Conference on Neural Information Processing Systems · 2283 papers
Memory Decoder: A Pretrained, Plug-and-Play Memory for Large Language Models
Jiaqi Cao (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
CodeRetrievalDomain AdaptationTransformerLarge Language ModelTextBiomedical DataFinance RelatedRetrieval-Augmented Generation
🎯 What it does: A pre-trained, pluggable Memory Decoder is proposed, which achieves non-parametric domain adaptation for large language models by learning to mimic the distribution of non-parametric retrievers within a small Transformer decoder.
🎯 What it does: This work extends the Memory Mosaics architecture to the 8B scale (Memory Mosaics v2) and conducts pre-training and fine-tuning on real-world data of 1 trillion tokens, validating its capabilities in new task learning, knowledge storage, and contextual learning.
🎯 What it does: This paper proposes the Memory-Augmented Potential Field Theory (MAPFT) and implements a Memory-Augmented MPPI (MA-MPPI) controller based on this theory to achieve online experiential learning and adaptive optimization in highly non-convex state spaces.
🎯 What it does: Designed and implemented a scale-aware KV cache compression framework called ScaleKV for visual autoregressive models, significantly reducing KV cache usage while maintaining pixel-level fidelity.
MemSim: A Bayesian Simulator for Evaluating Memory of LLM-based Personal Assistants
Zeyu Zhang (Renmin University of China), Ji-Rong Wen (Renmin University of China)
CodeTransformerLarge Language ModelText
🎯 What it does: A Bayesian simulator named MemSim has been developed to automatically generate reliable, scalable, and diverse user messages and question-answer datasets, thereby objectively assessing the memory capabilities of personal assistants based on large language models (LLMs).
Mesh Interpolation Graph Network for Dynamic and Spatially Irregular Global Weather Forecasting
Zinan Zheng (Hong Kong University of Science and Technology), Jia Li (Hong Kong University of Science and Technology)
CodeGraph Neural NetworkTabularTime Series
🎯 What it does: Designed and implemented the Mesh Interpolation Graph Network (MIGN) for global dynamic and spatially irregular weather station predictions.
MESS+: Dynamically Learned Inference-Time LLM Routing in Model Zoos with Service Level Guarantees
Herbert Woisetschläger (Technical University of Munich), Hans Arno Jacobsen
CodeOptimizationTransformerLarge Language ModelText
🎯 What it does: Proposes MESS+, a dynamic learning inference LLM routing algorithm that ensures minimum SLA compliance while minimizing operational costs.
Meta Guidance: Incorporating Inductive Biases into Deep Time Series Imputers
Jiacheng You (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
CodeAnomaly DetectionData-Centric LearningTransformerTime Series
🎯 What it does: By designing learnable Non-Stationary Guidance (NSG) and Periodic Guidance (PG) matrices, and automatically assigning weights to each sequence using Meta Guidance, the performance of deep missing value imputation models is enhanced.
🎯 What it does: We propose BraInCoRL, a Transformer-based meta-learning framework that can generate individualized voxel-level encoding models without fine-tuning, given only a small number of image-brain response examples for target individuals.
MetaDefense: Defending Fine-tuning based Jailbreak Attack Before and During Generation
Weisen Jiang (Chinese University of Hong Kong), Sinno Jialin Pan (Chinese University of Hong Kong)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes MetaDefense, a two-stage defense framework that detects whether LLM responses are harmful before and after generation, preventing finetuning-based jailbreak attacks;
MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems
Xuanming Zhang (University of Wisconsin-Madison), Sharon Li (University of Wisconsin-Madison)
CodeTransformerLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Proposes the MetaMind multi-agent framework, which achieves human social thinking modeling through a three-stage collaboration of theoretical mind, moral constraints, and response validation.
Metis: A Foundation Speech Generation Model with Masked Generative Pre-training
Yuancheng Wang (Chinese University of Hong Kong), Zhizheng Wu (Chinese University of Hong Kong)
CodeGenerationTransformerMultimodalityAudio
🎯 What it does: Metis is proposed, a unified speech generation framework based on large-scale unsupervised speech pre-training and fine-tuning, which performs excellently across various tasks such as zero-shot, voice conversion, target speaker extraction, speech enhancement, and lip reading generation.
