ICLR 2025 Papers — Page 20
International Conference on Learning Representations · 3704 papers
LLaMaFlex: Many-in-one LLMs via Generalized Pruning and Weight Sharing
Ruisi Cai (NVIDIA), Pavlo Molchanov (NVIDIA)
CompressionTransformerLarge Language ModelText
🎯 What it does: A flexible large language model system called LLAMAFLEX has been constructed, which can generate compressed models with any parameter budget in a zero-shot manner after a single training session, without the need for additional fine-tuning or distillation.
LLaRA: Supercharging Robot Learning Data for Vision-Language Policy
Xiang Li (Stony Brook University), Michael S Ryoo
Robotic IntelligenceTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Utilizing a pre-trained Vision-Language Model (VLM) and transforming it into a robot control strategy through dialog-based instruction tuning on behavior cloning data;
LLaVA-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models
Feng Li, Chunyuan Li
RecognitionGenerationDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageVideoMultimodalityBenchmark
🎯 What it does: We propose LLaVA-Interleave, a large-scale multimodal model capable of simultaneously handling multiple image, video, 3D, and single image tasks, and uniformly using an interleaved image-text format.
LLaVA-Mini: Efficient Image and Video Large Multimodal Models with One Vision Token
Shaolei Zhang (Chinese Academy of Sciences), Yang Feng (Chinese Academy of Sciences)
CompressionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageVideoMultimodality
🎯 What it does: An efficient multimodal model LLaVA-Mini has been developed, which allows for image and video understanding with only 1 visual token per image.
LLaVA-MoD: Making LLaVA Tiny via MoE-Knowledge Distillation
Fangxun Shu (Alibaba Group), Hao Jiang (Alibaba Group)
Computational EfficiencyKnowledge DistillationTransformerMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a method to transfer knowledge from large-scale multimodal language models to smaller models through knowledge distillation, and achieves model sparsification using a mixture of experts (MoE) architecture, balancing performance and efficiency.
LLM Unlearning via Loss Adjustment with Only Forget Data
Yaxuan Wang (University of California), Wei Wei (Accenture)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a 'flat' loss adjustment method called FLAT, which utilizes only forgotten data and template responses to achieve unlearning in large language models without using retained data or reference models.
LLM-based Typed Hyperresolution for Commonsense Reasoning with Knowledge Bases
Armin Toroghi (University of Toronto), Scott Sanner (University of Toronto)
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A typed hyperresolution (LLM-TH) framework based on large language models is proposed for reliable common sense reasoning in large-scale or incomplete knowledge bases.
LLM-SR: Scientific Equation Discovery via Programming with Large Language Models
Parshin Shojaee (Virginia Tech), Chandan K. Reddy (Virginia Tech)
Large Language ModelPrompt EngineeringPhysics Related
🎯 What it does: A framework for scientific equation discovery using large language models (LLM), called LLM-SR, is proposed, which combines programmatic equation representation, LLM-generated structural skeletons, parameter optimization, and iterative search with experience buffering.
LLM-wrapper: Black-Box Semantic-Aware Adaptation of Vision-Language Models for Referring Expression Comprehension
Amaia Cardiel (Valeo), Matthieu Cord (Valeo)
RecognitionObject DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: This paper proposes LLM-wrapper, a black-box adaptation method that utilizes large language models to perform natural language reasoning on the outputs of open-source visual-language models (VLMs), thereby completing the referential expression comprehension (REC) task without accessing the internal weights of the VLM.
LLMOPT: Learning to Define and Solve General Optimization Problems from Scratch
Caigao JIANG, Yang Yu (East China Normal University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A learning-based framework LLMOPT is constructed to automatically define and solve multi-type optimization problems from natural language descriptions.
LLMs Can Plan Only If We Tell Them
Bilgehan Sel (Virginia Tech), Ming Jin (Virginia Tech)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The AoT+ prompting method is proposed, which utilizes large language models to automatically generate long-term planning without external validation tools, significantly improving planning effectiveness.
LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations
Hadas Orgad (Technion), Yonatan Belinkov (Technion)
GenerationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper analyzes and predicts the errors and error types of large language models (LLMs) by probing their internal representations, and selects more reliable answers using internal information.
LoCA: Location-Aware Cosine Adaptation for Parameter-Efficient Fine-Tuning
Zhekai Du (University of Electronic Science and Technology of China), Mingming Gong (University of Melbourne)
Domain AdaptationOptimizationTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes LoCA - a frequency domain parameter-efficient fine-tuning method based on inverse discrete cosine transform (iDCT);
Local convergence of simultaneous min-max algorithms to differential equilibrium on Riemannian manifold
Sixin Zhang (Universite de Toulouse)
GenerationOptimizationGenerative Adversarial NetworkImage
🎯 What it does: This study investigates the local convergence properties of two simultaneous min-max algorithms (τ-GDA and τ-SGA) on Riemannian manifolds and applies the theory to the training of orthogonal Wasserstein GANs, exploring how to overcome oscillations caused by rotational dynamics by adjusting the learning rate ratio τ.
