NeurIPS 2025 Papers — Page 27
Conference on Neural Information Processing Systems · 5275 papers
Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials
Yifan Pu (Tsinghua University), Xiu Li (Tsinghua University)
ClassificationGenerationComputational 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;
Linear Mixture Distributionally Robust Markov Decision Processes
Zhishuai Liu (Duke University), Pan Xu (Duke University)
Reinforcement LearningSequential
🎯 What it does: A Linear Mixture Distributed Robust Markov Decision Process (DRMDP) framework is proposed, along with theoretical guarantees for offline robust policy learning.
Linear Transformers Implicitly Discover Unified Numerical Algorithms
Patrick Lutz (Boston University), Venkatesh Saligrama (Boston University)
OptimizationTransformerTabular
🎯 What it does: Train a linear attention Transformer to solve the low-rank matrix completion task and extract a unified numerical algorithm EAGLE from the trained weights.
Linearization Explains Fine-Tuning in Large Language Models
Zahra Rahimi Afzal (University of Illinois Chicago), Mesrob I Ohannessian (University of Illinois Chicago)
TransformerLarge 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.
Linearly Constrained Diffusion Implicit Models
Vivek Jayaram (University of Washington), John Thickstun (Cornell University)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: A new linear constraint diffusion implicit model (CDIM) is proposed to quickly and accurately solve noisy linear inverse problems, significantly reducing the number of projection steps.
LinEAS: End-to-end Learning of Activation Steering with a Distributional Loss
Pau Rodriguez, Xavier Suau
GenerationData 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.
LinPrim: Linear Primitives for Differentiable Volumetric Rendering
Nicolas von Lützow (Technical University of Munich), Matthias Nießner (Technical University of Munich)
GenerationOptimizationNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes linear primitives based on transparent polyhedra (octahedra and tetrahedra) to achieve real-time visualization of new view synthesis using differentiable rasterization rendering, and subsequently performs scene reconstruction.
List-Level Distribution Coupling with Applications to Speculative Decoding and Lossy Compression
Joseph Rowan (University of Toronto), Ashish J Khisti
CompressionLarge 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.
Listwise Preference Diffusion Optimization for User Behavior Trajectories Prediction
Hongtao Huang (University of New South Wales), Lina Yao (CSIRO's Data61 and University of New South Wales)
Recommendation SystemOptimizationTransformerDiffusion modelTime SeriesSequential
🎯 What it does: This paper proposes the User Behavior Trajectory Prediction (UBTP) task, utilizing a diffusion model to generate future multi-step interaction sequences, and designs the Listwise Preference Diffusion Optimization (LPDO) framework;
LiteReality: Graphic-Ready 3D Scene Reconstruction from RGB-D Scans
Zhening Huang (University of Cambridge), Joan Lasenby (University of Cambridge)
Object DetectionGenerationRetrievalLarge Language ModelVision Language ModelPoint CloudMesh
🎯 What it does: LiteReality is proposed, a complete pipeline that transforms indoor RGB-D scans into graphics-ready 3D scenes that can be directly used for rendering, simulation, and AR/VR.
LittleBit: Ultra Low-Bit Quantization via Latent Factorization
Banseok Lee (Samsung Research), Young-Min Kim
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper proposes the LittleBit method, which can compress large language models to a level of 0.1 bits per weight while maintaining performance close to full precision.
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)
RecognitionGenerationOptimizationComputational 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 at Network Edge: A Layer-wise Efficient Federated Fine-tuning Approach
Jinglong Shen (Xidian University), Jiajie Xu (Xidian University)
Federated LearningComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: LEFF is proposed, a federated fine-tuning framework that selects layers and customizes according to client resources;
LLM Interpretability with Identifiable Temporal-Instantaneous Representation
Xiangchen Song (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
Explainability and InterpretabilityRepresentation LearningLarge Language ModelAuto EncoderTextTime Series
🎯 What it does: A recognizable time-instant causal representation learning framework is proposed, combining sparse autoencoders with linear dynamic models to learn the time delays and instantaneous causal relationships in the activations of large language models (LLMs);
LLM Layers Immediately Correct Each Other
Arjun Patrawala, Jacob Steinhardt
TransformerLarge 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)
ClassificationExplainability 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)
GenerationData 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 Query Scheduling with Prefix Reuse and Latency Constraints
Gregory Dexter (LinkedIn), Aman Gupta (Nubank)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the query queuing scheduling problem when utilizing the RadixAttention prefix reuse technique in online LLM inference, and proposes a new scheduling algorithm k-LPM, aimed at improving inference throughput and latency performance while satisfying strict delay constraints of TTFT (Time to First Token) and TPOT (Time per Output Token).