🎯 What it does: A new unbiased microcanonical Markov chain sampler (MAMS) is proposed and implemented, achieving high-dimensional unbiased sampling competitive with NUTS by incorporating Metropolis-Hastings steps into microcanonical dynamics.
MF-LLM: Simulating Population Decision Dynamics via a Mean-Field Large Language Model Framework
Qirui Mi (Chinese Academy of Sciences), Jun Wang (University College London)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextTime SeriesFinance Related
🎯 What it does: By combining mean field theory with large language models (LLM), a scalable simulation of population-level decision dynamics has been achieved.
🎯 What it does: A selective update mechanism based on momentum-gradient consistency (MGUP) is proposed, which allows for using a larger learning rate for a portion of the parameters at each step while applying a smaller learning rate to the remaining parameters, thus achieving efficient and stable training.
MI-TRQR: Mutual Information-Based Temporal Redundancy Quantification and Reduction for Energy-Efficient Spiking Neural Networks
Dengfeng Xue (Xidian University), Zhetao Li (Jinan University)
CodeSpiking Neural NetworkImageTime Series
🎯 What it does: A parameter-free, pluggable MI-TRQR module is proposed to quantify and eliminate redundant synapses across time steps in SNNs, thereby reducing energy consumption and improving accuracy.
MIBP-Cert: Certified Training against Data Perturbations with Mixed-Integer Bilinear Programs
Tobias Lorenz (CISPA Helmholtz Center for Information Security), Mario Fritz (CISPA Helmholtz Center for Information Security)
CodeOptimizationTabular
🎯 What it does: A training-time robustness certification method based on Mixed Integer Bilinear Programming (MIBP), called MIBP-Cert, is proposed, which can accurately compute the upper and lower bounds of model parameters under constrained data perturbations and provide provable robustness.
MigGPT: Harnessing Large Language Models for Automated Migration of Out-of-Tree Linux Kernel Patches Across Versions
Pucheng Dang (Institute of Computing Technology, Chinese Academy of Sciences), Xing Hu (Institute of Computing Technology, Chinese Academy of Sciences)
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the MIGGPT framework for the automatic migration of offline patches (out-of-tree patches) in the Linux kernel, significantly reducing manual maintenance costs.
Mind the Quote: Enabling Quotation-Aware Dialogue in LLMs via Plug-and-Play Modules
Yueqi Zhang (Beijing Institute of Technology), Kan Li (Xiaohongshu Inc)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes a span-conditioned generation task that allows LLMs to accurately locate and use referenced text in conversations, and builds an automated data generation and validation pipeline, releasing a benchmark that includes five sub-scenes.
🎯 What it does: A method is proposed to decompose the features of a pre-trained diffusion model into semantic and visual sub-features, and based on this, a quantifiable and locatable visual inconsistency assessment metric, VSM, is designed.
🎯 What it does: Proposes the MIND algorithm, which directly generates material interfaces (MI) from unsigned distance fields (UDF), achieving non-manifold mesh reconstruction.
MindOmni: Unleashing Reasoning Generation in Vision Language Models with RGPO
Yicheng Xiao (Tsinghua University), Ying Shan (The University of Hong Kong)
CodeGenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: MindOmni is proposed, a unified multimodal large language model that combines reinforcement learning to achieve reasoning generation capabilities, and has advantages in tasks such as image understanding, text generation, and visual editing.
MINGLE: Mixture of Null-Space Gated Low-Rank Experts for Test-Time Continual Model Merging
Zihuan Qiu (University of Electronic Science and Technology of China), Hongliang Li (University of Electronic Science and Technology of China)
CodeMixture of ExpertsMultimodality
🎯 What it does: A framework for continuous model merging using a small number of unlabeled samples during the inference phase (TTCMM) is proposed, along with the implementation scheme MINGLE.
miniF2F-Lean Revisited: Reviewing Limitations and Charting a Path Forward
Azim Ospanov (Huawei Hong Kong Research Center), Roozbeh Yousefzadeh (Chinese University of Hong Kong)
CodeLarge Language ModelTextBenchmark
🎯 What it does: This paper presents miniF2F-v2, correcting and validating over 300 errors and simplifications in the original miniF2F dataset, and constructing a complete automated reasoning pipeline from natural language to Lean proofs, along with an evaluation of its performance.