Local Loss Optimization in the Infinite Width: Stable Parameterization of Predictive Coding Networks and Target Propagation
Satoki Ishikawa (Science Tokyo), Ryo Karakida (National Institute of Advanced Industrial Science and Technology)
OptimizationImage
🎯 What it does: This paper studies the maximum update parameterization (µP) for two local loss optimization algorithms, prediction coding (PC) and target propagation (TP), under the condition of infinite width, and achieves the transfer of hyperparameters across networks of different widths under this parameterization.
Local Patterns Generalize Better for Novel Anomalies
Yalong Jiang (Beihang University)
Anomaly DetectionTransformerVision Language ModelVideoText
🎯 What it does: Through a two-stage image-text alignment and cross-modal attention mechanism, local patterns capable of cross-domain generalization are identified and enhanced, achieving video anomaly detection by combining temporal sentence generation and motion estimation.
Local Steps Speed Up Local GD for Heterogeneous Distributed Logistic Regression
Michael Crawshaw (George Mason University), Mingrui Liu (George Mason University)
OptimizationTabular
🎯 What it does: This study investigates the acceleration effects of Local Gradient Descent (Local GD) in distributed logistic regression, proposing a two-stage learning rate warm start and a variant of Local Gradient Flow (Local GF);
Local-Prompt: Extensible Local Prompts for Few-Shot Out-of-Distribution Detection
Fanhu Zeng (Institute of Automation, Chinese Academy of Sciences), Xu-Yao Zhang (Southern University of Science and Technology)
Anomaly DetectionPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: Proposes Local-Prompt, which enhances few-shot OOD detection performance by freezing global prompts and learning local prompts, utilizing random cropping of negative samples for augmentation and regional positive-negative contrast regularization.
Locality Alignment Improves Vision-Language Models
Ian Connick Covert, Tatsunori Hashimoto (Stanford University)
RecognitionObject DetectionSegmentationRepresentation LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Locality alignment is applied to the pre-trained ViT, and self-supervised optimization through MaskEmbed enables the model to capture local semantics of images, thereby enhancing the spatial understanding ability of the visual language model.
Locality Sensitive Avatars From Video
Chunjin Song (University of British Columbia), Helge Rhodin (University of British Columbia)
GenerationPose EstimationGraph Neural NetworkNeural Radiance FieldVideo
🎯 What it does: A local sensitive human avatar model based on NeRF is proposed, which can learn character motion from monocular videos and decouple rigid skeletal motion from local non-rigid deformation in the pose space, achieving high-fidelity rendering.
Locality-aware Gaussian Compression for Fast and High-quality Rendering
Seungjoo Shin (POSTECH), Sunghyun Cho (POSTECH)
CompressionNeural Radiance FieldPoint Cloud
🎯 What it does: LocoGS is proposed, a framework that utilizes 3D Gaussian local consistency for compression, capable of significantly reducing storage space while maintaining high rendering quality.
Locally Connected Echo State Networks for Time Series Forecasting
Filip Matzner (Charles University), František Mráz (Charles University)
Time Series
🎯 What it does: This paper proposes and implements a Local Connected Echo State Network (LCESN), which enhances the scalability and stability of traditional ESNs through local topology and enforced memory, and evaluates it on the NARMA10 and nine real-world time series datasets.
LoCoDL: Communication-Efficient Distributed Learning with Local Training and Compression
Laurent Condat (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationFederated LearningComputational EfficiencyTabular
🎯 What it does: The LoCoDL algorithm is designed and implemented for distributed/federated learning scenarios, achieving linear convergence communication efficiency by combining local training and unbiased compressors.
LocoVR: Multiuser Indoor Locomotion Dataset in Virtual Reality
Kojiro Takeyama (University of California), Misha Sra (University of California)
Data SynthesisPose EstimationConvolutional Neural NetworkMultimodalityBenchmark
🎯 What it does: The LocoVR dataset is proposed, which collected over 7,000 dual walking trajectories in 131 indoor home scenes using VR, and its effectiveness was validated in three types of tasks.
Logic-Logit: A Logic-Based Approach to Choice Modeling
Shuhan Zhang (Chinese University of Hong Kong Shenzhen), Shuang Li (Chinese University of Hong Kong Shenzhen)
Recommendation SystemOptimizationExplainability and InterpretabilityTabularBiomedical Data
🎯 What it does: An interpretable choice model based on logical rules, Logic-Logit, is proposed, which utilizes OR-of-ANDs rules to learn human decision-making patterns and predict choice behavior.
Logical Consistency of Large Language Models in Fact-Checking
Bishwamittra Ghosh (Max Planck Institute for Software Systems), Arijit Khan (Aalborg University)
TransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a method to evaluate and enhance the consistency of large language models in the knowledge graph (KG) fact-checking task for propositional logic queries (including negation, conjunction, and disjunction), and significantly improves consistency through supervised fine-tuning.
