ICLR 2024 Papers — Page 13
International Conference on Learning Representations · 2260 papers
LLM-Assisted Code Cleaning For Training Accurate Code Generators
Naman Jain (University of California, Berkeley), Ion Stoica (University of California, Berkeley)
Data-Centric LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Using instruction-tuned LLMs (such as GPT-3.5-Turbo) to clean existing code data, first renaming variables, then breaking complex functions into smaller helper functions, and finally adding natural language planning comments at the beginning of the code to generate more structured and readable training samples.
LLM-CXR: Instruction-Finetuned LLM for CXR Image Understanding and Generation
Suhyeon Lee (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelGenerative Adversarial NetworkImageTextMultimodalityBiomedical DataComputed Tomography
🎯 What it does: A multimodal LLM called LLM-CXR is proposed, which achieves understanding of chest X-ray (CXR) images, report generation, and report-to-image generation all within the same model without the need for additional adapter networks or external generative models through instruction finetuning.
LLM-grounded Video Diffusion Models
Long Lian (University of California Berkeley), Boyi Li (University of California Berkeley)
GenerationData SynthesisRetrievalLarge Language ModelDiffusion modelVideoText
🎯 What it does: This paper proposes a framework called LLM-grounded Video Diffusion (LVD), which significantly enhances the semantic and motion alignment between text and video without additional training by first allowing a large language model to generate dynamic scene layouts (DSL) and then using it as a control signal to guide a text-to-video diffusion model in generating videos.
LLMCarbon: Modeling the End-to-End Carbon Footprint of Large Language Models
Ahmad Faiz (Indiana University), Lei Jiang (Indiana University)
Large Language ModelMixture of ExpertsText
🎯 What it does: This paper presents LLMCARBON, an end-to-end carbon footprint prediction model for dense and MoE LLMs.
LLMs Meet VLMs: Boost Open Vocabulary Object Detection with Fine-grained Descriptors
Sheng Jin (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
Object DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: A new open vocabulary object detection method called DVDet is proposed, which utilizes fine-grained descriptors and conditional context prompts to enhance the region-text alignment of visual-language models, thereby improving detection accuracy.
LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset
Lianmin Zheng (University of California Berkeley), Hao Zhang (University of California San Diego)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The LMSYS-Chat-1M dataset has been collected and released, containing 1 million real user dialogues from 25 types of LLMs, covering 210K users and 154 languages. Based on this, four application cases are presented: content moderation, instruction fine-tuning, safety benchmarks, and high-difficulty evaluations.
LMUFormer: Low Complexity Yet Powerful Spiking Model With Legendre Memory Units
Zeyu Liu (University of Southern California), Peter Anthony Beerel
Spiking Neural NetworkTransformerSequentialAudio
🎯 What it does: This paper proposes LMUFormer and its spiking version, which integrates Legendre Memory Units, convolutional patch embeddings, and channel mixers to achieve a Transformer alternative model that allows for parallel training, sequential inference, and a significant reduction in parameters.
Local Composite Saddle Point Optimization
Site Bai (Purdue University), Brian Bullins (Purdue University)
OptimizationFederated LearningImage
🎯 What it does: The Federated Dual Extrapolation (FeDualEx) algorithm is proposed for distributed composite saddle point optimization, combining local updates with an additional step size in the primal-dual method.
Local Graph Clustering with Noisy Labels
Artur Back de Luca (University of Waterloo), Shenghao Yang (University of Waterloo)
OptimizationGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: This paper proposes the construction of a weighted graph using noisy labels, and performs flow diffusion (ℓ2-norm flow diffusion) on this weighted graph to achieve local graph clustering, aiming to improve the identification accuracy of target clusters when only seed nodes are available.
Local Search GFlowNets
Minsu Kim (Korea Advanced Institute of Science and Technology), Jinkyoo Park (Korea Advanced Institute of Science and Technology)
OptimizationDrug DiscoveryReinforcement LearningBiomedical Data
🎯 What it does: An algorithm called Local Search GFlowNets (LS-GFN) is proposed, which incorporates a local search step into GFlowNet training, improving the sampling paths through backward backtracking and forward reconstruction.
Locality Sensitive Sparse Encoding for Learning World Models Online
Zichen Liu (Sea AI Lab), Min Lin (National University of Singapore)
Computational EfficiencyReinforcement LearningWorld ModelImage
🎯 What it does: A world model method called Losse-FTL is proposed, which utilizes sparse nonlinear random feature encoding (Locality Sensitive Sparse Encoding) combined with linear regression to achieve incremental updates of the Follow-The-Leader (FTL) strategy, avoiding catastrophic forgetting in neural networks.
Locality-Aware Graph Rewiring in GNNs
Federico Barbero (University of Oxford), Francesco Di Giovanni (University of Oxford)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes a Local-Aware Serialization Graph Reconnection (LASER) framework to improve message propagation in graph neural networks and reduce over-compression phenomena.
