Chuan Zhan (Nanjing University)

CodeAutonomous DrivingRobotic IntelligenceReinforcement Learning from Human FeedbackRecurrent Neural NetworkReinforcement LearningAuto EncoderWorld ModelVideo

🎯 What it does: A recursive state space model based on collision structure (CsRSSM) and an extended Dreamer framework (CsDreamer) are proposed for denoising observations and extracting task-relevant information in environments with noise interference.

Leveraging Depth and Language for Open-Vocabulary Domain-Generalized Semantic Segmentation

Siyu Chen (Jimei University), Meiliu Wu (University of Glasgow)

CodeSegmentationDomain AdaptationAutonomous DrivingTransformerPrompt EngineeringImageMultimodality

🎯 What it does: A single-stage framework named Vireo is proposed to achieve the joint task of open-vocabulary semantic segmentation (OVSS) and domain generalization semantic segmentation (DGSS) without accessing target domain samples or new category annotations—open-vocabulary domain generalization semantic segmentation (OV-DGSS).

Leveraging semantic similarity for experimentation with AI-generated treatments

Lei Shi (University of California), Avi Feller (University of California)

CodeOptimizationRepresentation LearningTransformerLarge Language ModelTextTabular

🎯 What it does: This paper proposes a dual-core representation learning framework for estimating causal effects and achieving adaptive allocation in online experiments, targeting the variable content generated by large language models.

LeVo: High-Quality Song Generation with Multi-Preference Alignment

Shun Lei (Shenzhen International Graduate School, Tsinghua University), Dong Yu (Shenzhen International Graduate School, Tsinghua University)

CodeGenerationTransformerReinforcement LearningTextAudio

🎯 What it does: Developed the LeVo framework, which implements parallel prediction of mixed tokens and dual-track tokens for language models, generating high-quality songs and supporting text or audio prompt control.

Lifelong Safety Alignment for Language Models

Haoyu Wang (Tsinghua University), Tianyu Pang (Sea AI Lab)

CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: A lifelong safety alignment framework is proposed, allowing large language models to continuously learn and resist the ever-evolving jailbreak attacks.

Lifelong Test-Time Adaptation via Online Learning in Tracked Low-Dimensional Subspace

Dexin Duan (Shanghai Jiao Tong University), Fei Wen (Shanghai Jiao Tong University)

CodeDomain AdaptationContrastive LearningImage

🎯 What it does: A lifelong testing-time adaptation (LCoTTA) method based on gradient low-dimensional subspaces is proposed to continuously update the model in continuously evolving target domains without the need for source data or labels.

Light-Weight Diffusion Multiplier and Uncertainty Quantification for Fourier Neural Operators

Albert Matveev (PhysicsX), Michalis Michaelides (PhysicsX)

CodeDiffusion modelTabularBenchmark

🎯 What it does: A lightweight Fourier Neural Operator (DINOZAUR) based on thermal kernels is proposed, significantly reducing the number of parameters while maintaining prediction accuracy and supporting Bayesian uncertainty quantification.

LightFair: Towards an Efficient Alternative for Fair T2I Diffusion via Debiasing Pre-trained Text Encoders

Boyu Han (Institute of Computing Technology, Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)

CodeGenerationData SynthesisComputational EfficiencyTransformerSupervised Fine-TuningDiffusion modelImageText

🎯 What it does: This paper proposes a lightweight debiasing method called LightFair, which reduces gender and racial bias in text-to-image diffusion models by fine-tuning a pre-trained text encoder and combining two-stage text-guided sampling.

Limited Preference Data? Learning Better Reward Model with Latent Space Synthesis

Leitian Tao (University of Wisconsin Madison), Sharon Li (University of Wisconsin Madison)

CodeData SynthesisRecommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelAuto EncoderText

🎯 What it does: This paper studies a framework called LENS that synthesizes preference data in the latent space of LLMs using VAE, training a reward model with the generated embedding pairs.

Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials

Yifan Pu (Tsinghua University), Xiu Li (Tsinghua University)

CodeClassificationGenerationComputational EfficiencyTransformerDiffusion modelContrastive LearningImage

🎯 What it does: Proposed and implemented Visual-Contrast Attention (VCA) to replace traditional Multi-Head Self-Attention (MHSA), achieving linear complexity contrast attention in visual Transformers;

Linearization Explains Fine-Tuning in Large Language Models

Zahra Rahimi Afzal (University of Illinois Chicago), Mesrob I Ohannessian (University of Illinois Chicago)

CodeTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: This paper studies and demonstrates that the regularization-induced approximate linearization can effectively explain the fine-tuning process by mapping Parameter-Efficient Fine-Tuning (PEFT) to the Neural Tangent Kernel (NTK) regression framework, and utilizes NTK spectrum to predict fine-tuning performance.

LinEAS: End-to-end Learning of Activation Steering with a Distributional Loss

Pau Rodriguez, Xavier Suau

CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: This paper proposes and implements LinEAS, a method for end-to-end learning of lightweight activation interventions across all layers through global distribution loss, capable of fine-tuning generative models under low data, unpaired, and no reward model conditions.

List-Level Distribution Coupling with Applications to Speculative Decoding and Lossy Compression

Joseph Rowan (University of Toronto), Ashish J Khisti

CodeCompressionLarge Language ModelImageText

🎯 What it does: A Gumbel-max based list sampling (GLS) method is proposed to maximize the matching probability when one party generates multiple samples and the other party generates a single sample in a communication-constrained scenario.

LiveStar: Live Streaming Assistant for Real-World Online Video Understanding

Zhenyu Yang (Institute of Automation, Chinese Academy of Sciences), Changsheng Xu (Institute of Automation, Chinese Academy of Sciences)

CodeRecognitionGenerationOptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextMultimodality

🎯 What it does: LiveStar is proposed, an online understanding assistant for real-time video streams that can adaptively generate responses in continuous streaming and determine the best output timing.

LLM Layers Immediately Correct Each Other

Arjun Patrawala, Jacob Steinhardt

CodeTransformerLarge Language ModelText

🎯 What it does: Investigate and describe the correction mechanism between Transformer layers, TLCM, revealing that adjacent layers partially offset the contributions of the previous layer.

LLM Meeting Decision Trees on Tabular Data

Hangting Ye (Jilin University), Yi Chang (Jilin University)

CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelTabular

🎯 What it does: The DeLTa method is proposed, utilizing LLM to generate improved decision tree rules and enhancing table data prediction through an error correction network, without the need to serialize the table or fine-tune the LLM.

LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation

Subhojyoti Khastagir (Indian Institute of Technology), Niloy Ganguly (Indian Institute of Technology)

CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelText

🎯 What it does: This paper proposes a hybrid framework CrysLLMGen, which first uses a fine-tuned LLM to generate the atomic types, coordinates, and lattice parameters of crystals, and then refines the coordinates and lattice using a pre-trained equivariant diffusion model to obtain crystals that are effective in both structure and composition.

LLM-Driven Treatment Effect Estimation Under Inference Time Text Confounding

Yuchen Ma (Munich Center for Machine Learning), Stefan Feuerriegel (Munich Center for Machine Learning)

CodeTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataElectronic Health Records

🎯 What it does: A clinical causal inference framework, called TCA, is proposed, which can only obtain text descriptions during inference and cannot access complete confounding variables.

LLM-Explorer: A Plug-in Reinforcement Learning Policy Exploration Enhancement Driven by Large Language Models

Qianyue Hao (Tsinghua University), Yong Li (Tsinghua University)

CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText

🎯 What it does: This paper proposes a plugin module called LLM-Explorer, which dynamically generates task-specific exploration strategies using large language models to improve policy exploration in reinforcement learning.

LLMs Encode Harmfulness and Refusal Separately

Jiachen Zhao (Northeastern University), Weiyan Shi

CodeTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: The study found that LLMs encode two independent concepts of harmfulness and refusal in their hidden layers, and proposed a method to regulate model behavior by extracting the harmfulness direction at the end of the instruction token.