LLM Safety Alignment is Divergence Estimation in Disguise
Rajdeep Haldar (Purdue University), Qifan Song (Purdue University)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a unified theoretical framework that views LLM alignment methods as divergence estimation between aligned (safe) and unaligned (harmful) distributions, and based on this framework, designs new KL-based alignment methods KLDO and the generalized FDO family.
LLM Strategic Reasoning: Agentic Study through Behavioral Game Theory
Jingru Jia (University of Illinois), Deming Chen (University of Illinois)
Large Language ModelAgentic AITextTabularChain-of-Thought
🎯 What it does: Evaluate the strategic reasoning ability of LLMs in a series of abstract games, proposing an assessment framework based on behavioral game theory to quantify reasoning depth and analyze the impact of the model's thought chain and identity context on decision-making.
LLM Unlearning via Neural Activation Redirection
William F. Shen (University of Cambridge), Nicholas D. Lane (University of Cambridge)
Large Language ModelText
🎯 What it does: This paper studies a method called LUNAR for unlearning in LLMs through neural activation redirection, which can completely delete specified knowledge while maintaining model utility.
LLM-DAMVC: A Large Language Model Assisted Dynamic Agent for Multi-View Clustering
HaiMing Xu, Qianqian Wang (Xidian University)
TransformerLarge Language ModelContrastive LearningMultimodality
🎯 What it does: A framework called LLM-DAMVC is proposed, which transforms multi-view clustering into a dynamic decision-making problem, utilizing large language models to dynamically evaluate and fuse different views;
LLM-Driven Treatment Effect Estimation Under Inference Time Text Confounding
Yuchen Ma (Munich Center for Machine Learning), Stefan Feuerriegel (Munich Center for Machine Learning)
TransformerLarge 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)
TransformerLarge 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.
LLM-PySC2: Starcraft II learning environment for Large Language Models
Zongyuan Li (Nankai University), Xuebo Zhang (Nankai University)
Large Language ModelReinforcement LearningTextMultimodality
🎯 What it does: This paper proposes the LLM-PySC2 environment, enabling large language models to fully operate the action space of StarCraft II's PySC2 and support multi-agent collaboration.
LLMs Encode Harmfulness and Refusal Separately
Jiachen Zhao (Northeastern University), Weiyan Shi
TransformerLarge 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.
LMFusion: Adapting Pretrained Language Models for Multimodal Generation
Weijia Shi (University of Washington), LILI YU
GenerationTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Proposes the LMFusion framework, which achieves multimodal generation and understanding by integrating a pre-trained text LLM (Llama-3) with a parallel image Transformer and diffusion model.
Local Curvature Descent: Squeezing More Curvature out of Standard and Polyak Gradient Descent
Peter Richtárik (King Abdullah University of Science and Technology), Robin Yadav (University of British Columbia)
OptimizationTabular
🎯 What it does: Three variants of gradient descent based on local curvature information (LCD1, LCD2, LCD3) are proposed, along with convergence theory and experimental validation; the gradient descent and Polyak step size are generalized to matrix step sizes through the local curvature mapping C(x).
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)
GraphTabular
🎯 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.
Local-Global Associative Frames for Symmetry-Preserving Crystal Structure Modeling
Haowei Hua (Polytechnic University), Wanyu Lin (Polytechnic University)
Graph Neural NetworkTransformerTabularBenchmarkPhysics Related
🎯 What it does: This paper proposes a local-global associative symmetry-preserving framework (SPFrame) for crystal structure prediction, ensuring that the model remains rotation-invariant while preserving the symmetry of the crystal.
Local-Global Coupling Spiking Graph Transformer for Brain Disorders Diagnosis from Two Perspectives
Geng Zhang (Xi'an Jiaotong University), Badong Chen (Xi'an Jiaotong University)
ClassificationSpiking Neural NetworkTransformerBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A Local-Global Coupling Spiking Graph Transformer (LGC‑SGT) is proposed, which jointly models the firing rates of neuronal populations in brain regions and the differences in functional connectivity to diagnose brain diseases (ASD, MDD, schizophrenia) and uncover dual-perspective biomarkers.
