π― What it does: A low-bit quantization method named LeanQuant is proposed, which maintains higher accuracy in post-training quantization of LLMs by utilizing a loss error-aware quantization grid (non-uniform and adaptively uniform) and achieves scalable acceleration through the integration of GPU cores.
π― What it does: The STEEL algorithm is proposed, achieving unsupervised representation learning in a single continuous trajectory, non-resetting Ex-BMDP environment, and providing an upper bound on sample complexity.
CodeOptimizationComputational EfficiencyTime SeriesPhysics Related
π― What it does: A physics-informed neural iterative solver is proposed to quickly solve parameterized partial differential equations (PDEs), achieving convergence solely through PDE parameters during inference.
π― What it does: A medical diagnosis framework based on causal alignment is proposed, utilizing counterfactual generation to identify the causal chain of model decisions, and employing causal alignment loss to focus the model on diagnostic factors consistent with radiologists.
Learning Diagrams: A Graphical Language for Compositional Training Regimes
Mason Lary (University at Buffalo), James Fairbanks
CodeKnowledge DistillationImageText
π― What it does: This paper proposes the Learning Diagrams, a graphical and composable framework for describing and constructing multi-model deep learning training processes, providing a unified semantics and an operable DSL;
Learning Distributions of Complex Fluid Simulations with Diffusion Graph Networks
Mario Lino Valencia, Nils Thuerey (Technical University of Munich)
CodeGenerationData SynthesisOptimizationGraph Neural NetworkDiffusion modelTime Series
π― What it does: A diffusion model based on graph neural networks (DGN and LDGN) is proposed, which can directly learn and sample the equilibrium state distribution of fluid simulations from incomplete short time series data, thereby quickly obtaining flow field statistics.
Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning
Seanie Lee (Korea Advanced Institute of Science and Technology), Moksh Jain (Mila - Quebec Artificial Intelligence Institute)
CodeAdversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: By using a two-stage training approach (GFlowNet fine-tuning + MLE smoothing), diverse and effective attack prompts are generated, which are then used for automated red teaming tests on various large language models, verifying the cross-model transferability of the generated prompts and the enhancement of security fine-tuning.
Yi Ren (University of British Columbia), Danica J. Sutherland (University of British Columbia)
CodeOptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This study unifies the learning dynamics of large language models (LLMs) at different fine-tuning stages (SFT, DPO, etc.) and provides a step-by-step impact decomposition formula; based on this, it explains various phenomena (such as repetitive nonsense, confidence decay, compression effects, etc.) and proposes a simple improvement method to alleviate the compression effect by expanding the dataset during the SFT stage.
Learning Efficient Positional Encodings with Graph Neural Networks
Charilaos Kanatsoulis, Alejandro Ribeiro (University of Pennsylvania)
CodeGraph Neural NetworkGraph
π― What it does: A learnable position encoding framework based on graph neural networks (PEARL) is proposed, which achieves transferable position encoding for graphs through random or basis vector initialization and statistical pooling after multiple message passing.
Learning Equivariant Non-Local Electron Density Functionals
Nicholas Gao (Munich Data Science Institute), Stephan GΓΌnnemann (Munich Data Science Institute)
CodeGraph Neural NetworkPoint CloudPhysics Related
π― What it does: A non-local electronic density functional EG-XC based on SO(3) equivariant graph neural networks has been designed, which can efficiently learn non-local interactions by compressing the electronic density into a point cloud centered at the nucleus.
π― What it does: This paper proposes the Gain Map-based Inverse Tone Mapping (GM-ITM) task and designs a dual-branch network GMNet to learn the Gain Map corresponding to SDR images, achieving more efficient HDR up-conversion.
π― What it does: Proposes the HiSSD framework, which combines common and task-specific skills in collaborative learning to enhance offline multi-task multi-agent cooperation strategy transfer.
Learning Geometric Reasoning Networks For Robot Task And Motion Planning
Smail Ait Bouhsain (National Center for Scientific Research), Thierry Simeon (National Center for Scientific Research)
CodeRobotic IntelligenceGraph Neural NetworkGraph
π― What it does: A geometric reasoning network (GRN) based on graph neural networks is proposed to quickly predict the feasibility of robot operations (pick/put) and grasping types in a 3D environment, and to provide reasons for infeasibility.
π― What it does: This paper proposes the LIRS framework, which achieves graph invariant feature learning by first learning the outlier (spuriosity) characteristics in the graph and removing them from the features learned through ERM, significantly improving OOD generalization performance.