Logically Consistent Language Models via Neuro-Symbolic Integration
Diego Calanzone (University of Trento), Antonio Vergari (University of Edinburgh)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A method for fine-tuning LoCo-LLMs based on neural-symbolic reasoning is proposed, allowing large language models to maintain factuality and consistency under knowledge base facts and logical constraints.
Logicbreaks: A Framework for Understanding Subversion of Rule-based Inference
Anton Xue (University of Pennsylvania), Eric Wong (University of Pennsylvania)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A framework based on Horn logic is proposed to study the reasoning behavior of large language models (LLMs) when following prompt rules, and to theoretically and experimentally investigate how adversarial suffix (jailbreak) attacks can disrupt rule adherence.
LOIRE: LifelOng learning on Incremental data via pre-trained language model gRowth Efficiently
Xue Han (China Mobile Research Institute), Chao Deng (China Mobile Research Institute)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: A LOIRE framework has been developed, utilizing functional-preserving multidimensional model expansion and iterative distillation to achieve lifelong learning for language models.
LOKI: A Comprehensive Synthetic Data Detection Benchmark using Large Multimodal Models
Junyan Ye (Sun Yat-sen University), Weijia Li (Sun Yat-sen University)
Data SynthesisAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageVideoTextMultimodalityBenchmarkChain-of-ThoughtAudio
🎯 What it does: This paper proposes the LOKI multimodal synthetic data detection benchmark and evaluates the detection and interpretability capabilities of various large-scale multimodal models on it.
LoLCATs: On Low-Rank Linearizing of Large Language Models
Michael Zhang (Stanford University), Christopher Re (Stanford University)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The LOLCATS two-step method is proposed, which first uses learnable linear attention to approximate the softmax attention of the Transformer, and then fine-tunes with LoRA to restore model quality.
Long Context Compression with Activation Beacon
Peitian Zhang (Beijing Academy of Artificial Intelligence), Zhicheng Dou (Renmin University of China)
CompressionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the Activation Beacon plugin, which implements long context compression in Transformer-based LLMs (such as Llama-2, Qwen-2): it splits long texts into chunks and finely inserts beacon tokens, ultimately compressing the context into the activation of beacon tokens, significantly reducing KV cache and inference computation.
Long-Context Linear System Identification
Oğuz Kaan Yüksel (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
Time Series
🎯 What it does: This paper studies the identification of long-context linear systems and provides non-asymptotic convergence rates parameterized by the number of samples NT, dimension d, and context length p.
Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG
Bowen Jin (University of Illinois Urbana-Champaign), Sercan O Arik
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper explores the effect of increasing the number of retrievals in long-context LLM-assisted retrieval-augmented generation (RAG) and proposes three methods to enhance robustness.
Long-horizon Visual Instruction Generation with Logic and Attribute Self-reflection
Yucheng Suo (Zhejiang University), Yi Yang (Zhejiang University)
GenerationTransformerLarge Language ModelDiffusion modelTextMultimodality
🎯 What it does: The LIGER framework is proposed to achieve visual instruction generation for long-sequence tasks, enhancing image coherence and attribute accuracy through historical prompts, visual memory, and self-reflection mechanisms.
Long-Sequence Recommendation Models Need Decoupled Embeddings
Ningya Feng (Tsinghua University), Mingsheng Long (Tsinghua University)
Recommendation SystemSequential
🎯 What it does: This paper addresses the conflict between attention and representation learning in long sequence recommendation by proposing the DARE model, which separates attention and representation into two independent embedding tables.
Long-Short Decision Transformer: Bridging Global and Local Dependencies for Generalized Decision-Making
Jincheng Wang (University College London), Dimitrios Kanoulas (University College London)
TransformerReinforcement LearningSequentialBenchmark
🎯 What it does: Designed and implemented the Long-Short Decision Transformer (LSDT), which incorporates parallel long-term and short-term branches in the decision transformer, using self-attention to capture global dependencies and dynamic convolution to capture local Markovian features, achieving a unified general architecture across various offline RL tasks.
Long-tailed Adversarial Training with Self-Distillation
Seungju Cho (Korea Advanced Institute of Science and Technology), Changick Kim (Korea Advanced Institute of Science and Technology)
Knowledge DistillationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a self-distillation framework to enhance adversarial robustness under long-tail distribution data. It first trains a self-teacher model on a balanced subset obtained through resampling, and then performs self-distillation on the main model, significantly improving the robust accuracy of tail classes.
Long-time asymptotics of noisy SVGD outside the population limit
Victor Priser (Telecom Paris), Adil Salim (Microsoft Research)
OptimizationTabularStochastic Differential Equation
🎯 What it does: A variant of SVGD with added Langevin noise (NSVGD) is proposed, and it is proven that the limit set formed at the iteration endpoint with a finite number of particles approaches the target distribution as the number of particles increases, thus avoiding the variance collapse of traditional SVGD.