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
Convolutional 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.
Locality in Image Diffusion Models Emerges from Data Statistics
Artem Lukoianov (Massachusetts Institute of Technology), Vincent Sitzmann (Massachusetts Institute of Technology)
RestorationGenerationTransformerDiffusion modelImage
🎯 What it does: The locality of image diffusion models is determined by the pixel correlation of the training data, rather than the prior of the network structure; a new analytical denoiser based on the high signal-to-noise ratio principal components (Wiener filter) is proposed, and the locality can be controlled by modifying the data statistics.
Localized Data Shapley: Accelerating Valuation for Nearest Neighbor Algorithms
Guangyi Zhang (Shenzhen Technology University), Wei Wang (HONG KONG UNIVERSITY OF SCIENCE AND TECHNOLOGY)
ClassificationComputational 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.
Localizing Knowledge in Diffusion Transformers
Arman Zarei (University of Maryland), Soheil Feizi (University of Maryland)
GenerationOptimizationComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: This study investigates the localization methods of different types of knowledge in Diffusion Transformers (DiT) and validates their effectiveness on models such as PixArtα, FLUX, and SANA.
Locally Optimal Private Sampling: Beyond the Global Minimax
Hrad Ghoukasian (McMaster University), Shahab Asoodeh (McMaster University)
OptimizationSafty 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.
LocDiff: Identifying Locations on Earth by Diffusing in the Hilbert Space
Zhangyu Wang (University of Maine), Gengchen Mai (University of Texas at Austin)
RecognitionRetrievalDiffusion modelImage
🎯 What it does: This paper proposes a multi-scale Latent Diffusion model called LocDiff, based on a Spherical Harmonics Dirac Delta encoding-decoding framework, for image geolocation.
LODGE: Level-of-Detail Large-Scale Gaussian Splatting with Efficient Rendering
Jonas Kulhanek (Czech Technical University in Prague), Federico Tombari (Google DeepMind)
Computational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: Proposes Level of Detail (LOD) and chunk caching techniques to achieve real-time rendering of large-scale 3D Gaussian Splatting scenes on mobile devices.
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)
Object 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
Large 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)
Drug 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.
LogicTree: Improving Complex Reasoning of LLMs via Instantiated Multi-step Synthetic Logical Data
Zehao Wang (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
TransformerLarge Language ModelText
🎯 What it does: Proposes the LogicTree framework, which generates complex logic trees through backward reasoning and contextualizes instantiation using a two-stage LLM to produce multi-step natural language reasoning data.
LOMIA: Label-Only Membership Inference Attacks against Pre-trained Large Vision-Language Models
Yihao LIU, Bin Xiao (Hong Kong Polytechnic University)
Adversarial AttackTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A member inference attack framework LOMIA for pre-trained visual-language models (VLLM) is proposed, which only allows querying the model's top-level predictions (label-only), and three attack variants are implemented: Text-Text Feature Attack (TTFA), Image-Text Feature Attack (ITFA), and Dual Feature Attack (DUFA).
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)
SegmentationNeural 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)
ClassificationRecognitionContrastive LearningImage
🎯 What it does: A two-stage learning framework for information retention is proposed for long-tail classification;
Long-tailed Recognition with Model Rebalancing
Jiaan Luo (Shanghai Jiao Tong University), Jiangchao Yao (Shanghai Jiao Tong University)
ClassificationRecognitionContrastive LearningImage
🎯 What it does: A MOdel REbalancing (MORE) framework is proposed, which directly reallocates the model space through low-rank parameter decomposition to enhance the performance of tail categories.
Longer Context, Deeper Thinking: Uncovering the Role of Long-Context Ability in Reasoning
Van Yang, Xiaotian Han (Case Western Reserve University)
TransformerLarge 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.
LongMagpie: A Self-synthesis Method for Generating Large-scale Long-context Instructions
Chaochen Gao (Institute of Information Engineering Chinese Academy of Sciences), Songlin Hu
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study proposes LongMagpie, a self-synthesis method that automatically generates document-related questions using aligned large language models, given only the document and user tags, and combines model answers to generate high-quality large-scale long-context instruction data.