Minimizing False-Positive Attributions in Explanations of Non-Linear Models
Anders Gjølbye (Technical University of Denmark), Lars Kai Hansen (Technical University of Denmark)
CodeExplainability and InterpretabilityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A method named PatternLocal is proposed, which can suppress erroneous positive attribution caused by confounding variables in local explanations of nonlinear models.
MINT-CoT: Enabling Interleaved Visual Tokens in Mathematical Chain-of-Thought Reasoning
Xinyan Chen (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
CodeReinforcement LearningVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: The MINT-CoT method is proposed, which dynamically inserts fine-grained visual tokens during the chain of thought (CoT) process to enhance multimodal mathematical reasoning capabilities.
Mint: A Simple Test-Time Adaptation of Vision-Language Models against Common Corruptions
Wenxuan Bao (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
CodeDomain AdaptationOptimizationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: We propose Mint, an algorithm that adapts only during testing on LayerNorm parameters, maximizing the inter-class variance of pseudo-labels using cumulative means and gradient accumulators to enhance the robustness of CLIP on corrupted data.
MISA: Memory-Efficient LLMs Optimization with Module-wise Importance Sampling
Yuxi Liu (Peking University), Kun Yuan (Peking University)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A modular importance sampling method MISA has been developed for efficiently training large language models in memory-constrained environments.
🎯 What it does: A missing value imputation method called MIRI is proposed, which iteratively reduces the mutual information between the filled data and the missing mask.
Mitigating Forgetting in LLM Fine-Tuning via Low-Perplexity Token Learning
Chao-Chung Wu (Appier AI Research), Hung-yi Lee (National Taiwan University)
CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies how using data generated by LLM during fine-tuning can effectively reduce forgetting of non-target tasks, and proposes a Selective Token Masking (STM) method to achieve the same effect by masking high perplexity tokens.
Mitigating Hallucination Through Theory-Consistent Symmetric Multimodal Preference Optimization
Wenqi Liu (Shandong University), Liqiang Nie (Harbin Institute of Technology)
CodeRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposes Symmetric Multimodal Preference Optimization (SymMPO), which reduces hallucinations in multimodal large language models (MLLMs) through symmetric preference learning and preference margin consistency regularization;
🎯 What it does: This paper proposes an adaptive sampling method based on Langevin dynamics (LAS) to improve the training stability and convergence speed of Physics-Informed Neural Networks (PINNs).
Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models
Xiwen Wei (University of Texas at Austin), Radu Marculescu (University of Texas at Austin)
CodeGenerationKnowledge DistillationRepresentation LearningMixture of ExpertsImageTextMultimodality
🎯 What it does: The Modality-Decoupled Experts (MoDE) framework is proposed to address the intra-modal and inter-modal catastrophic forgetting issues in Unified Multimodal Generative Models (UMGM) during continual instruction fine-tuning.
Mitigating Overthinking in Large Reasoning Models via Manifold Steering
Yao Huang (Beihang University), Yinpeng Dong (Tsinghua University)
CodeLarge Language ModelText
🎯 What it does: This paper studies the phenomenon of overthinking in large reasoning models (LRM) during the reasoning process through mechanism interpretability methods, and proposes an intervention strategy based on low-dimensional activation subspace—Manifold Steering—to suppress overthinking.
Mitigating Reward Over-optimization in Direct Alignment Algorithms with Importance Sampling
Nguyen Minh Phuc, Khoa D Doan
CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: This paper proposes an Importance Sampling Direct Alignment Algorithm (IS-DAAs), which approximates the KL regularization in online RLHF by multiplying the importance ratio in the objective of direct alignment algorithms (such as DPO), thereby alleviating the issue of reward over-optimization that arises during offline alignment.
MixAT: Combining Continuous and Discrete Adversarial Training for LLMs
Csaba Dékány (INSAIT), Martin Vechev (ETH Zurich)
CodeAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper proposes a novel adversarial training method called MIXAT, which combines continuous and discrete perturbations to enhance the robustness of large language models.
Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go
Yichuan Ma (Fudan University), Kai Chen (Shanghai AI Laboratory)
CodeLarge Language ModelReinforcement LearningMixture of ExpertsTextChain-of-Thought
🎯 What it does: By mixing general LLMs with structured Go expert data, combined with long-chain thinking data for cold start fine-tuning, and then using GRPO reinforcement learning for self-exploration, the LoGos model is ultimately launched, which possesses professional Go level while maintaining general reasoning abilities.
MixPrompt: Efficient Mixed Prompting for Multimodal Semantic Segmentation
Zhiwei Hao (Beijing Institute of Technology), Dan Zeng (Shanghai University)
CodeSegmentationPrompt EngineeringMultimodality
🎯 What it does: The MixPrompt framework is proposed, which integrates auxiliary modalities into a pre-trained RGB segmentation model through a lightweight prompting module, achieving efficient multi-modal semantic segmentation.
🎯 What it does: This paper proposes MixSignGraph, which enhances the performance of gesture recognition and translation models by combining cross-region graph convolution and hierarchical graph convolution. It also introduces the TCTC pre-training scheme to improve translation effectiveness in the absence of gloss annotations.
Mixture of Inputs: Text Generation Beyond Discrete Token Sampling
Yufan Zhuang (University of California San Diego), Jianfeng Gao (Microsoft Research)
CodeGenerationTransformerLarge Language ModelText
🎯 What it does: The Mixture of Inputs (MOI) method is proposed, which mixes the sampled discrete tokens with their corresponding probability distributions into continuous inputs during autoregressive generation, preserving the rich information of the model's predictive distribution; it can be used directly in the existing LLM inference process without additional training.
Mixture of Scope Experts at Test: Generalizing Deeper Graph Neural Networks with Shallow Variants
Gangda Deng (University of Southern California), Viktor Prasanna (University of Southern California)
CodeGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: A framework for mixing different depth GNN experts during the testing phase (Moscat) is proposed, which enhances the generalization ability of deep GNNs on heterogeneous graphs by decoupling expert training and gating learning.
Wenhao Wu (Nanjing University), Zhi Wang (Nanjing University)
CodeRobotic IntelligenceTransformerReinforcement LearningMixture of ExpertsContrastive LearningMultimodality
🎯 What it does: A framework is proposed that introduces Token-wise and Task-wise two-layer Mixture-of-Experts (MoE) in In-Context Reinforcement Learning (ICRL), enhancing the ability to handle multimodal inputs (state, action, reward) and task diversity.
Mixture-of-Experts Operator Transformer for Large-Scale PDE Pre-Training
Hong Wang (University of Science and Technology of China), Yan Jiang
CodeTransformerMixture of ExpertsPhysics Related
🎯 What it does: A pre-training model for PDE based on a sparse Mixture-of-Experts (MoE) Transformer, named MoE-POT, is proposed, which can expand model capacity without increasing inference costs.
Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation
Sangmin Bae (KAIST), Se-Young Yun (KAIST)
CodeComputational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: A Mixture-of-Recursions (MoR) framework is proposed, which efficiently combines parameter sharing, token-level adaptive recursion depth, and KV caching to achieve a lighter Transformer;
MLE-STAR: Machine Learning Engineering Agent via Search and Targeted Refinement
Jaehyun Nam (KAIST), Tomas Pfister (Google Cloud)
CodeOptimizationAI Code AssistantTransformerLarge Language ModelAgentic AIImageTextMultimodalityTabularRetrieval-Augmented GenerationAudio
🎯 What it does: A machine learning engineering agent named MLE-STAR has been developed, which utilizes LLMs and search engines to first retrieve task-related models to generate initial code. It then locates key code blocks through ablation analysis for deep iterative refinement and introduces an automated integration strategy to further enhance performance, suitable for multimodal tasks (tables, images, text, audio).