Localizing and Editing Knowledge In Text-to-Image Generative Models
Samyadeep Basu (University of Maryland), Varun Manjunatha (Adobe Research)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: This paper explores the storage location of visual attribute knowledge in text-to-image diffusion models through causal mediation analysis, and based on this, proposes a fast, data-free model editing method called DIFF-QUICKFIX.
LoftQ: LoRA-Fine-Tuning-aware Quantization for Large Language Models
Yixiao Li (Georgia Institute of Technology), Tuo Zhao (Georgia Institute of Technology)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The LoftQ framework is proposed, which combines quantization and low-rank decomposition to provide a better initialization for LoRA fine-tuning after quantization in large language models.
Logical Languages Accepted by Transformer Encoders with Hard Attention
Pablo Barcelo (Pontifical Catholic University of Chile), Vladimir Podolskii
Transformer
🎯 What it does: This paper characterizes the language sets that two types of hard attention Transformer encoders (UHAT and AHAT) can recognize through the combination of formal languages and circuit complexity theory, clarifying their correspondence with AC⁰ and TC⁰.
LogicMP: A Neuro-symbolic Approach for Encoding First-order Logic Constraints
Weidi Xu (INFLY TECH), Wei Chu (BioMap Research)
ClassificationOptimizationComputational EfficiencyGraph Neural NetworkReinforcement LearningTextGraph
🎯 What it does: A pluggable neural layer called LogicMP is proposed, which couples any neural network with first-order logic constraints (MLN) using efficient mean-field inference to achieve neural-symbolic integration.
Long-Short-Range Message-Passing: A Physics-Informed Framework to Capture Non-Local Interaction for Scalable Molecular Dynamics Simulation
Yunyang Li (Yale University), Tie-Yan Liu (Microsoft Research AI4Science)
Graph Neural NetworkGraphPhysics Related
🎯 What it does: Proposes the Long-Short-Range Message-Passing (LSR-MP) framework, which combines chemical segmentation with long-range message passing to enhance the prediction accuracy of EGNN in large molecular dynamics.
Long-tailed Diffusion Models with Oriented Calibration
Tianjiao Zhang (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: For training diffusion models under long-tail distributions, a head-to-tail (T2H) knowledge transfer and batch re-sampling strategy based on multi-objective score estimation is proposed to enhance the generation diversity and quality of tail classes.
Long-Term Typhoon Trajectory Prediction: A Physics-Conditioned Approach Without Reanalysis Data
Young-Jae Park (Gwangju Institute of Science and Technology), Yeji Choi (SI Analytics)
Convolutional Neural NetworkTransformerTime SeriesPhysics Related
🎯 What it does: The LT3P model is proposed, utilizing real-time Unified Model (UM) data to achieve 72-hour long-term typhoon trajectory prediction without relying on traditional physical models.
LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models
Yukang Chen (Chinese University of Hong Kong), Jiaya Jia (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the LongLoRA method, which efficiently extends the context length of LLMs while maintaining the original architecture, using sparse local attention and trainable embedding/normalization layers.
Look, Remember and Reason: Grounded Reasoning in Videos with Language Models
Apratim Bhattacharyya (Qualcomm AI Research), Roland Memisevic (Qualcomm AI Research)
RecognitionObject DetectionObject TrackingTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: A 'Look, Remember, Reason' framework is proposed, embedding low-level visual tasks (detection, re-identification, tracking) into a multimodal language model to achieve reasoning about videos based on fine-grained visual information.
Looped Transformers are Better at Learning Learning Algorithms
Liu Yang (University of Wisconsin), Dimitris Papailiopoulos (University of Wisconsin)
TransformerTabular
🎯 What it does: The research proposes a cyclic Transformer architecture and training method, enabling the Transformer to learn and approximate iterative algorithms for various data fitting problems while significantly reducing the number of parameters.
LOQA: Learning with Opponent Q-Learning Awareness
Milad Aghajohari (University of Montreal), Aaron Courville (University of Montreal)
Reinforcement Learning
🎯 What it does: A decentralized reinforcement learning algorithm LOQA is proposed, which models the opponent's action value function Q to shape the opponent's behavior and achieve mutual cooperation in partially competitive environments.
LoTa-Bench: Benchmarking Language-oriented Task Planners for Embodied Agents
Jae-Woo Choi (Electronics and Telecommunications Research Institute), Minsu Jang (Electronics and Telecommunications Research Institute)
Robotic IntelligenceTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A benchmark framework for the automatic evaluation of LLM-driven language task planners is proposed, and systematic experiments are conducted in the context of home service robot tasks.