LongGenBench: Benchmarking Long-Form Generation in Long Context LLMs
Yuhao Wu (Singapore University of Technology and Design), Roy Ka-Wei Lee (Singapore University of Technology and Design)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: The LongGenBench benchmark is proposed and made public, specifically designed to evaluate the ability of large language models to generate long texts that comply with complex instructions at lengths of 16K and 32K tokens.
Longhorn: State Space Models are Amortized Online Learners
Bo Liu (Meta), qiang liu
ImageText
🎯 What it does: A new state space model called Longhorn is proposed, aimed at achieving efficient sequence mixing through an online learning framework.
LongMamba: Enhancing Mamba's Long-Context Capabilities via Training-Free Receptive Field Enlargement
Zhifan Ye (Georgia Institute of Technology), Yingyan Celine Lin (NVIDIA)
TransformerLarge Language ModelTextSequential
🎯 What it does: Improvement of the Mamba model's ability to handle long sequences in a training-independent manner.
LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory
Di Wu (University of California Los Angeles), Dong Yu (Tencent AI Lab)
RetrievalOptimizationTransformerTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the LONGMEMEVAL benchmark for the systematic evaluation of chat assistants' long-term memory capabilities.
LongPO: Long Context Self-Evolution of Large Language Models through Short-to-Long Preference Optimization
Guanzheng Chen (National University of Singapore), Lidong Bing (Shanda AI Research Institute)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The LongPO method is proposed, which evolves a short-context LLM into a long-context model through internally generated short-long comparative preference data without relying on manual long-text annotations.
LongVILA: Scaling Long-Context Visual Language Models for Long Videos
Yukang Chen, Song Han
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: The LongVILA model is proposed, establishing a five-stage training process that supports long video (up to 2048 frames) visual language understanding and achieves reasoning with a context of 1 million tokens.
LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs
Yushi Bai (Tsinghua University), Juanzi Li (Tsinghua University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmark
🎯 What it does: This study addresses the length limitation issue of long-context LLMs when generating texts exceeding 2000 words. It proposes AgentWrite, an agent-based planning and writing pipeline that automatically constructs ultra-long SFT data, and utilizes this data to train the LongWriter model, enabling it to generate high-quality texts of over 10k words. Ultimately, it builds the LongBench-Write benchmark evaluation set.
Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models
Alexandre Variengien (European Commission), Eric Winsor (UK AI Security Institute)
RetrievalTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates the internal working mechanism of language models in retrieval tasks, proposes the ORION dataset, and discovers the modularity of the request-retrieval module within the model through causal intervention (residual flow patches).
Looking Backward: Retrospective Backward Synthesis for Goal-Conditioned GFlowNets
Haoran He (Hong Kong University of Science and Technology), Ling Pan (Tsinghua University)
Data SynthesisReinforcement LearningSequential
🎯 What it does: A method called 'Retrospective Backward Synthesis (RBS)' is proposed, which synthesizes new positive reward trajectories using the backward strategy of GFlowNet under target conditions, thereby enriching the training data and addressing the sparse reward problem.
Looking Backward: Streaming Video-to-Video Translation with Feature Banks
Feng Liang (University of Texas at Austin), Diana Marculescu (University of Texas at Austin)
Image TranslationGenerationDiffusion modelVideoStochastic Differential Equation
🎯 What it does: Proposes StreamV2V, a diffusion model for real-time video-to-video translation that can handle videos of unlimited length.
Looking into User’s Long-term Interests through the Lens of Conservative Evidential Learning
Dingrong Wang (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
Recommendation SystemReinforcement LearningSequential
🎯 What it does: A new Evidential Conservative Q-Learning (ECQL) framework is proposed for dynamic recommendation systems, combining Sequence State Encoder (SSE) and Conservative Evidence Actor-Critic (CEAC) to learn a recommendation strategy that explores long-term interests while maintaining safety in an offline reinforcement learning environment.
Looking Inward: Language Models Can Learn About Themselves by Introspection
Felix Jedidja Binder (University of California San Diego), Owain Evans (Truthful AI)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the introspective ability of LLMs through self-prediction training, assessing their self-awareness of behavioral characteristics in hypothetical scenarios, and validating the introspective advantage by comparing with cross-prediction models.
Looped Transformers for Length Generalization
Ying Fan (University of Wisconsin Madison), Kangwook Lee (UC Berkeley)
TransformerSequential
🎯 What it does: This study investigates the effectiveness of the Looped Transformer in length generalization, proposing to supervise only the final answer during training without requiring intermediate steps, and achieving adaptive depth through variable-step supervision.
Loopy: Taming Audio-Driven Portrait Avatar with Long-Term Motion Dependency
Jianwen Jiang (ByteDance), Yanbo Zheng (ByteDance)
GenerationData SynthesisDiffusion modelVideoAudio
🎯 What it does: We propose Loopy, an end-to-end audio-driven portrait video generation framework that removes traditional motion constraints, capable of generating natural, stable, and detail-rich dialogue portrait videos solely from audio.