MLLM-For3D: Adapting Multimodal Large Language Model for 3D Reasoning Segmentation
Jiaxin Huang (MBZUAI), Tongliang Liu (MBZUAI)
CodeSegmentationTransformerLarge Language ModelMultimodalityPoint Cloud
🎯 What it does: The MLLM-For3D framework is proposed, which combines a pre-trained two-dimensional multimodal large language model (MLLM) with SAM to generate multi-view pseudo-segmentation masks and project them into three-dimensional space, achieving label-free three-dimensional inference segmentation.
MLZero: A Multi-Agent System for End-to-end Machine Learning Automation
Haoyang Fang (Amazon Web Services), George Karypis (Amazon Web Services)
CodeLarge Language ModelAgentic AIImageTextMultimodalityTabularBenchmark
🎯 What it does: A multi-agent framework called MLZero is proposed, achieving end-to-end automation from raw multimodal data to a complete machine learning model with zero human intervention.
Ling Yang (Princeton University), Mengdi Wang (Princeton University)
CodeGenerationTransformerLarge Language ModelReinforcement LearningDiffusion modelTextMultimodalityChain-of-Thought
🎯 What it does: MMaDA is proposed—a unified multimodal diffusion language model that can simultaneously perform text reasoning, multimodal understanding, and text-to-image generation;
MoBA: Mixture of Block Attention for Long-Context LLMs
Enzhe Lu (Moonshot AI), Jiezhong Qiu (Hangzhou Institute of Medicine)
CodeTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Proposes Mixture of Block Attention (MoBA), which achieves sparsification of long sequence attention by dynamically routing to select historical KV blocks, and supports unstructured, switchable sparse and full attention.
🎯 What it does: A lightweight network called MobileODE is proposed, which utilizes discrete neural ODEs to replace the 1×1 convolution of MobileNet with a learnable Channelwise ODE Solver (COS).
MobileUse: A Hierarchical Reflection-Driven GUI Agent for Autonomous Mobile Operation
Ning Li (Shanghai Jiao Tong University), Weinan Zhang (Shanghai Jiao Tong University)
CodeRobotic IntelligenceTransformerLarge Language ModelAgentic AIMultimodality
🎯 What it does: A GUI mobile agent named MobileUse is proposed, achieving robust execution and error recovery for long-sequence tasks through hierarchical reflection and active exploration.
🎯 What it does: A batch multi-objective Bayesian optimization algorithm MOBO-OSD based on orthogonal search directions is proposed to efficiently discover the multi-objective Pareto front under a limited evaluation budget.
🎯 What it does: A hierarchical model inversion (PMI) is proposed, combining feature distribution modeling and alignment to achieve synthetic data generation and replay in data-free continual learning, avoiding the storage of historical data.
Model Reconciliation via Cost-Optimal Explanations in Probabilistic Logic Programming
Yinxu Tang (Washington University in St. Louis), William Yeoh (Washington University in St. Louis)
CodeOptimizationExplainability and InterpretabilityTabular
🎯 What it does: This paper proposes a model reconciliation method within the framework of the probabilistic logic program ProbLog, addressing the inconsistency between agent and human models in terms of the most probable explanation (MPE) probabilities by generating cost-optimal explanations.
Model Selection for Off-policy Evaluation: New Algorithms and Experimental Protocol
Pai Liu (University of Illinois), Nan Jiang (University of Illinois)
CodeReinforcement Learning
🎯 What it does: This paper studies how to perform model selection for different OPE (Offline Policy Evaluation) methods in offline reinforcement learning, proposing a new selector based on LSTD called LSTD-Tournament, as well as several model-based selection methods, and designing a controllable and stable experimental protocol.
🎯 What it does: This paper proposes MODEL SHAPLEY, which uses Shapley values to measure the importance of parameters in large neural networks at the parameter level, and achieves a scalable closed-form approximation through a single gradient backpropagation.
🎯 What it does: This paper presents MGAudio, a video-to-audio generation framework based on flow-matching Transformer, which enhances the quality and synchronization of audio through model guidance and dual-role alignment.
🎯 What it does: A framework for the Unbalanced Mean Field Schrödinger Bridge (UMFSB) has been constructed, and a deep learning method called CytoBridge has been implemented to simultaneously learn the stochastic dynamics, proliferation/death, and intercellular interactions of cells from sparse temporal snapshot data.