Low Rank Matrix Completion via Robust Alternating Minimization in Nearly Linear Time
Yuzhou Gu (Institute of Advanced Study), Lichen Zhang (Massachusetts Institute of Technology)
OptimizationComputational Efficiency
🎯 What it does: A robust alternating minimization algorithm is proposed for low-rank matrix completion in nearly linear time.
lpNTK: Better Generalisation with Less Data via Sample Interaction During Learning
Shangmin Guo (University of Edinburgh), Kenny Smith (University of Edinburgh)
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Analyzed inter-sample interactions through first-order Taylor approximation to study their impact on the generalization of neural networks;
LQ-LoRA: Low-rank plus Quantized Matrix Decomposition for Efficient Language Model Finetuning
Han Guo (Carnegie Mellon University), Yoon Kim (Massachusetts Institute of Technology)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a low-rank + quantized matrix decomposition method (LQ-LoRA), which splits the weights of the pre-trained model into quantizable low-precision matrices and learnable low-rank matrices, updating only the low-rank part during training; it also dynamically allocates quantization configurations for different layers through integer linear programming and provides a data-aware version that utilizes Fisher information for weighted reconstruction.
LRM: Large Reconstruction Model for Single Image to 3D
Yicong Hong (Australian National University), Hao Tan (Adobe Research)
GenerationData SynthesisTransformerNeural Radiance FieldImageVideo
🎯 What it does: We propose LRM, a large-scale Transformer model that can generate high-quality 3D NeRF from a single image in 5 seconds;
LRR: Language-Driven Resamplable Continuous Representation against Adversarial Tracking Attacks
Jianlang Chen (Kyushu University), Jianjun Zhao (Kyushu University)
Object TrackingAdversarial AttackContrastive LearningImageVideo
🎯 What it does: A language-driven resampling continuous representation (LRR) is proposed for preprocessing frames during tracking to resist adversarial attacks.
LUM-ViT: Learnable Under-sampling Mask Vision Transformer for Bandwidth Limited Optical Signal Acquisition
Lingfeng Liu (Zhejiang University), Hangjie Yuan (Zhejiang University)
ClassificationTransformerImage
🎯 What it does: A learnable downsampling mask Vision Transformer (LUM-ViT) is proposed, which achieves efficient acquisition of broadband constrained optical signals through pre-sampling optical modulation.
LUT-GEMM: Quantized Matrix Multiplication based on LUTs for Efficient Inference in Large-Scale Generative Language Models
Gunho Park (Pohang University of Science and Technology), Dongsoo Lee (NAVER Cloud)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A lookup table (LUT)-based quantized matrix multiplication kernel LUT-GEMM has been developed, which directly uses BCQ quantized weights, eliminating the dequantization step and accelerating inference for large-scale language models.
M3C: A Framework towards Convergent, Flexible, and Unsupervised Learning of Mixture Graph Matching and Clustering
Jiaxin Lu (University of Texas), Junchi Yan (Shanghai Jiao Tong University)
Graph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes M3C and its unsupervised extension UM3C for joint learning of mixed graph matching and clustering, balancing matching accuracy and clustering quality.
Machine Unlearning for Image-to-Image Generative Models
Guihong Li (University of Texas at Austin), Radu Marculescu (JPMorgan Chase)
Image TranslationGenerationData SynthesisDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A machine forgetting framework for image-to-image generation models is proposed, along with an efficient forgetting algorithm based solely on the encoder, which can completely erase the memory of forgotten samples without degrading the performance of retained samples.
MaGIC: Multi-modality Guided Image Completion
Hao Wang (University of Chinese Academy of Sciences), Libo Zhang (Institute of Software, Chinese Academy of Sciences)
RestorationGenerationDiffusion modelImageMultimodality
🎯 What it does: This paper proposes a scalable multimodal guided image inpainting framework called MaGIC, which first injects modality information into a frozen Stable Diffusion model using a single-modal conditional U-Net MCU-Net, and then achieves the fusion of arbitrary modality combinations through a training-free CMB algorithm to complete image inpainting for large occlusions.
Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors
Guocheng Qian (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
GenerationData SynthesisDiffusion modelNeural Radiance FieldImageMesh
🎯 What it does: The MAGIC123 method is proposed, which achieves high-quality texture 3D mesh generation from a single image through a two-stage coarse-to-fine pipeline and a joint 2D/3D pre-trained diffusion model.
MagicDrive: Street View Generation with Diverse 3D Geometry Control
Ruiyuan Gao (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
Object DetectionSegmentationGenerationData SynthesisAutonomous DrivingDiffusion modelImageVideoText
🎯 What it does: Developed the MAGICDRIVE framework, which can generate multi-view street scene images and videos based on multi-source 3D geometric information such as road maps, 3D bounding boxes, and camera poses.
Magnitude Invariant Parametrizations Improve Hypernetwork Learning
Jose Javier Gonzalez Ortiz (Massachusetts Institute of Technology), Adrian V Dalca (Massachusetts Institute of Technology)
SegmentationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This study addresses and solves the amplitude ratio problem in the training of super networks, proposing MIP for stable and efficient training.