Local Learning for Covariate Selection in Nonparametric Causal Effect Estimation with Latent Variables

Zheng Li (Beijing Technology and Business University), Zhi Geng (Beijing Technology and Business University)

CodeGraphTabular

🎯 What it does: A local learning-based covariate selection algorithm LSAS is proposed for estimating non-parametric causal effects in the presence of latent variables.

Localist Topographic Expert Routing: A Barrel Cortex-Inspired Modular Network for Sensorimotor Processing

Tianfang Zhu (Wuhan National Laboratory for Optoelectronics Huazhong University of Science and Technology), Anan LI

CodeConvolutional Neural NetworkGraph Neural NetworkMultimodality

🎯 What it does: A modular network with a local expert routing module that mimics the rodent barrel cortex is proposed for handling tactile perception tasks.

Localized Data Shapley: Accelerating Valuation for Nearest Neighbor Algorithms

Guangyi Zhang (Shenzhen Technology University), Wei Wang (HONG KONG UNIVERSITY OF SCIENCE AND TECHNOLOGY)

CodeClassificationComputational EfficiencyData-Centric LearningImage

🎯 What it does: A framework is proposed to accelerate the computation of data Shapley values for KNN models using the local structure of data space, along with an efficient algorithm for threshold KNN.

Locally Optimal Private Sampling: Beyond the Global Minimax

Hrad Ghoukasian (McMaster University), Shahab Asoodeh (McMaster University)

CodeOptimizationSafty and PrivacyTabular

🎯 What it does: This paper addresses the sampling problem under local differential privacy. Given a reference distribution P0, it defines the neighborhood of adjacent distributions N_γ(P0) and solves for the local minimization risk within that neighborhood. It also extends the global optimal sampler to the functional LDP framework, obtaining a closed-form optimal sampler, and further constructs a point-to-point superior nonlinear sampler under pure LDP.

Logic-in-Frames: Dynamic Keyframe Search via Visual Semantic-Logical Verification for Long Video Understanding

Weiyu Guo (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)

CodeObject DetectionRetrievalComputational EfficiencyLarge Language ModelVision Language ModelVideoBenchmark

🎯 What it does: A visual semantic-logic search framework (VSLS) is proposed for efficiently selecting key frames related to queries in long videos.

Logic.py: Bridging the Gap between LLMs and Constraint Solvers

Pascal Kesseli, Ricardo Silveira Cabral

CodeLarge Language ModelAgentic AITextBenchmark

🎯 What it does: This paper proposes a search problem-solving method that combines large language models (LLM) with constraint solvers. It first allows the LLM to convert natural language logic puzzles into a form that can be processed by the solver using a custom domain-specific language, Logic.py, and then the constraint solver obtains the answer.

Logical Expressiveness of Graph Neural Networks with Hierarchical Node Individualization

Arie Soeteman (Institute for Logic Language and Computation University of Amsterdam), Balder ten Cate (Institute for Logic Language and Computation University of Amsterdam)

CodeDrug DiscoveryGraph Neural NetworkGraph

🎯 What it does: This paper proposes and studies a hierarchical Ego Graph Neural Network (HE-GNN) and its subgraph-limited version (HES-GNN). It characterizes their expressive power through logical methods and conducts experimental validation on tasks such as molecular property prediction and strong regular graph isomorphism classification.

LoMix: Learnable Weighted Multi-Scale Logits Mixing for Medical Image Segmentation

Md Mostafijur Rahman (University of Texas at Austin), Radu Marculescu (University of Texas at Austin)

CodeSegmentationNeural Architecture SearchImageBiomedical DataMagnetic Resonance ImagingUltrasound

🎯 What it does: This paper proposes LoMix, a learnable weighted multi-scale logits mixing method for U-shaped networks;

Long-Tailed Recognition via Information-Preservable Two-Stage Learning

Fudong Lin (University of Delaware), Xu Yuan (University of Delaware)

CodeClassificationRecognitionContrastive LearningImage

🎯 What it does: A two-stage learning framework for information retention is proposed for long-tail classification;

Longer Context, Deeper Thinking: Uncovering the Role of Long-Context Ability in Reasoning

Van Yang, Xiaotian Han (Case Western Reserve University)

CodeTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: This paper verifies through systematic experiments that placing the enhancement of long-context capabilities after the preliminary steps of the reasoning model can significantly improve the accuracy of various reasoning tasks, especially performing better in long-context and multi-model scenarios.