LoR-VP: Low-Rank Visual Prompting for Efficient Vision Model Adaptation
Can Jin (Rutgers University), Dimitris N. Metaxas (Rutgers University)
Domain AdaptationComputational EfficiencyPrompt EngineeringImage
🎯 What it does: A low-rank matrix multiplication visual prompt (LOR-VP) method is proposed for efficient and comprehensive task adaptation of pre-trained visual models.
LoRA Done RITE: Robust Invariant Transformation Equilibration for LoRA Optimization
Jui-Nan Yen (University of California Los Angeles), Sanjiv Kumar (Google)
OptimizationLarge Language ModelSupervised Fine-TuningText
🎯 What it does: An adaptive matrix preprocessing optimizer for LoRA fine-tuning, LoRA-RITE, is proposed, which maintains transformation invariance during the optimization process for LoRA factorization.
LoRA-Pro: Are Low-Rank Adapters Properly Optimized?
Zhengbo Wang (University of Science and Technology of China), Tieniu Tan (Institute of Automation, Chinese Academy of Sciences)
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: Proposes LoRA-Pro, which adjusts the low-rank matrix gradients of LoRA to approximate the full fine-tuning gradients, thereby narrowing the performance gap between LoRA and full fine-tuning.
LoRA-X: Bridging Foundation Models with Training-Free Cross-Model Adaptation
Farzad Farhadzadeh (Qualcomm AI Research), Fatih Porikli (Qualcomm AI Research)
GenerationDomain AdaptationTransformerDiffusion modelImageText
🎯 What it does: A LoRA-X adapter is proposed, enabling transfer across different base models without the need for original training data or additional training, and its effectiveness is validated in text-to-image generation tasks.
LoRA3D: Low-Rank Self-Calibration of 3D Geometric Foundation models
Ziqi Lu (Massachusetts Institute of Technology), Yue Wang (University of Southern California)
Pose EstimationAutonomous DrivingSupervised Fine-TuningSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Self-calibration of a pre-trained 3D geometric foundation model is performed, utilizing sparse RGB images and multi-view consistency to automatically generate pseudo-labels and fine-tune the model using LoRA, enhancing performance in the target scene.
LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual Learning
Liangzu Peng (University of Pennsylvania), Rene Vidal
ClassificationRecognitionSupervised Fine-TuningImage
🎯 What it does: A continuous learning method LoRanPAC based on pre-trained models, random ReLU features, and low-rank truncation has been designed.
Loss Landscape of Shallow ReLU-like Neural Networks: Stationary Points, Saddle Escape, and Network Embedding
Zhengqing Wu, François Gaston Ged (University of Vienna)
OptimizationTabular
🎯 What it does: This study investigates the loss landscape of single hidden layer ReLU-like networks, revealing directional stationary points and escape neurons, explaining how they lead to stagnation in gradient descent and saddle point escape.
Lossy Compression with Pretrained Diffusion Models
Jeremy Vonderfecht (Portland State University), Feng Liu (Portland State University)
CompressionDiffusion modelImageOrdinary Differential Equation
🎯 What it does: A complete implementation of the DiffC algorithm has been achieved, and it has been applied to mainstream pre-trained diffusion models such as Stable Diffusion 1.5, 2.1, XL, and Flux-dev, enabling lossless image compression without training and zero-shot.
Lotus: Diffusion-based Visual Foundation Model for High-quality Dense Prediction
Jing He (Hong Kong University of Science and Technology), Ying-Cong Chen
SegmentationGenerationDepth EstimationComputational EfficiencyDiffusion modelImage
🎯 What it does: This paper proposes a vision foundation model called Lotus based on diffusion models for dense prediction tasks (such as monocular depth and normal vector estimation), achieving zero-shot generalization by minimizing training data.
LR0.FM: Low-Res Benchmark and Improving robustness for Zero-Shot Classification in Foundation Models
Priyank Pathak, Yogesh S Rawat
ClassificationRecognitionTransformerDiffusion modelContrastive LearningImageBenchmark
🎯 What it does: Evaluate the zero-shot classification performance of visual-language foundation models on low-resolution images and propose the LR0.FM benchmark.
LucidPPN: Unambiguous Prototypical Parts Network for User-centric Interpretable Computer Vision
Mateusz Pach (Jagiellonian University), Dawid Damian Rymarczyk (Jagiellonian University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes LucidPPN, a prototype part network that separates color from shape/texture features through a two-branch structure, achieving clearer interpretability.