Magnushammer: A Transformer-Based Approach to Premise Selection
Maciej Mikuła (Google DeepMind), Yuhuai Wu (xAI)
RetrievalTransformerContrastive LearningText
🎯 What it does: In this paper, the authors propose a Transformer-based premise selection method called Magnushammer and construct the largest Isabelle premise selection dataset.
Making LLaMA SEE and Draw with SEED Tokenizer
Yuying Ge (Tencent AI Lab), Ying Shan (Tencent AI Lab)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageVideoTextMultimodality
🎯 What it does: This paper proposes the SEED tokenizer and SEED-LLaMA, which integrates large language models with discrete visual codes to construct a unified multimodal autoregressive model that supports text generation, image understanding, image generation, and even multi-turn interaction and zero-shot compositional generation.
Making Pre-trained Language Models Great on Tabular Prediction
Jiahuan Yan (Zhejiang University), Jintai Chen (University of Illinois at Urbana-Champaign)
ClassificationOptimizationTransformerLarge Language ModelSupervised Fine-TuningTabular
🎯 What it does: A pre-trained language model for table prediction, TP-BERTa, is proposed.
Making Retrieval-Augmented Language Models Robust to Irrelevant Context
Ori Yoran (Tel Aviv University), Jonathan Berant (Tel Aviv University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Research and address the performance decline issue caused by retrieving irrelevant contexts in Retrieval-Augmented Language Models (RALM).
Making RL with Preference-based Feedback Efficient via Randomization
Runzhe Wu (Cornell University), Wen Sun (Cornell University)
Computational EfficiencyReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: A randomized reinforcement learning algorithm is proposed, which can achieve a balance of sample efficiency, computational efficiency, and query efficiency in environments using preference-based feedback.
MAMBA: an Effective World Model Approach for Meta-Reinforcement Learning
Zohar Rimon (Technion - Israel Institute of Technology), Aviv Tamar (Technion - Israel Institute of Technology)
Meta LearningRecurrent Neural NetworkReinforcement LearningWorld ModelSequential
🎯 What it does: MAMBA is proposed, a meta-reinforcement learning algorithm based on Dreamer, designed for efficient learning of Bayesian optimal policies in multi-task environments.
MAmmoTH: Building Math Generalist Models through Hybrid Instruction Tuning
Xiang Yue (Ohio State University), Wenhu Chen (University of Waterloo)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: This paper constructs a specialized instruction tuning dataset for mathematical reasoning, MathInstruct, and trains the MAmmoTH series of open-source large language models, significantly enhancing their reasoning capabilities on various mathematical problems.
Manifold Diffusion Fields
Ahmed A. A. Elhag, Miguel Ángel Bautista (Apple)
GenerationData SynthesisDiffusion modelPoint CloudMesh
🎯 What it does: This paper proposes a Manifold Diffusion Fields (MDF) based on diffusion models, used for learning the distribution of continuous functions on arbitrary Riemannian manifolds.
Manifold Preserving Guided Diffusion
Yutong He (Carnegie Mellon University), Stefano Ermon (Stanford University)
GenerationData SynthesisOptimizationComputational EfficiencyDiffusion modelAuto EncoderImage
🎯 What it does: A training-free conditional generation framework MPGD is proposed, utilizing pre-trained diffusion models and autoencoders to achieve various conditional generation tasks at low computational cost.
Manipulating dropout reveals an optimal balance of efficiency and robustness in biological and machine visual systems
Jacob S. Prince (Harvard University), Talia Konkle (Harvard University)
ClassificationRepresentation LearningConvolutional Neural NetworkGenerative Adversarial NetworkImageMagnetic Resonance Imaging
🎯 What it does: By systematically varying the dropout rate in the AlexNet model, this study explores the impact of dropout on the dimensionality of the representational space, robustness, and its correspondence with the human brain.
MAP IT to Visualize Representations
Robert Jenssen (UiT Arctic University of Norway & University of Copenhagen & Norwegian Computing Center)
OptimizationRepresentation LearningImage
🎯 What it does: A novel visualization dimensionality reduction method called MAP IT is proposed, based on projective divergence and marginal probability alignment.
MAPE-PPI: Towards Effective and Efficient Protein-Protein Interaction Prediction via Microenvironment-Aware Protein Embedding
Lirong Wu (Westlake University), Stan Z. Li (Westlake University)
Protein Structure PredictionGraph Neural NetworkAuto EncoderGraphBiomedical Data
🎯 What it does: A microenvironment-based protein embedding framework MAPE-PPI is proposed to efficiently predict protein-protein interactions.
Mask-Based Modeling for Neural Radiance Fields
Ganlin Yang (University of Science and Technology of China), Dong Liu (University of Science and Technology of China)
GenerationRepresentation LearningTransformerNeural Radiance FieldContrastive LearningPoint Cloud
🎯 What it does: A self-supervised masking ray and view modeling framework for general NeRF (MRVM-NeRF) is proposed, which significantly enhances the prior learning and inference accuracy of 3D scenes by randomly masking features and cross-view features during the fine-grained sampling stage and aligning targets with the online branch in the latent space.