LongVPO: From Anchored Cues to Self-Reasoning for Long-Form Video Preference Optimization
Zhenpeng Huang (Nanjing University), Limin Wang (Nanjing University)
Recommendation 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.
Look Before You Leap: A GUI-Critic-R1 Model for Pre-Operative Error Diagnosis in GUI Automation
Yuyang Wanyan (Chinese Academy of Sciences), Changsheng Xu (Chinese Academy of Sciences)
OptimizationTransformerLarge Language ModelReinforcement LearningMultimodalityChain-of-Thought
🎯 What it does: A pre-operation critic GUI, named GUI-Critic-R1, has been designed to diagnose and provide error correction suggestions before the execution of GUI automation.
Look-Ahead Reasoning on Learning Platforms
Haiqing Zhu (Australian National University), Celestine Mendler-Dünner (ELLIS Institute)
Recommendation SystemOptimizationTabularFinance Related
🎯 What it does: This paper proposes and analyzes the 'look-ahead reasoning' framework, studying how users predict and influence future model updates on learning platforms; it examines the effects of two strategies, level-k thinking and collective reasoning, on learning dynamics, convergence speed, and final equilibrium; corresponding theoretical theorems, upper bounds, and simulation validations are provided.
Lookahead Routing for Large Language Models
Canbin Huang (Sun Yat-sen University), Xiaojun Quan (Shenzhen Loop Area Institute)
GenerationOptimizationTransformerLarge 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
Object 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)
RestorationDepth 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
RecognitionComputational 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.
LOPT: Learning Optimal Pigovian Tax in Sequential Social Dilemmas
Yun Hua (Shanghai Jiao Tong University), Hongyuan Zha (Chinese University of Hong Kong)
OptimizationReinforcement LearningSequential
🎯 What it does: Proposes a method to internalize externalities in multi-agent reinforcement learning using Pigovian taxes to address social dilemma problems.
Loquetier: A Virtualized Multi-LoRA Framework for Unified LLM Fine-tuning and Serving
Yuchen Zhang (Nanjing University), Jingwei Xu (Nanjing University)
TransformerLarge 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.
LoRA vs Full Fine-tuning: An Illusion of Equivalence
Reece S Shuttleworth, Pratyusha Sharma (Massachusetts Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the different update methods of LoRA and full fine-tuning in large language models, comparing their spectral structures and forgetting behaviors.
LoRA-EnVar: Parameter-Efficient Hybrid Ensemble Variational Assimilation for Weather Forecasting
Yi Xiao (Tsinghua University), LEI BAI
Auto EncoderTime Series
🎯 What it does: A LoRA-EnVar hybrid ensemble variational data assimilation framework is proposed, which learns the climate background error distribution through VAE and adapts flow-dependent errors online using the LoRA module.
LoRAShop: Training-Free Multi-Concept Image Generation and Editing with Rectified Flow Transformers
Yusuf Dalva (Virginia Tech), Pinar Yanardag (Virginia Tech)
GenerationData SynthesisTransformerFlow-based ModelRectified FlowImage
🎯 What it does: This paper proposes LoRAShop, which enables multi-concept image generation and editing using a pre-trained LoRA module without training, segmentation, or auxiliary guidance.
LoRASuite: Efficient LoRA Adaptation Across Large Language Model Upgrades
Yanan Li (Beijing University of Posts and Telecommunications), Mengwei Xu
TransformerLarge 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)
Object TrackingTransformerVideo
🎯 What it does: Proposes LoRATv2, a Transformer framework for multi-frame visual object tracking.
LORE: Lagrangian-Optimized Robust Embeddings for Visual Encoders
Borna khodabandeh, Seyed-Mohsen Moosavi-Dezfooli (Apple)
OptimizationRepresentation LearningAdversarial AttackConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes an unsupervised adversarial fine-tuning framework called LORE, which enhances the robustness of visual encoders through constrained optimization while maintaining clean data performance.
Lorentz Local Canonicalization: How to make any Network Lorentz-Equivariant
Jonas Spinner (Heidelberg University), Fred A. Hamprecht (Heidelberg University)
Graph 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)
Safty 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.