Lumina-T2X: Scalable Flow-based Large Diffusion Transformer for Flexible Resolution Generation
Peng Gao (Shanghai AI Laboratory), Hongsheng Li (Chinese University of Hong Kong)
GenerationData SynthesisTransformerDiffusion modelFlow-based ModelImageVideoTextMultimodalityAudio
🎯 What it does: The Lumina-T2X series model is proposed, with the core being the Flow-based Large Diffusion Transformer (Flag-DiT), achieving unified multimodal (image, video, multi-view, audio) generation and editing while maintaining stable training with large-scale parameters (up to 7B) and long context (128K tokens).
LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias
Haian Jin (Cornell University), Zexiang Xu (Adobe Research)
GenerationData SynthesisTransformerImage
🎯 What it does: Two full Transformer architectures (Encoder-Decoder and Decoder-Only) are proposed for generating new perspective images from sparse perspective images, completely eliminating traditional 3D priors and achieving scalable, general perspective synthesis.
M^3PC: Test-time Model Predictive Control using Pretrained Masked Trajectory Model
Kehan Wen (ETH Zurich), Lei Ke (Carnegie Mellon University)
OptimizationTransformerReinforcement LearningSequential
🎯 What it does: This paper proposes a method for model predictive control (MPC) using a pre-trained Masked Trajectory Transformer during the testing phase, which includes forward MPC for reward maximization and backward MPC for goal achievement, achieving action sampling, state prediction, and evaluation through various mask combinations.
MA-RLHF: Reinforcement Learning from Human Feedback with Macro Actions
Yekun Chai (Baidu Inc), Hua Wu (Baidu Inc)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The MA-RLHF framework is proposed, introducing macro actions (serialized token segments) into RLHF to address the credit assignment problem of token-level RLHF in long sequences.
MA$^2$E: Addressing Partial Observability in Multi-Agent Reinforcement Learning with Masked Auto-Encoder
Sehyeok Kang (KAIST AI), Se-Young Yun (KAIST AI)
TransformerReinforcement LearningAuto EncoderSequential
🎯 What it does: By using Masked Auto-Encoder (MAE) within the CTDE framework, multi-agent systems can infer global information solely based on their local observations, thus addressing the issue of partial observability.
Machine Unlearning Fails to Remove Data Poisoning Attacks
Martin Pawelczyk (Harvard University), Seth Neel (Google)
ClassificationAdversarial AttackData-Centric LearningConvolutional Neural NetworkTransformerImageText
🎯 What it does: Evaluated the removal effects of eight mainstream machine unlearning algorithms on three types of data poisoning attacks: target, non-target, and Gaussian noise in image classification and text classification models.
Machine Unlearning via Simulated Oracle Matching
Kristian Georgiev (Massachusetts Institute of Technology), Seth Neel (Harvard University)
ClassificationOptimizationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new machine model forgetting algorithm—Datamodel Matching (DMM), which first estimates the model's output after deleting part of the training samples using predictive data attribution (datamodelling) techniques, and then fine-tunes the model using Oracle Matching (OM) to make the final model's prediction distribution close to that of the 'oracle' model after complete retraining;
MACPO: Weak-to-Strong Alignment via Multi-Agent Contrastive Preference Optimization
Yougang Lyu (University of Amsterdam), Zhaochun Ren (Leiden University)
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningContrastive LearningText
🎯 What it does: A multi-agent comparative preference optimization (MACPO) framework is proposed, utilizing weak teachers and strong students to learn from each other during the training phase, iteratively improving the alignment performance of the strong student.
MAD-TD: Model-Augmented Data stabilizes High Update Ratio RL
Claas A Voelcker (University of Toronto), Igor Gilitschenski (University of Toronto)
Reinforcement LearningWorld ModelSequential
🎯 What it does: By utilizing a small amount of on-policy data generated by a learned world model to supplement the experience replay of TD3, the issue of value function misgeneralization to unseen actions under a high update-to-data (UTD) ratio is alleviated, stabilizing training and improving sample efficiency.
MADGEN: Mass-Spec attends to De Novo Molecular generation
Yinkai Wang (Tufts University), Soha Hassoun (Tufts University)
GenerationDrug DiscoveryGraph Neural NetworkTransformerContrastive LearningMultimodality
🎯 What it does: This paper proposes MADGEN, a two-stage framework for generating novel molecules based on chemical scaffolds: first, scaffold retrieval of MS/MS spectra is performed using contrastive learning, and then a attention-driven graph generation model conditionally generates complete molecules on the retrieved scaffolds.
MAESTRO: Masked Encoding Set Transformer with Self-Distillation
Matthew Eric Lee, Dokyoon Kim (University of Pennsylvania)
Knowledge DistillationRepresentation LearningTransformerBiomedical Data
🎯 What it does: Developed the MAESTRO self-supervised set representation learning model, which compresses the entire human immune cell population in high-dimensional cell count data into fixed-dimensional vectors. During training, it learns global and local features of the entire sample by masking a large number of cells and reconstructing their expression, ultimately predicting clinical indicators such as disease diagnosis, age, and gender.