Masked Audio Generation using a Single Non-Autoregressive Transformer
Alon Ziv (Meta), Yossi Adi (Meta)
GenerationData SynthesisTransformerAudio
🎯 What it does: This paper proposes a single-stage, non-autoregressive transformer model called MAGNET, which directly generates audio on multi-stream audio discrete representations using mask generation techniques.
Masked Autoencoders with Multi-Window Local-Global Attention Are Better Audio Learners
Sarthak Yadav (Aalborg University), Zheng-Hua Tan (Aalborg University)
Representation LearningTransformerAuto EncoderAudio
🎯 What it does: Design and evaluate a Multi-Window Masked Autoencoder (MW-MAE) to learn general audio representations.
Masked Completion via Structured Diffusion with White-Box Transformers
Druv Pai (University of California Berkeley), Yi Ma (University of California Berkeley)
CompressionRepresentation LearningTransformerDiffusion modelAuto EncoderImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a white-box Transformer-like autoencoder CRATE-MAE, which constructs interpretable encoder and decoder layers using the equivalence of structured diffusion and denoising/compression, and performs unsupervised pre-training through a masked autoencoding (MAE) task.
Masked Distillation Advances Self-Supervised Transformer Architecture Search
Caixia Yan (Xi'an Jiaotong University), Qinghua Zheng (Xi'an Jiaotong University)
Knowledge DistillationNeural Architecture SearchTransformerImage
🎯 What it does: This paper proposes MaskTAS, an unsupervised visual Transformer architecture search framework based on Masked Image Modeling (MIM), which can automatically search for efficient Transformer structures without the need for labels.
Masked Structural Growth for 2x Faster Language Model Pre-training
Yiqun Yao (Beijing Academy of Artificial Intelligence), Yequan Wang (Beijing Academy of Artificial Intelligence)
Hyperparameter SearchTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The Masked Structural Growth (MSG) scheme is proposed, which gradually expands the model structure in a masked manner during the pre-training process of the Transformer while strictly maintaining function invariance.
Masks, Signs, And Learning Rate Rewinding
Advait Harshal Gadhikar (CISPA Helmholtz Center for Information Security), Rebekka Burkholz (CISPA Helmholtz Center for Information Security)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the mechanisms of Learning Rate Replay (LRR) and Iterative Magnitude Pruning (IMP) in sparse network learning. It first provides a proof of gradient flow analysis and symbolic switching within the theoretical framework of single hidden layer ReLU neurons, and then conducts experimental validation on standard benchmarks such as CIFAR-10, CIFAR-100, and Tiny ImageNet.
Massive Editing for Large Language Models via Meta Learning
Chenmien Tan (University of Edinburgh), Jie Fu (Hong Kong University of Science and Technology)
Meta LearningLarge Language ModelText
🎯 What it does: Utilizing hypernetworks for large-scale factual editing of large language models
Massively Scalable Inverse Reinforcement Learning in Google Maps
Matt Barnes (Google Research), Shawn O'Banion (Google Research)
Autonomous DrivingOptimizationReinforcement LearningMixture of ExpertsGraph
🎯 What it does: Achieving scalable inverse reinforcement learning routing on a global road network (approximately 200 million nodes), proposing Receding Horizon Inverse Planning (RHIP) and completing full training;
Mastering Memory Tasks with World Models
Mohammad Reza Samsami (Mila Quebec AI Institute), Sarath Chandar (Mila Quebec AI Institute)
Robotic IntelligenceRecurrent Neural NetworkReinforcement LearningWorld ModelSequential
🎯 What it does: A novel model-based reinforcement learning (MBRL) framework called Recall to Imagine (R2I) is implemented by introducing a state space model (SSM) into the world model of DreamerV3.
Mastering Symbolic Operations: Augmenting Language Models with Compiled Neural Networks
Yixuan Weng (University of Chinese Academy of Sciences), Jun Zhao (University of Chinese Academy of Sciences)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Developed the Neural Comprehension framework, which integrates Compiled Neural Networks (CoNNs) with pre-trained language models to achieve symbolic reasoning and rule execution, supporting end-to-end differentiable symbolic computation.
Matcher: Segment Anything with One Shot Using All-Purpose Feature Matching
Yang Liu (Zhejiang University), Chunhua Shen (Zhejiang University)
SegmentationContrastive LearningImageVideo
🎯 What it does: Proposes the Matcher framework, which utilizes pre-trained visual foundation models (DINOv2+SAM) to complete various segmentation tasks under the guidance of one-shot examples, completely without training.
MathCoder: Seamless Code Integration in LLMs for Enhanced Mathematical Reasoning
Ke Wang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Introducing a closed loop of natural language reasoning, code generation, and execution feedback in open-source large language models, we construct the MathCoder model and its corresponding MathCodeInstruct dataset to enhance mathematical reasoning capabilities.