LoSplit: Loss-Guided Dynamic Split for Training-Time Defense Against Graph Backdoor Attacks
Di Jin (Tianjin University), Zhen Wang (Northwestern Polytechnical University)
Graph Neural NetworkGraph
🎯 What it does: During the training phase of graph neural networks, the LoSplit framework is proposed, which dynamically partitions target nodes using early loss drift and eliminates backdoor effects through a Decoupling-Forgetting strategy.
Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation
François Rozet (Polymathic AI), Shirley Ho (Princeton University)
GenerationOptimizationComputational 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)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Constructed and validated the BAPO model, analyzing the internal information flow limitations of LLMs.
LoTA-QAF: Lossless Ternary Adaptation for Quantization-Aware Fine-Tuning
Junyu Chen (Southwestern University of Finance and Economics), Xuming Hu (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes LoTA-QAF, a quantization-aware fine-tuning method that achieves lossless ternary adaptation on quantized LLMs;
Low Precision Streaming PCA
Sanjoy Dasgupta (University of California San Diego), Purnamrita Sarkar (University of Texas at Austin)
Optimization
🎯 What it does: This paper studies the impact of low-precision quantization (both linear and logarithmic) on the Oja algorithm in streaming principal component analysis (PCA), and provides corresponding lower and upper bounds on the error.
Low Rank Gradients and Where to Find Them
Rishi Sonthalia (Boston College), Guido Montufar
Tabular
🎯 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-degree evidence for computational transition of recovery rate in stochastic block model
Jingqiu Ding, David Steurer (ETH Zürich)
Graph
🎯 What it does: This paper provides a rigorous lower bound on the computational difficulty of weak recovery in symmetric random block models at the Kesten–Stigum threshold using low-degree polynomial methods.
Low-Rank Graphon Learning for Networks
Xinyuan Fan (Tsinghua University), Weichi Wu (Tsinghua University)
Graph Neural NetworkGraph
🎯 What it does: A low-rank graphical learning framework is proposed, jointly estimating the low-rank connectivity probability matrix and graphical function;
Low-Rank Head Avatar Personalization with Registers
Sai Tanmay Reddy Chakkera (Stony Brook University), Dimitris Samaras (Stony Brook University)
GenerationMeta LearningConvolutional Neural NetworkSupervised Fine-TuningVideo
🎯 What it does: This paper proposes a method that utilizes Low-Rank Adaptation (LoRA) combined with a 3D registration module to personalize the general head avatar generation model, enhancing the reproduction of high-frequency facial details (such as wrinkles and tattoos) for individuals.
LT-Soups: Bridging Head and Tail Classes via Subsampled Model Soups
Masih Aminbeidokhti (École de technologie supérieure), Marco Pedersoli (École de technologie supérieure)
ClassificationTransformerSupervised 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.
Lua-LLM: Learning Unstructured-Sparsity Allocation for Large Language Models
Mingge Lu (University of Science and Technology of China), Guangzhong Sun (University of Science and Technology of China)
TransformerLarge Language ModelText
🎯 What it does: A gradient-based global sparsification framework called Lua-LLM is proposed, which learns the unstructured sparse allocation of LLMs.
Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement
Derong Kong (National University of Defense Technology), Jingyuan Xia (National University of Defense Technology)
RestorationGenerationDiffusion 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)
Computational 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.
LuxDiT: Lighting Estimation with Video Diffusion Transformer
Ruofan Liang (NVIDIA), Zian Wang (University of Toronto)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageVideo
🎯 What it does: This paper proposes LuxDiT, which utilizes video diffusion transformers to generate high dynamic range environment maps from a single image or video for light estimation.
LVLM-Driven Attribute-Aware Modeling for Visible-Infrared Person Re-Identification
Zhiqi Pang (Harbin Institute of Technology), Chunyu Wang (Harbin Institute of Technology)
RecognitionRetrievalTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes an unsupervised method called LVLM-AAM for improving visible-infrared person re-identification (VI-ReID) by extracting attribute information using a large visual-language model (LVLM).
Lyapunov-Stable Adaptive Control for Multimodal Concept Drift
Tianyu Pan, Damon L. Woodard
ClassificationOptimizationExplainability 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)
GenerationData 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.