Modeling Neural Activity with Conditionally Linear Dynamical Systems
Victor Geadah (Princeton University), Alex H Williams
CodeOptimizationComputational EfficiencyTime Series
🎯 What it does: A conditional linear dynamical system (CLDS) model is proposed and implemented, using Gaussian process priors to smoothly vary the parameters of the linear dynamical system (LDS) with experimental conditions. By combining Kalman filtering and EM inference, it can explain the nonlinear dynamics of neural population activity with very little data.
🎯 What it does: The ModHiFi algorithm is proposed, which utilizes Subset Fidelity to identify high-fidelity (HiFi) components in models through synthetic data without training data or loss functions, thereby achieving structured pruning and class-level forgetting.
MoE-Gyro: Self-Supervised Over-Range Reconstruction and Denoising for MEMS Gyroscopes
Feiyang Pan (Southeast University), Guangbin Dou (Southeast University)
CodeRestorationOptimizationMixture of ExpertsAuto EncoderTime SeriesBenchmark
🎯 What it does: Designed and implemented a self-supervised Mixture-of-Experts framework MoE-Gyro for over-range reconstruction and noise suppression of MEMS gyroscopes, addressing the challenges of traditional hardware upgrades.
MoEMeta: Mixture-of-Experts Meta Learning for Few-Shot Relational Learning
Han Wu (University of Sydney), Jie Yin (University of Sydney)
CodeMeta LearningMixture of ExpertsGraph
🎯 What it does: Proposes the MoEMeta framework, which combines Mixture-of-Experts to learn global relationship prototypes and achieves local adaptation through task-specific projections, enhancing few-shot knowledge graph relation learning effectiveness.
🎯 What it does: Proposes the Momentum-SAM (MSAM) optimizer, which imposes sharpness constraints on deep networks without increasing additional forward/backward propagation, achieving better generalization performance.
MonarchAttention: Zero-Shot Conversion to Fast, Hardware-Aware Structured Attention
Can Yaras (University of Michigan), Laura Balzano (University of Michigan)
CodeGenerationOptimizationComputational EfficiencyTransformerLarge Language ModelImageText
🎯 What it does: This paper proposes MonarchAttention, a zero-shot sub-quadratic complexity attention approximation method based on the Monarch structured matrix, and implements an efficient hardware-friendly version aimed at GPUs.
MoniTor: Exploiting Large Language Models with Instruction for Online Video Anomaly Detection
Shengtian Yang (Nanjing University of Aeronautics and Astronautics), Jie Qin (Nanjing University of Aeronautics and Astronautics)
CodeAnomaly DetectionRecurrent Neural NetworkTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: This paper proposes MoniTor, an online, training-free video anomaly detection framework that utilizes large language models (LLM) and visual language models (VLM) to detect and locate anomalous events in real-time streaming video.
Zeqian Ju (University of Science and Technology of China), Xu Tan (Microsoft Research)
CodeGenerationTransformerLarge Language ModelTextAudio
🎯 What it does: MoonCast provides a zero-shot, long-duration, two-speaker podcast generation system that can automatically generate podcast audio with a natural improvisational dialogue style from non-speech sources such as text, web pages, and PDFs.
MoORE: SVD-based Model MoE-ization for Conflict- and Oblivion-Resistant Multi-Task Adaptation
Shen Yuan (Renmin University of China), Hongteng Xu
CodeLarge Language ModelMixture of ExpertsText
🎯 What it does: A model MoE strategy based on Singular Value Decomposition (SVD) called MoORE is proposed for parameter-efficient adaptation of large-scale pre-trained models in multi-task scenarios.
CodeOptimizationDrug DiscoveryTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This paper proposes a fine-grained scientific hypothesis generation task, modeling it as a combinatorial optimization problem, and designs a Hierarchical Heuristic Search (HHS) method to implement it.
More of the Same: Persistent Representational Harms Under Increased Representation
Jennifer Mickel (University of Texas at Austin), Kevin Tian (University of Texas at Austin)
CodeGenerationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the GAS(P) evaluation method, which reveals the representation differences of different groups in generated texts without explicitly prompting group identities, and conducts empirical research in the context of gender and occupation.