MaestroMotif: Skill Design from Artificial Intelligence Feedback
Martin Klissarov (McGill University), Pierluca D'Oro (Mila)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Developed the MaestroMotif method, which utilizes large language models (LLM) to automatically generate reward functions, skill initiation/termination code, and skill strategies during training, thereby achieving zero-shot control based on natural language descriptions.
MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation
Zhaoning Yu (Iowa State University), Hongyang Gao (Iowa State University)
Explainability and InterpretabilityDrug DiscoveryGraph Neural NetworkAuto EncoderGraphBiomedical Data
🎯 What it does: MAGE is proposed, a model-agnostic explanation method based on molecular motifs, which constructs explanatory molecules using motifs to ensure chemical validity.
MagicDec: Breaking the Latency-Throughput Tradeoff for Long Context Generation with Speculative Decoding
Ranajoy Sadhukhan (Carnegie Mellon University), Beidi Chen (Carnegie Mellon University)
GenerationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By introducing a compressible KV cache in long-context scenarios through Speculative Decoding (MagicDec), it achieves simultaneous improvements in throughput and latency during large batch requests without compromising generation quality.
MagicPIG: LSH Sampling for Efficient LLM Generation
Zhuoming Chen (Carnegie Mellon University), Beidi Chen (Carnegie Mellon University)
GenerationRetrievalComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a sampling attention approximation method based on LSH called MAGICPIG, which addresses the KV cache bottleneck in long-context LLM generation and significantly reduces attention computation costs.
MAGNet: Motif-Agnostic Generation of Molecules from Scaffolds
Leon Hetzel (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
GenerationDrug DiscoveryGraph Neural NetworkTransformerAuto EncoderGraph
🎯 What it does: This paper proposes the idea of separating molecular structure from features, using a more general scaffold to replace traditional motifs for molecular generation, and based on this, constructs the MAGNet model.
Magnetic Preference Optimization: Achieving Last-iterate Convergence for Language Model Alignment
Mingzhi Wang (Peking University), Yaodong Yang (Peking University)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A Magnetic Preference Optimization (MPO) framework is proposed, utilizing self-play to achieve consistency with human preferences in large language models, allowing the final strategy (last-iterate) to converge to the Nash equilibrium of the original preference game.
Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing
Zhangchen Xu (University of Washington), Bill Yuchen Lin (University of Washington)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Using aligned large language models (such as Llama‑3‑Instruct) to automatically generate high-quality instructions and their responses by only providing pre-question templates, thus eliminating the need for manual annotation or complex prompt engineering, to construct large-scale aligned datasets (MAGPIE-Air and MAGPIE-Pro).
MAI: A Multi-turn Aggregation-Iteration Model for Composed Image Retrieval
Yanzhe Chen (Peking University), Yuxin Peng (University of Science and Technology Beijing)
RetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a task framework for Multi-Turn Combined Image Retrieval (MTCIR) and constructs a large-scale, historically rich FashionMT dataset; it designs the Multi-turn Aggregation-Iteration (MAI) model, which employs two-stage semantic aggregation, cyclic combination loss, and multi-turn iterative optimization to achieve efficient multimodal information aggregation and historical information compression.
Maintaining Structural Integrity in Parameter Spaces for Parameter Efficient Fine-tuning
Chongjie Si (Shanghai Jiao Tong University), Wei Shen (Shanghai Jiao Tong University)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImageTextMultimodality
🎯 What it does: A general low-rank tensor adaptation method called FLoRA is proposed for parameter-efficient fine-tuning in parameter spaces of different dimensions, while preserving the topological structure of the original space.
Make Haste Slowly: A Theory of Emergent Structured Mixed Selectivity in Feature Learning ReLU Networks
Devon Jarvis (University of the Witwatersrand), Andrew M Saxe
OptimizationRepresentation LearningConvolutional Neural NetworkTabular
🎯 What it does: This study investigates the feature learning of limited-width ReLU networks and provides a complete learning dynamics by mapping them to an analyzable Gated Deep Linear Network (ReLN); it also reveals that ReLU networks spontaneously generate structured mixed selectivity and node reuse mechanisms in multi-context tasks.
Making Text Embedders Few-Shot Learners
Chaofan Li (Beijing University of Posts and Telecommunications), Zheng Liu (Beijing Academy of Artificial Intelligence)
Representation LearningMeta LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: The ICL Embedder is proposed, which incorporates task descriptions and a variable number of examples into the embedding input, allowing the LLM to maintain zero-shot performance while gaining few-shot enhancement capabilities.
Making Transformer Decoders Better Differentiable Indexers
Wuchao Li (University of Science and Technology of China), Kun Gai (Kuaishou)
RetrievalRecommendation SystemTransformerTextSequential
🎯 What it does: A unified framework URI is proposed, integrating the retrieval model and index construction into the same Transformer decoder for end-to-end training.
MallowsPO: Fine-Tune Your LLM with Preference Dispersions
Haoxian Chen (Columbia University), Wenpin Tang (Columbia University)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The MallowsPO method is proposed, which optimizes the preference learning of LLMs through the Mallows ranking model and the divergence index, incorporating divergence as a weighting factor into DPO.