Mathematical Justification of Hard Negative Mining via Isometric Approximation Theorem
Albert Xu (Carnegie Mellon University), Howie Choset (Carnegie Mellon University)
RetrievalOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proves that the triplet loss under hard negative sampling is equivalent to minimizing a Hausdorff-like distance through the isometric approximation theorem, thereby providing a theoretical explanation for network collapse (where all embeddings converge to the same point) and further points out that batch size and embedding dimension are the main factors leading to collapse; this conclusion is also verified through experiments.
MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts
Pan Lu (University of California Los Angeles), Jianfeng Gao (Microsoft Research)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: A benchmark called MATHVISTA has been constructed to evaluate the mathematical reasoning capabilities of models in visual scenes.
Matrix Manifold Neural Networks++
Xuan Son Nguyen (CY Cergy Paris University), Aymeric Histace (CY Cergy Paris University)
ClassificationRecognitionGraph Neural NetworkGraph
🎯 What it does: This paper constructs fully connected layers, convolutional layers, and polynomial logistic regression layers for positive semi-definite matrices in Riemannian random spaces, and implements gradient propagation of logarithmic mapping from the perspective of Grassmann projection, subsequently applying it to construct Grassmann graph convolutional networks.
Matryoshka Diffusion Models
Jiatao Gu (Apple), Navdeep Jaitly (Apple)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: Proposes the Matryoshka Diffusion (MDM) framework, which directly trains high-resolution image and video generation models in pixel space using a multi-resolution joint diffusion process;
Maximum Entropy Heterogeneous-Agent Reinforcement Learning
Jiarong Liu (Peking University), Yaodong Yang (Peking University)
Reinforcement Learning
🎯 What it does: This paper proposes a heterogeneous agent soft Actor-Critic algorithm based on maximum entropy (HASAC), which addresses the issues of high sample complexity, unstable training, and the tendency to fall into suboptimal Nash equilibria in traditional multi-agent reinforcement learning.
Maximum Entropy Model Correction in Reinforcement Learning
Amin Rakhsha (University of Toronto), Amir-massoud Farahmand (University of Toronto)
Reinforcement Learning
🎯 What it does: A reinforcement learning method based on maximum entropy model correction (MoCoVI and MoCoDyna) is proposed, which dynamically corrects the approximate model to reduce model error.
Maximum Likelihood Estimation is All You Need for Well-Specified Covariate Shift
Jiawei Ge (Princeton University), Chi Jin (Princeton University)
Domain AdaptationOptimization
🎯 What it does: It is proven that under the regularization assumption, traditional maximum likelihood estimation can achieve the optimal bound for minimizing risk under covariate shift.
Mayfly: a Neural Data Structure for Graph Stream Summarization
Yuan Feng (University of Science and Technology of China), S Kevin Zhou
CompressionMeta LearningGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A neural data structure named Mayfly is proposed for approximate compression and real-time querying of graph streams under single pass and limited space conditions.
MBR and QE Finetuning: Training-time Distillation of the Best and Most Expensive Decoding Methods
Mara Finkelstein (Google), Markus Freitag (Google)
Computational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningText
🎯 What it does: Proposes quality improvement of MBR and QE decoding during the training phase for neural machine translation tasks, while maintaining efficient decoding during inference.
MCM: Masked Cell Modeling for Anomaly Detection in Tabular Data
Jiaxin Yin (Beijing University of Posts and Telecommunications), Jie Yang (Beijing University of Posts and Telecommunications)
Anomaly DetectionTabularFinance Related
🎯 What it does: This study investigates anomaly detection in tabular data and proposes a self-supervised method based on Masked Cell Modeling (MCM).
Mean Field Theory in Deep Metric Learning
Takuya Furusawa (ZOZO Research)
ClassificationRetrievalTransformerContrastive LearningImage
🎯 What it does: By introducing mean field theory from statistical physics into deep metric learning, two new classification-based loss functions (MeanFieldContrastive and MeanFieldClassWiseMultiSimilarity) have been designed, significantly reducing the training complexity of traditional contrastive-based losses.
Meaning Representations from Trajectories in Autoregressive Models
Tian Yu Liu (University of California Los Angeles), Stefano Soatto (Amazon Web Services AI Labs)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: A method is proposed to represent sentence meaning using the trajectory distribution generated by an autoregressive language model, completely without the need for fine-tuning or specific prompts;
Measuring Vision-Language STEM Skills of Neural Models
Jianhao Shen (Peking University), Chenguang Wang (Washington University in St. Louis)
TransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmark
🎯 What it does: A multimodal STEM benchmark dataset covering the four major fields of science, technology, engineering, and mathematics (STEM) has been proposed and constructed, containing 1,073,146 multiple-choice questions based on 448 skills from the K-12 curriculum. A systematic evaluation of multimodal models (CLIP) and large language models (GPT-3.5-Turbo) was then conducted on this dataset under zero-shot and fine-tuning scenarios, compared to the performance of elementary school students.