Machine Unlearning under Overparameterization
Jacob L. Block (University of Texas Austin), Sanjay Shakkottai (University of Texas Austin)
OptimizationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: In response to the problem of machine unlearning in over-parameterized models, a new definition of unlearning (minimal complexity interpolator) is proposed, along with a set of algorithmic frameworks centered on gradient orthogonal constraints, which can achieve effective 'forgetting' without complete retraining.
Machine Unlearning via Task Simplex Arithmetic
Junhao Dong (Nanyang Technological University), Piotr Koniusz (Data61 CSIRO)
OptimizationKnowledge DistillationData-Centric LearningTransformerVision Language ModelImageMultimodality
🎯 What it does: A reversible model forgetting method based on simple task x is proposed, achieving efficient and training-free VLM forgetting through countless function-level integrations on the multi-dimensional simplex of task vectors.
macOSWorld: A Multilingual Interactive Benchmark for GUI Agents
Pei Yang (National University of Singapore), Mike Zheng Shou (National University of Singapore)
Safty and PrivacyAgentic AITextBenchmark
🎯 What it does: This paper presents macOSWorld, a multilingual interactive GUI benchmark for macOS, which includes 202 tasks, 30 applications, 5 languages, and incorporates a security subset for evaluating deception attacks.
MACS: Multi-Agent Reinforcement Learning for Optimization of Crystal Structures
Elena Zamaraeva (University of Liverpool), Matthew Rosseinsky
OptimizationReinforcement 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)
ClassificationRecognitionAnomaly 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.)
GenerationComputational 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)
GenerationRecommendation 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.
MAGNET: A Multi-agent Framework for Finding Audio-Visual Needles by Reasoning over Multi-Video Haystacks
Sanjoy Chowdhury (University of Maryland), Dinesh Manocha (University of Maryland)
RetrievalTransformerLarge Language ModelAgentic AIVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationAudio
🎯 What it does: A new multi-video audio-visual retrieval and reasoning task, AVHaystacksQA, is proposed, along with the construction of a benchmark dataset AVHaystacks containing 3,100 QA pairs. Subsequently, a multi-agent retrieval-enhanced generation framework, MAGNET, is designed to locate and integrate key information within massive video collections, generating structured, temporally aligned answers.
MaintainCoder: Maintainable Code Generation Under Dynamic Requirements
Zhengren Wang (Peking University), Wentao Zhang (Peking University)
GenerationAI 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)
Large 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;
Make Information Diffusion Explainable: LLM-based Causal Framework for Diffusion Prediction
Wenbo Shang (Hong Kong Baptist University), Xin Huang (Hong Kong Baptist University)
Explainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: This paper proposes a causal framework MILD based on large language models to explicitly infer the 'who influences whom' relationship in information diffusion and construct a diffusion influence graph, thereby enhancing the interpretability and accuracy of diffusion predictions.
Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness–Generalization Perspective
Ming Gu (Zhejiang University), Jiajun Bu (Zhejiang University)
Graph 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.
MALinZero: Efficient Low-Dimensional Search for Mastering Complex Multi-Agent Planning
Sizhe Tang (George Washington University), Tian Lan (George Washington University)
OptimizationReinforcement LearningAgentic AITabularBenchmark
🎯 What it does: This paper proposes MALinZero, a method that utilizes low-dimensional representations and improves MCTS in multi-agent planning.
Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation
Szymon Plotka, Arkadiusz Sitek (Massachusetts General Hospital)
SegmentationMixture of ExpertsImageBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
🎯 What it does: A 3D medical image segmentation framework called Mamba-HoME, which combines Hierarchical Soft Mixture-of-Experts (HoME) with Mamba, is proposed to efficiently capture spatial hierarchical relationships from local to global.
Mamba Modulation: On the Length Generalization of Mamba Models
Peng Lu (Université de Montréal), Yufei Cui (Noah's Ark Lab)
RetrievalOptimizationTransformerLarge 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)
RecognitionObject 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)
Meta 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.
MANGO: Multimodal Attention-based Normalizing Flow Approach to Fusion Learning
Thanh-Dat Truong (University of Arkansas), Khoa Luu (University of Arkansas)
Image TranslationSegmentationFlow-based ModelAuto EncoderImageMultimodality
🎯 What it does: A multi-modal fusion framework MANGO based on Invertible Cross Attention (ICA) is proposed, which captures high-order correlations across modalities using explicit invertible normalization flows.