LongVPO: From Anchored Cues to Self-Reasoning for Long-Form Video Preference Optimization

Zhenpeng Huang (Nanjing University), Limin Wang (Nanjing University)

CodeRecommendation SystemOptimizationTransformerLarge Language ModelVision Language ModelVideoText

🎯 What it does: A two-stage Direct Preference Optimization (DPO) framework called LongVPO is proposed, which synthesizes preference triplets from short video data to extend short-context VLM to long video understanding.

Lookahead Routing for Large Language Models

Canbin Huang (Sun Yat-sen University), Xiaojun Quan (Shenzhen Loop Area Institute)

CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: The Lookahead framework is proposed for routing decisions in multi-model LLM systems. This framework makes routing decisions by predicting the potential representations of responses that each candidate model may generate, rather than simply classifying based on the input query.

Looking Beyond the Known: Towards a Data Discovery Guided Open-World Object Detection

Anay Majee (University of Texas at Dallas), Rishabh K Iyer

CodeObject DetectionRepresentation LearningImageBenchmark

🎯 What it does: This paper proposes the CROWD framework, which combines data discovery and representation learning to address the issues of unknown class discovery and model forgetting in open-world object detection.

Looking Into the Water by Unsupervised Learning of the Surface Shape

Ori Lifschitz (University of Haifa), Dan Rosenbaum (University of Haifa)

CodeRestorationDepth EstimationNeural Radiance FieldImage

🎯 What it does: This paper proposes an unsupervised learning method that simultaneously estimates water surface height and reconstructs undistorted images using two SIREN neural field networks, addressing the refraction distortion problem when observing underwater objects from the air.

LookWhere? Efficient Visual Recognition by Learning Where to Look and What to See from Self-Supervision

Anthony Fuller, Evan Shelhamer

CodeRecognitionComputational EfficiencyKnowledge DistillationTransformerContrastive LearningImage

🎯 What it does: Designed and implemented the selector-extractor architecture and the what-where distillation method, utilizing self-supervised models to learn 'where to look' and 'what to look at', thereby achieving efficient visual recognition without fully processing high-resolution inputs.

Loquetier: A Virtualized Multi-LoRA Framework for Unified LLM Fine-tuning and Serving

Yuchen Zhang (Nanjing University), Jingwei Xu (Nanjing University)

CodeTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: Proposes the Loquetier framework, which unifies efficient fine-tuning and inference of LoRA parameters, and implements virtualization and dynamic loading of multiple adapters in a single runtime.

LoRASuite: Efficient LoRA Adaptation Across Large Language Model Upgrades

Yanan Li (Beijing University of Posts and Telecommunications), Mengwei Xu

CodeTransformerLarge Language ModelText

🎯 What it does: This study investigates efficient methods for transferring LoRA weights during the upgrade of large language models and proposes the LoRASuite framework.

LoRATv2: Enabling Low-Cost Temporal Modeling in One-Stream Trackers

Liting Lin (Pengcheng Laboratory), Haibin Ling (Westlake University)

CodeObject TrackingTransformerVideo

🎯 What it does: Proposes LoRATv2, a Transformer framework for multi-frame visual object tracking.

Lorentz Local Canonicalization: How to make any Network Lorentz-Equivariant

Jonas Spinner (Heidelberg University), Fred A. Hamprecht (Heidelberg University)

CodeGraph Neural NetworkTransformerGraphPhysics Related

🎯 What it does: A general framework LLoCa is proposed to convert any non-Lorentz invariant network into a Lorentz invariant network.

LoRO: Real-Time on-Device Secure Inference for LLMs via TEE-Based Low Rank Obfuscation

Gaojian Xiong (Beihang University), Jianwei Liu (Beihang University)

CodeSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: This paper proposes LoRO, a low-rank masking and factor reuse scheme based on TEE for secure edge inference of LLMs, preventing model theft.

Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation

François Rozet (Polymathic AI), Shirley Ho (Princeton University)

CodeGenerationOptimizationComputational EfficiencyTransformerDiffusion modelAuto EncoderTime SeriesPhysics RelatedOrdinary Differential Equation

🎯 What it does: In physical simulation, the study investigates the transfer of diffusion models from pixel space to the latent space of autoencoders for dynamic system simulation, significantly improving inference speed while maintaining or even enhancing accuracy.

Lost in Transmission: When and Why LLMs Fail to Reason Globally

Tobias Schnabel (Microsoft Research), Jennifer Neville (Microsoft Research)

CodeTransformerLarge Language ModelTextChain-of-Thought

🎯 What it does: Constructed and validated the BAPO model, analyzing the internal information flow limitations of LLMs.

Low Rank Gradients and Where to Find Them

Rishi Sonthalia (Boston College), Guido Montufar

CodeTabular

🎯 What it does: This paper studies the low-rank structure of gradients in two-layer neural networks with training data that has spike covariance, deriving the theory that gradients approximate rank two and analyzing the effects of activation functions and regularization on gradient components.

Low-Rank Graphon Learning for Networks

Xinyuan Fan (Tsinghua University), Weichi Wu (Tsinghua University)

CodeGraph Neural NetworkGraph

🎯 What it does: A low-rank graphical learning framework is proposed, jointly estimating the low-rank connectivity probability matrix and graphical function;

LT-Soups: Bridging Head and Tail Classes via Subsampled Model Soups

Masih Aminbeidokhti (École de technologie supérieure), Marco Pedersoli (École de technologie supérieure)

CodeClassificationTransformerSupervised Fine-TuningImage

🎯 What it does: A two-stage model fusion framework called LT-Soups is proposed for long-tail distributions, which first fine-tunes multiple models on different sub-samples (with varying imbalance ratios) and averages them, and then fine-tunes only the classification head on the complete data to restore the performance of the head class.

Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement

Derong Kong (National University of Defense Technology), Jingyuan Xia (National University of Defense Technology)

CodeRestorationGenerationDiffusion modelGenerative Adversarial NetworkImage

🎯 What it does: This paper proposes an unsupervised hierarchical learning framework called LASQ, which reconstructs low-light image enhancement using statistical sampling methods.

LUNA: Efficient and Topology-Agnostic Foundation Model for EEG Signal Analysis

Berkay Döner (Integrated Systems Laboratory, ETH Zürich), Yawei Li (Integrated Systems Laboratory, ETH Zürich)

CodeComputational EfficiencyRepresentation LearningTransformerTime SeriesBiomedical Data

🎯 What it does: The LUNA model is proposed to address the issue of topological heterogeneity caused by different electrode layouts in EEG, achieving self-supervised pre-training and efficient inference.

Lyapunov-Stable Adaptive Control for Multimodal Concept Drift

Tianyu Pan, Damon L. Woodard

CodeClassificationOptimizationExplainability and InterpretabilityMultimodality

🎯 What it does: An online adaptive control framework based on Lyapunov stability, LS-OGD, is proposed to address the problem of concept drift in multimodal learning.

Machine Unlearning in 3D Generation: A Perspective-Coherent Acceleration Framework

Shixuan Wang (National University of Singapore), Xinchao Wang (National University of Singapore)

CodeGenerationData SynthesisComputational EfficiencyDiffusion modelPoint CloudMesh

🎯 What it does: This paper explores the machine unlearning technology of 3D generative models for the first time, achieving selective forgetting for specific objects or attributes.

MACS: Multi-Agent Reinforcement Learning for Optimization of Crystal Structures

Elena Zamaraeva (University of Liverpool), Matthew Rosseinsky

CodeOptimizationReinforcement Learning

🎯 What it does: A multi-agent reinforcement learning method called MACS is proposed to efficiently optimize the geometric shape of periodic crystal structures.