Mechanistically analyzing the effects of fine-tuning on procedurally defined tasks
Samyak Jain (University of Cambridge), David Krueger (University of Cambridge)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: This paper provides an explanatory analysis of the fine-tuning mechanism of large-scale pre-trained models, finding that fine-tuning typically adds only a minimal 'wrapper' to the pre-trained capabilities rather than changing the core abilities.
Mediator Interpretation and Faster Learning Algorithms for Linear Correlated Equilibria in General Sequential Games
Brian Hu Zhang (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
Reinforcement LearningSequentialBenchmark
🎯 What it does: This paper studies and implements an efficient learning algorithm for linear-swap regret, providing both theoretical and experimental validation in generalized games.
Mega-TTS 2: Boosting Prompting Mechanisms for Zero-Shot Speech Synthesis
Ziyue Jiang (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationData SynthesisTransformerPrompt EngineeringAuto EncoderGenerative Adversarial NetworkAudio
🎯 What it does: This paper presents Mega-TTS 2, a universal prompting mechanism for zero-shot TTS that supports multi-sentence voice prompts and controllable prosody transfer.
Memorization Capacity of Multi-Head Attention in Transformers
Sadegh Mahdavi (University of British Columbia), Christos Thrampoulidis (University of British Columbia)
TransformerImage
🎯 What it does: This paper studies the memory capacity of the multi-head attention mechanism and provides a lower bound under the assumption of linear independence.
Memorization in Self-Supervised Learning Improves Downstream Generalization
Wenhao Wang (CISPA), Franziska Boenisch (CISPA)
SegmentationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: A memory-based definition for self-supervised learning (SSL) called SSLMem is proposed, which quantifies the degree of memory of training samples across different encoders, further demonstrating that memorization can enhance the generalization performance of downstream tasks.
Memory-Assisted Sub-Prototype Mining for Universal Domain Adaptation
Yuxiang Lai (Southeast University), Tao Zhou (Nanjing University of Science and Technology)
Domain AdaptationContrastive LearningImage
🎯 What it does: The Memory-Assisted Sub-Prototype Mining (MemSPM) method is proposed, which utilizes a learnable memory structure to mine sub-prototypes within the same category, thereby better capturing the conceptual differences within categories in general domain adaptation and achieving inter-domain alignment through task-oriented embedding.
Memory-Consistent Neural Networks for Imitation Learning
Kaustubh Sridhar (University of Pennsylvania), Insup Lee (University of Pennsylvania)
Autonomous DrivingRobotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: Proposes the Memory Consistent Neural Network (MCNN) and applies it to behavior cloning, significantly improving the generalization and robustness of imitation learning across various tasks.
MEND: Meta Demonstration Distillation for Efficient and Effective In-Context Learning
Yichuan Li (Worcester Polytechnic Institute), Chenlei Guo (Amazon Alexa AI)
Computational EfficiencyKnowledge DistillationMeta LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A demonstration compression method called MEND is proposed, which utilizes a hypernetwork to compress long demonstrations into vectors and directly feeds them to a large language model for efficient context learning without retraining for new tasks.
Merge, Then Compress: Demystify Efficient SMoE with Hints from Its Routing Policy
Pingzhi Li (University of North Carolina at Chapel Hill), Tianlong Chen (Harvard University)
CompressionComputational EfficiencyKnowledge DistillationSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: A routing strategy-based Sparse Mixture of Experts (SMoE) merging and compression framework, MC-SMoE, is proposed. It first merges redundant experts into fewer but more knowledgeable experts through M-SMoE, and then compresses the merged experts using low-rank and sparse decomposition.
MERT: Acoustic Music Understanding Model with Large-Scale Self-supervised Training
Yizhi LI, Jie Fu (Beijing Academy of Artificial Intelligence)
Representation LearningTransformerContrastive LearningAudio
🎯 What it does: This paper proposes MERT—a multi-task audio pre-training model based on self-supervised learning, designed to learn general representations of music audio and evaluate downstream performance on various MIR tasks.
Meta Continual Learning Revisited: Implicitly Enhancing Online Hessian Approximation via Variance Reduction
Yichen Wu (City University of Hong Kong), Ying Wei (Nanyang Technological University)
OptimizationMeta LearningSequential
🎯 What it does: This paper reinterprets the Meta-CL method as a technique for online approximate Hessian and proposes the Momentum-Based Variance-Reduced Meta-CL (VR-MCL) to reduce the variance of the hypergradient, thereby enhancing the stability of model updates in continual learning.
Meta Inverse Constrained Reinforcement Learning: Convergence Guarantee and Generalization Analysis
Shicheng Liu (Pennsylvania State University), Minghui Zhu (Pennsylvania State University)
Meta LearningReinforcement LearningSequential
🎯 What it does: The M-ICRL framework is proposed, implementing a meta-learning method to learn expert reward functions and constraints from a small number of demonstrations, with convergence and generalization theoretical guarantees provided.