MambaExtend: A Training-Free Approach to Improve Long Context Extension of Mamba
Seyedarmin Azizi (University of Southern California), Massoud Pedram (University of Southern California)
RetrievalOptimizationTransformerLarge Language ModelText
🎯 What it does: A framework called MambaExtend is designed to extend the context length of the Mamba model by 32 times with only calibration of the scale factor and no training.
MambaPEFT: Exploring Parameter-Efficient Fine-Tuning for Mamba
Masakazu Yoshimura (Sony Group Corporation), Yota Maeda (Sony Group Corporation)
TransformerSupervised Fine-TuningImageText
🎯 What it does: A systematic exploration and evaluation of parameter-efficient fine-tuning (PEFT) for the Mamba model.
MambaQuant: Quantizing the Mamba Family with Variance Aligned Rotation Methods
Zukang Xu (Houmo AI), Sifan Zhou (Houmo AI)
ClassificationOptimizationImageVideo
🎯 What it does: A post-training quantization framework called MambaQuant is proposed for the Mamba model, addressing the issues of extreme values and uneven channel variance during the quantization process.
MamBEV: Enabling State Space Models to Learn Birds-Eye-View Representations
Hongyu Ke (Georgia State University), Yi Ding (Georgia State University)
Object DetectionAutonomous DrivingComputational EfficiencyConvolutional Neural NetworkContrastive LearningImageMultimodality
🎯 What it does: A BEV generation framework called MamBEV based on the state space model (Mamba) is proposed, utilizing linear spatiotemporal SSM to achieve a unified bird's-eye view representation, supporting multi-frame and multi-camera 3D detection and other visual perception tasks.
MamKO: Mamba-based Koopman operator for modeling and predictive control
ZHAOYANG LI, Xunyuan Yin (Nanyang Technological University)
OptimizationReinforcement LearningTime SeriesSequential
🎯 What it does: A model that uses the Mamba structure to generate time-varying Koopman operators (MamKO) is proposed, and an MPC controller is designed based on this model.
Manifold Constraint Reduces Exposure Bias in Accelerated Diffusion Sampling
Yuzhe YAO, Jingdong Wang (Baidu)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A training-free, almost hyperparameter-free manifold constraint method (MCDO) is proposed, which alleviates exposure bias in accelerated sampling of diffusion models by projecting denoised observations back to their corresponding manifolds, thereby improving sampling quality.
Manifold Induced Biases for Zero-shot and Few-shot Detection of Generated Images
Jonathan Brokman (Fujitsu Research of Europe), Guy Gilboa (Technion - Israel Institute of Technology)
Object DetectionDiffusion modelScore-based ModelImage
🎯 What it does: A zero-shot and few-shot AI-generated image detection method is proposed, utilizing the bias features of the implicit probability manifold of a pre-trained diffusion model for judgment.
Manifolds, Random Matrices and Spectral Gaps: The geometric phases of generative diffusion
Enrico Ventura (Bocconi University), Luca Ambrogioni (Radboud University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: In this paper, the author explores the potential geometric structure in the generation process through spectral analysis of the score function Jacobian of diffusion models, revealing a three-stage evolution process.
ManiSkill-HAB: A Benchmark for Low-Level Manipulation in Home Rearrangement Tasks
Arth Shukla (Hillbot Inc.), Hao Su (University of California San Diego)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningPoint CloudBenchmark
🎯 What it does: The MS-HAB benchmark is proposed, providing GPU-accelerated low-level control implementations, reinforcement learning and imitation learning baselines, automatic trajectory filtering, and large-scale demonstration data generation.
MANTRA: The Manifold Triangulations Assemblage
Rubén Ballester (Universitat de Barcelona), Bastian Rieck (Technical University of Munich)
Graph Neural NetworkMeshGraphBenchmark
🎯 What it does: A large-scale, purely high-order dataset MANTRA has been proposed and made public, containing over 43,000 surface triangulations and 250,000 three-dimensional manifold triangulations for high-order model benchmarking;
Many-Objective Multi-Solution Transport
Ziyue Li (University of Maryland), Tianyi Zhou (University of Maryland)
OptimizationFederated LearningText
🎯 What it does: The MosT framework is proposed, which simultaneously trains m models in multi-objective optimization where n≫m, allowing them to form diverse and complementary solutions on the Pareto front, thereby covering all objectives.
MAP: Low-compute Model Merging with Amortized Pareto Fronts via Quadratic Approximation
Lu Li (University of Pennsylvania), Yoshua Bengio (MILA)
OptimizationComputational EfficiencyImageBiomedical Data
🎯 What it does: This paper proposes a low-computation model merging method called MAP, which directly infers the multi-task Pareto front by performing quadratic approximation on the proportional coefficients of task vectors, achieving efficient integration of multiple models.