MAESTRO : Adaptive Sparse Attention and Robust Learning for Multimodal Dynamic Time Series

Payal Mohapatra (Northwestern University), Qi Zhu (Northwestern University)

CodeClassificationRecognitionAnomaly DetectionComputational EfficiencyMixture of ExpertsMultimodalityTime SeriesBiomedical Data

🎯 What it does: Developed the MAESTRO framework for multimodal time series learning, capable of automatically adapting to any missing perceptual modalities;

MagCache: Fast Video Generation with Magnitude-Aware Cache

Zehong Ma (Peking University), Qi Tian (Huawei Inc.)

CodeGenerationComputational EfficiencyTransformerDiffusion modelVideo

🎯 What it does: Proposes MagCache, an adaptive caching strategy based on the residual amplitude ratio, to accelerate the inference of video diffusion models.

Magical: Medical Lay Language Generation via Semantic Invariance and Layperson-tailored Adaptation

Weibin Liao (Peking University), Liantao Ma (Peking University)

CodeGenerationRecommendation SystemTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBiomedical Data

🎯 What it does: This study investigates the limitations of using LoRA for medical-oriented language generation for non-professionals (MLLG) in the context of multi-source heterogeneous data and proposes the Magical framework.

MaintainCoder: Maintainable Code Generation Under Dynamic Requirements

Zhengren Wang (Peking University), Wentao Zhang (Peking University)

CodeGenerationAI Code AssistantLarge Language ModelAuto EncoderTextBenchmark

🎯 What it does: This paper proposes a system called MaintainCoder for generating maintainable code and a corresponding dynamic evaluation benchmark called MaintainBench, aimed at assessing and enhancing the maintainability of code during the process of requirement evolution.

Majority of the Bests: Improving Best-of-N via Bootstrapping

Amin Rakhsha (University of Toronto), Amir Khasahmadi (Autodesk)

CodeLarge Language ModelReinforcement LearningText

🎯 What it does: Proposes the Majority-of-the-Bests (MoB) method, which uses bootstrap to estimate the output distribution of Best-of-N and selects the mode to improve answer selection;

Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness–Generalization Perspective

Ming Gu (Zhejiang University), Jiajun Bu (Zhejiang University)

CodeGraph Neural NetworkGraph

🎯 What it does: This paper raises the universality issue of classical GNNs under different levels of homophily and discovers the smoothness-generalization dilemma through theoretical analysis. It then designs and implements the Inceptive Graph Neural Network (IGNN) framework, utilizing three main principles (SN, IN, NR) to address this dilemma, achieving multi-hop generalization and adaptive smoothness.

Mamba Modulation: On the Length Generalization of Mamba Models

Peng Lu (Université de Montréal), Yufei Cui (Noah's Ark Lab)

CodeRetrievalOptimizationTransformerLarge Language ModelText

🎯 What it does: This paper analyzes the spectral distribution of the state transition matrix of the Mamba structured state space model (SSM) and proposes enhancing the model's generalization ability beyond the training context length by adjusting the spectral scale of matrix A. Using the calibration framework of MambaExtend, learnable scaling factors are applied to A or the discrete step size ∆ at each layer, and evaluations are conducted on various long-context benchmark tasks.

Mamba Only Glances Once (MOGO): A Lightweight Framework for Efficient Video Action Detection

Yunqing Liu (Fujitsu R&D Center), Zhiming Tan (Fujitsu Research Japan)

CodeRecognitionObject DetectionComputational EfficiencyTransformerVideo

🎯 What it does: A lightweight video action detection framework called MOGO is proposed, which can perform keyframe action detection using only a single layer Mamba decoder and achieves efficient information retention through a video token construction mechanism.

MaNGO — Adaptable Graph Network Simulators via Meta-Learning

Philipp Dahlinger (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)

CodeMeta LearningGraph Neural NetworkAuto EncoderGraphPhysics Related

🎯 What it does: A graph network simulator that quickly adapts to unknown physical parameters using meta-learning can generate accurate dynamic trajectories with only a few context samples without retraining.