🎯 What it does: In a fixed text-to-image diffusion model, a pluggable transformer is utilized to achieve a unification of generation and depth estimation tasks;
More Than Just Functional: LLM-as-a-Critique for Efficient Code Generation
Derui Zhu (Technical University of Munich), Weiyi Shang (University of Waterloo)
CodeOptimizationAI Code AssistantTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Utilizing prompts to enable large language models (LLMs) to serve as 'efficiency judges', guiding the generation of more efficient programs through a reward mechanism during code generation, without the need for code execution verification.
MoRE-Brain: Routed Mixture of Experts for Interpretable and Generalizable Cross-Subject fMRI Visual Decoding
YUXIANG WEI, Vince Calhoun
CodeGenerationExplainability and InterpretabilityMixture of ExpertsDiffusion modelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This study proposes a mixed expert network based on brain functional hierarchy (MoRE-Brain), which maps fMRI signals to CLIP space and achieves high-fidelity, interpretable visual reconstruction through a dynamic time/space routing condition-controlled diffusion model.
🎯 What it does: The first algorithm for Multi-Agent Offline Safe Reinforcement Learning (MOSRL) called MOSDT is proposed, along with the first MOSRL dataset and benchmark, MOSDB.
MOTION: Multi-Sculpt Evolutionary Coarsening for Federated Continual Graph Learning
Guancheng Wan (Wuhan University), Mang Ye (Wuhan University)
CodeFederated LearningGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: This paper studies the problem of Federated Continual Graph Learning (FCGL) and proposes the MOTION framework, which retains historical graph topology using Multi-Expert Trimming (G-TMSC) on the client side and mitigates parameter conflicts using Topology-Sensitive Compatible Matrix Adaptive Aggregation (G-EPAE) on the server side.
🎯 What it does: Incorporating human motion into a multimodal learning framework, a unified embedding space for motion and modalities such as text, images, and audio is constructed, enabling cross-modal retrieval, zero-shot action recognition, and generation from any modality to motion.
🎯 What it does: By retrieving motion patterns from relevant videos and transferring them to the input image, more realistic motion generation is achieved.
MPS-Prover: Advancing Stepwise Theorem Proving by Multi-Perspective Search and Data Curation
Zhenwen Liang (Tencent), Dong Yu (Tencent)
CodeLarge Language ModelTextBenchmark
🎯 What it does: A stepwise automated theorem proving system named MPS-Prover is proposed, which combines post-training data filtering and multi-perspective tree search techniques.
MR. Video: MapReduce as an Effective Principle for Long Video Understanding
Ziqi Pang (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
CodeRecognitionRetrievalOptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: A long video understanding framework MR.Video based on the principles of MapReduce is proposed, which achieves detail perception and global reasoning of long videos through parallel processing of short segments and global aggregation;
MRO: Enhancing Reasoning in Diffusion Language Models via Multi-Reward Optimization
Chenglong Wang (Northeastern University), Tong Xiao (Northeastern University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningDiffusion modelText
🎯 What it does: This paper proposes a Multi-Reward Optimization (MRO) framework to enhance the performance of Diffusion Language Models (DLM) in reasoning tasks by strengthening the token correlations both within and between sequences during the decoding process.
🎯 What it does: Developed the msf-CNN scheme, which utilizes multi-stage patch-based fusion to achieve dual optimization of memory and latency during CNN inference on MCUs.
MSTAR: Box-free Multi-query Scene Text Retrieval with Attention Recycling
Liang Yin (Huazhong University of Science and Technology), Yuliang Liu (Huazhong University of Science and Technology)
CodeRetrievalTransformerVision Language ModelContrastive LearningImageTextBenchmark
🎯 What it does: MSTAR is proposed, a multi-query scene text retrieval method with box-free annotations, achieving unified retrieval for multiple query types.
Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree Search
Huanjin Yao (Tsinghua University), Dacheng Tao (Nanyang Technological University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageTextMultimodality
🎯 What it does: This study proposes the Collective Monte Carlo Tree Search (CoMCTS) method for learning-based reasoning in multimodal large language models (MLLMs) and constructs the Mulberry-260k dataset based on this method. Subsequently, the dataset is used for joint supervised fine-tuning of multimodal models, resulting in the Mulberry series models that possess o1-like step-by-step reasoning and reflection capabilities.