Meta-Evolve: Continuous Robot Evolution for One-to-many Policy Transfer
Xingyu Liu (Carnegie Mellon University), Ding Zhao (Carnegie Mellon University)
Robotic IntelligenceReinforcement LearningMultimodality
🎯 What it does: This study investigates how to efficiently transfer the expert policy of source robots to multiple target robots.
Meta-Learning Priors Using Unrolled Proximal Networks
Yilang Zhang (University of Minnesota), Georgios B. Giannakis (University of Minnesota)
OptimizationMeta LearningImage
🎯 What it does: A novel explainable prior learning framework called MetaProxNet is proposed, which utilizes a learnable piecewise linear function to approximate the projection operator for task-independent regularization.
Meta-VBO: Utilizing Prior Tasks in Optimizing Risk Measures with Gaussian Processes
Quoc Phong Nguyen (Massachusetts Institute of Technology), Patrick Jaillet (Massachusetts Institute of Technology)
Recommendation SystemOptimizationMeta LearningTabularFinance Related
🎯 What it does: A meta-Bayesian optimization algorithm (meta-VBO) for risk measures (VaR, CVaR) is proposed, which can accelerate the optimization of the current task by utilizing the results of previous tasks.
MetaCoCo: A New Few-Shot Classification Benchmark with Spurious Correlation
Min Zhang (Zhejiang University), Kun Kuang (Zhejiang University)
ClassificationMeta LearningConvolutional Neural NetworkSupervised Fine-TuningImageBenchmark
🎯 What it does: Proposed the MetaCoCo benchmark dataset and evaluated the performance of various few-shot learning methods under the presence of spurious correlation shifts.
MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework
Sirui Hong (DeepWisdom), Jürgen Schmidhuber (AI Initiative, King Abdullah University of Science and Technology)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: A multi-agent collaboration framework based on LLM, MetaGPT, has been constructed to simulate the software company process and complete end-to-end generation from requirements to code.
MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models
Longhui Yu (University of Cambridge), Weiyang Liu (Max Planck Institute for Intelligent Systems)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes the MetaMathQA dataset and the MetaMath LLM, utilizing forward rewriting and backward reasoning to expand mathematical problems from multiple perspectives, and trains LLaMA-2 to enhance mathematical reasoning capabilities.
MetaPhysiCa: Improving OOD Robustness in Physics-informed Machine Learning
S Chandra Mouli (Purdue University), Bruno Ribeiro (Purdue University)
Domain AdaptationOptimizationMeta LearningTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes MetaPhysiCa, a physical information machine learning framework that achieves robustness in predicting out-of-distribution (OOD) dynamical systems through meta-learning and causal structure discovery.
MetaTool Benchmark for Large Language Models: Deciding Whether to Use Tools and Which to Use
Yue Huang (Lehigh University), Lichao Sun (Lehigh University)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposes the METATOOL benchmark to evaluate the capabilities of large language models in tool awareness and tool selection; simultaneously constructs the TOOLE dataset, which includes 21,127 diverse user queries (single tool and multiple tools), and designs four sub-tasks to examine different dimensions of tool selection.
METRA: Scalable Unsupervised RL with Metric-Aware Abstraction
Seohong Park (University of California), Sergey Levine (University of California)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes a scalable unsupervised reinforcement learning framework called METRA, which utilizes time distance metrics to achieve a compact abstraction of the environment in latent space and learn diverse skills.
MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process
Xinyao Fan (University of British Columbia), Jiang Bian (Microsoft Research)
Recurrent Neural NetworkDiffusion modelTime Series
🎯 What it does: This paper proposes a Multi-Granularity Time Series Diffusion Model (MG-TSD), which guides the diffusion model learning by using time series of different granularities as targets in the intermediate steps of the diffusion process, thereby enhancing the stability and accuracy of probabilistic time series forecasting.
MgNO: Efficient Parameterization of Linear Operators via Multigrid
Juncai He (King Abdullah University of Science and Technology), Jinchao Xu (King Abdullah University of Science and Technology)
Convolutional Neural Network
🎯 What it does: A multi-grid based neural operator architecture MgNO is proposed, which parameterizes linear operators using multi-channel convolution without the need for traditional lifting/projection layers.
Mind Your Augmentation: The Key to Decoupling Dense Self-Supervised Learning
Congpei Qiu (Xi'an Jiaotong University), Sabine Süsstrunk
Object DetectionSegmentationConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This study investigates the object coupling problem in dense self-supervised learning and proposes a solution using Region Collaborative Cutout (RCC) and a decoupling branch.
MINDE: Mutual Information Neural Diffusion Estimation
Giulio Franzese (EURECOM), Pietro Michiardi (EURECOM)
Diffusion modelScore-based ModelImage
🎯 What it does: A mutual information (MI) and entropy estimation method called MINDE is proposed, based on diffusion models and the Girsanov theorem.