Manipulating 3D Molecules in a Fixed-Dimensional E(3)-Equivariant Latent Space

Zitao Chen (Tsinghua University), Yanyan Lan (Tsinghua University)

CodeOptimizationDrug DiscoveryGraph Neural NetworkFlow-based ModelAuto EncoderGraph

🎯 What it does: We propose MolFLAE, a VAE that can map 3D molecules to a fixed-dimensional, E(3) invariant latent space, and utilizes a Bayesian Flow Network for decoding, achieving zero-shot molecular editing and optimization.

Manipulating Feature Visualizations with Gradient Slingshots

Dilyara Bareeva (Fraunhofer Heinrich Hertz Institute), Kirill Bykov

CodeAdversarial AttackImageOrdinary Differential Equation

🎯 What it does: This paper studies an attack method called Gradient Slingshots, which can manipulate the feature visualization results of deep neural networks to create arbitrary target images without changing the network structure or significantly degrading performance.

Many LLMs Are More Utilitarian Than One

Anita Keshmirian (Forward College), Lav R. Varshney (University of Illinois at Urbana-Champaign)

CodeTransformerLarge Language ModelText

🎯 What it does: This study investigates whether large language models (LLMs) exhibit 'utilitarian enhancement' during moral reasoning in multi-agent (two or three person) discussions, comparing it to individual reasoning.

Many Minds, One Goal: Time Series Forecasting via Sub-task Specialization and Inter-agent Cooperation

Qihe Huang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)

CodeGraph Neural NetworkTransformerAgentic AITime SeriesFinance Related

🎯 What it does: A multi-agent time series forecasting framework called MAFS is proposed, utilizing multi-agent collaboration to address forecasting tasks with different time scales and signal characteristics.

MAPLE: Multi-scale Attribute-enhanced Prompt Learning for Few-shot Whole Slide Image Classification

Junjie Zhou (Nanjing University of Aeronautics and Astronautics), Daoqiang Zhang (Nanjing University of Aeronautics and Astronautics)

CodeClassificationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage

🎯 What it does: A multi-scale attribute-enhanced prompt learning framework called MAPLE is proposed for few-shot whole slide image classification.

MARS: A Malignity-Aware Backdoor Defense in Federated Learning

Wei Wan (City University of Macau), Leo Yu Zhang (Griffith University)

CodeFederated LearningImage

🎯 What it does: A federated learning backdoor defense framework named MARS is proposed, which identifies and filters backdoored models by measuring the backdoor energy of neurons.

Mask Image Watermarking

Runyi Hu (Nanyang Technological University), Tianwei Zhang (Nanyang Technological University)

CodeImage TranslationRestorationData SynthesisConvolutional Neural NetworkAuto EncoderImage

🎯 What it does: The MaskWM framework is proposed, supporting full-image and local watermark embedding, localization, and extraction, and achieving multi-watermark embedding.

MASTER: Enhancing Large Language Model via Multi-Agent Simulated Teaching

Liang Yue (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)

CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText

🎯 What it does: A MASTER multi-agent simulation teaching framework is proposed, utilizing interactions between teacher and student agents in three teaching scenarios: error correction, debate, and analogy reasoning, to generate a high-quality BOOST-QA instruction dataset, which is then used to fine-tune large language models.

MATCH: Multi-faceted Adaptive Topo-Consistency for Semi-Supervised Histopathology Segmentation

Meilong Xu (Stony Brook University), Chao Chen (Stony Brook University)

CodeSegmentationContrastive LearningImageBiomedical Data

🎯 What it does: A semi-supervised histopathological image segmentation framework is proposed, utilizing topological consistency in multiple perturbation predictions to enhance the topological accuracy of segmentation results.

MaxSup: Overcoming Representation Collapse in Label Smoothing

Yuxuan Zhou (University of Mannheim), Margret Keuper (CISPA Helmholtz Center for Information Security)

CodeClassificationRepresentation LearningConvolutional Neural NetworkSupervised Fine-TuningImage

🎯 What it does: Proposes the MaxSup method to address the issues of overconfidence in incorrect predictions and feature clustering compression in Label Smoothing, by penalizing the highest Logit instead.