🎯 What it does: The MAFIS framework is proposed for multi-agent imitation learning, combining soft Q function factorization with energy model sampling to achieve both distributed, online, and offline training modes.
Multi-Class Support Vector Machine with Differential Privacy
Jinseong Park (Korea Institute for Advanced Study), Jaewook Lee (Seoul National University)
CodeClassificationSafty and PrivacyGaussian SplattingTabular
🎯 What it does: A differential privacy multi-class support vector machine (PMSVM) based on all-in-one is proposed, achieving privacy protection through a single data access.
🎯 What it does: Designed and implemented Multi-Head Temporal Latent Attention (MTLA), a self-attention mechanism capable of dynamically compressing the time dimension of KV cache in the Transformer decoder, and evaluated its performance on various speech and text tasks.
Multi-modal contrastive learning adapts to intrinsic dimensions of shared latent variables
Yu Gui (University of Pennsylvania), Zongming Ma (Yale University)
CodeRepresentation LearningConvolutional Neural NetworkContrastive LearningMultimodalityBiomedical Data
🎯 What it does: This study investigates the theoretical properties of multimodal contrastive learning (CLIP) under nonlinear representations and general data distributions, demonstrating that it can adaptively share the intrinsic dimensionality of latent variables through temperature optimization.
Multi-Modal View Enhanced Large Vision Models for Long-Term Time Series Forecasting
ChengAo Shen (University of Houston), Jingchao Ni (NEC Laboratories America)
CodeTransformerVision Language ModelMultimodalityTime Series
🎯 What it does: A decomposition framework DMMV is proposed for long-term time series forecasting using multimodal perspectives and large visual models.
Multi-Objective Reinforcement Learning with Max-Min Criterion: A Game-Theoretic Approach
Woohyeon Byeon (Korea Advanced Institute of Science and Technology), Youngchul Sung (Korea Advanced Institute of Science and Technology)
CodeOptimizationReinforcement Learning
🎯 What it does: A new framework for max-min multi-objective reinforcement learning (max-min MORL) is proposed, utilizing a game-theoretic perspective of two-player zero-sum games to transform the problem into learning Nash equilibria, followed by the design of a single-loop ERAM/ARAM algorithm to achieve synchronized updates of policies and weights.
🎯 What it does: A knowledge distillation framework called TRUST is proposed for model heterogeneous federated graph learning, addressing the compatibility issue between different client models and the complexity of graph structures.
🎯 What it does: A multi-scale fine-tuning framework MSFT is proposed, which performs multi-scale fine-tuning on encoder-based time series models to enhance the predictive performance of downstream tasks.
Multi-scale Temporal Prediction via Incremental Generation and Multi-agent Collaboration
Zhitao Zeng (National University of Singapore), Yueming Jin (National University of Singapore)
CodeSegmentationGenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelMultimodalityTime SeriesBenchmark
🎯 What it does: This paper proposes a multi-scale temporal prediction task, constructs a synchronized multi-scale annotation benchmark (MSTP), and introduces the IG-MC framework to achieve incremental generation of states and visuals, as well as multi-agent collaborative prediction.
Multi-Task Vehicle Routing Solver via Mixture of Specialized Experts under State-Decomposable MDP
Yuxin Pan (Hong Kong University of Science and Technology), Fangzhen Lin (Hong Kong University of Science and Technology)
CodeOptimizationTransformerReinforcement LearningMixture of Experts
🎯 What it does: A unified multi-task vehicle routing problem (VRP) solver is proposed, utilizing the State-Decomposable MDP (SDMDP) and Latent-Space SDMDP (LS-SDMDP) frameworks to reuse benchmark VRP solvers and implement the Mixture-of-Specialized-Experts Solver (MoSES);
Anastasios Gerontopoulos (Athena Research Center), Nikos Komodakis (IACM-Forth)
CodeTransformerText
🎯 What it does: This paper proposes the MuToR method, which inserts learnable register tokens into autoregressive Transformer training to achieve multi-step prediction.