MDNS: Masked Diffusion Neural Sampler via Stochastic Optimal Control

Yuchen Zhu (Georgia Institute of Technology), Molei Tao (FAIR at Meta)

CodeOptimizationDiffusion modelGraphPhysics Related

🎯 What it does: A discrete neural sampler MDNS based on mask diffusion and stochastic optimal control has been developed for efficient sampling of high-dimensional multimodal target distributions.

MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification

Yingying Feng (Northeastern University), Jiayi Ji (National University of Singapore)

CodeRecognitionRetrievalTransformerImageMultimodality

🎯 What it does: A framework for image-level object re-identification across arbitrary modalities, MDReID, is proposed, which can handle both modality-matching and modality-mismatching retrieval tasks simultaneously.

Measure gradients, not activations! Enhancing neuronal activity in deep reinforcement learning

Jiashun Liu (Hong Kong University of Science and Technology), Ling Pan (Hong Kong University of Science and Technology)

CodeReinforcement LearningDiffusion modelSequential

🎯 What it does: This paper studies the problem of neuron activity decay in deep reinforcement learning, proposing a gradient magnitude-based neuron activity metric called GraMa, and implementing a neuron reset method called ReGraMa based on this metric;

Measure-Theoretic Anti-Causal Representation Learning

Arman Behnam (Illinois Institute of Technology), Binghui Wang (Illinois Institute of Technology)

CodeRepresentation LearningBiomedical Data

🎯 What it does: This paper proposes a measure-theory-based counterfactual representation learning framework called ACIA, which learns how observations are generated from labels at a low level and removes environmental noise at a high level.

Measuring and Guiding Monosemanticity

Ruben Härle (TU Darmstadt), Kristian Kersting (TU Darmstadt)

CodeSafty and PrivacyRepresentation LearningAuto EncoderText

🎯 What it does: This paper proposes the Feature Monosemanticity Score (FMS) to measure the unambiguity of latent representations, and based on this, introduces Guided Sparse Autoencoders (G‑SAE), which enforce feature unambiguity and local decoupling in the latent space during training through conditional loss. The detection and guidance capabilities of this method are then evaluated in tasks such as toxicity detection, Shakespearean writing style identification, and privacy attribute recognition.

MeCeFO: Enhancing LLM Training Robustness via Fault-Tolerant Optimization

Rizhen Hu (Peking University), Kun Yuan (Peking University)

CodeOptimizationTransformerLarge Language ModelText

🎯 What it does: A fault-tolerant optimization algorithm for large language model training, MeCeFO, is proposed, which enables rapid and low-overhead recovery during node failures while maintaining training stability.

Mechanistic Interpretability of RNNs emulating Hidden Markov Models

Elia Torre (Institute of Neuroinformatics, University of Zurich and ETH Zurich), Valerio Mante (Institute of Neuroinformatics, University of Zurich and ETH Zurich)

CodeExplainability and InterpretabilityRecurrent Neural NetworkSequential

🎯 What it does: Train standard RNNs to fit three types of HMMs (linear chain, fully connected, cyclic), and then reveal their internal implementations through reverse engineering: noise-maintained orbital dynamics, slow noise integration groups, and transfer mechanisms of fast kick-neuron interactions;

MEGADance: Mixture-of-Experts Architecture for Genre-Aware 3D Dance Generation

kaixing yang, Hongyan Liu (Tsinghua University)

CodeGenerationTransformerMixture of ExpertsVideoMultimodality

🎯 What it does: This paper proposes a music-driven 3D dance generation framework based on Mixture-of-Experts called MEGADance, which is divided into two stages: high-fidelity dance quantization (using FSQ and motion constraints) and music-based dance generation (utilizing MoE and a Mamba-Transformer hybrid backbone).

MemEIC: A Step Toward Continual and Compositional Knowledge Editing

Jin Seong (Electronics and Telecommunications Research Institute), Namhoon Lee (POSTECH)

CodeTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation

🎯 What it does: Proposes the MemEIC framework and the CCKEB benchmark for achieving continuous and combinatorial knowledge editing in large audiovisual language models.