CodeGraph Neural NetworkTransformerContrastive LearningGraphBenchmarkPhysics Related
π― What it does: A graph quantization tokenizer (GQT) based on multi-task self-supervised learning and residual vector quantization is proposed, which maps graph nodes to discrete and compressed tokens.
CodeGenerationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: The HASS (Harmonized Speculative Sampling) scheme is proposed, which achieves a higher acceptance rate by aligning the target distribution with the context during training and decoding, significantly accelerating the inference of LLMs;
CodeRecommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Using a pre-trained LLM to generate comparative judgments through self-sampling, training on preference data to obtain a decision model, Con-J, which can provide both binary preference judgments and explanations.
Gang Liu (University of Notre Dame), Shantanu Singh (Broad Institute of MIT and Harvard)
CodeRepresentation LearningDrug DiscoveryGraph Neural NetworkMultimodalityGraphBiomedical Data
π― What it does: The InfoAlign method is proposed, which learns molecular representations on a cell context graph through an information bottleneck, allowing molecular embeddings to fully decode multimodal features such as cell morphology and gene expression while maintaining minimal redundant information.
Learning on One Mode: Addressing Multi-modality in Offline Reinforcement Learning
Mianchu Wang (University of Warwick), Giovanni Montana (Alan Turing Institute)
CodeReinforcement LearningMultimodality
π― What it does: This paper proposes an offline reinforcement learning method called LOM, which models the behavior policy as a Gaussian mixture model and selects the single mode with the highest reward for weighted imitation learning, thereby achieving better policy learning on multimodal datasets.
Learning Partial Graph Matching via Optimal Partial Transport
Gathika Ratnayaka (Australian National University), Qing Wang (Australian National University)
CodeOptimizationGraph Neural NetworkGraph
π― What it does: Proposes an optimal partial transport-based partial graph matching optimization framework that automatically decides which nodes to match and provides the optimal partial mapping.
Learning Spatial-Semantic Features for Robust Video Object Segmentation
Xin Li (Harbin Institute of Technology), Ming-Hsuan Yang
CodeSegmentationTransformerVideo
π― What it does: A robust video object segmentation framework (S3) based on spatial-semantic feature learning and discriminative query propagation is proposed.
π― What it does: This paper proposes a Regularized Unbalanced Optimal Transport (DeepRUOT) method based on deep learning, which can learn continuous non-equilibrium stochastic dynamics from sparse temporal snapshot data.
π― What it does: This paper builds upon the existing CPCC regularization framework by introducing Optimal Transport (OT) distance to measure the similarity between class distributions, thereby achieving more fine-grained and accurate hierarchical embeddings.
π― What it does: This study investigates the Continuous Dynamic Learning (CDL) task and proposes the MS-GODE model, which can continuously learn across multiple systems without forgetting, and establishes the biological cell system benchmark Bio-CDL.
Learning to Discover Regulatory Elements for Gene Expression Prediction
Xingyu Su (Texas A&M University), Shuiwang Ji (Texas A&M University)
CodeBiomedical Data
π― What it does: The Seq2Exp framework is proposed to predict gene expression by learning and extracting regulatory elements from DNA sequences and epigenetic signals.
π― What it does: This study proposes a lightweight framework (LD3) for pre-trained diffusion models, which significantly reduces the number of neural network evaluations while maintaining generation quality by learning optimal time discretization strategies.
Learning to Explore and Exploit with GNNs for Unsupervised Combinatorial Optimization
Utku Umur ACIKALIN, Carla P Gomes
CodeOptimizationGraph Neural NetworkGraph
π― What it does: An unsupervised graph neural network framework named X2GNN is proposed for simultaneously exploring and exploiting in combinatorial optimization problems (maximum clique, maximum independent set, maximum cut) to generate high-quality solutions.
π― What it does: This paper proposes a Transformer-based world model called TWISTER, which uses action-conditioned contrastive predictive coding (AC-CPC) to learn high-quality temporal features, thereby enhancing the performance of model-based reinforcement learning.
Anders Aamand (University of Copenhagen), Hao WU (University of Waterloo)
CodeRecurrent Neural NetworkVideo
π― What it does: This paper proposes two learning-enhanced streaming algorithms, Misra-Gries and Frequent Directions, which guide memory allocation using predictors to achieve lower errors in frequency estimation and high-dimensional direction estimation tasks.
Learning-Guided Rolling Horizon Optimization for Long-Horizon Flexible Job-Shop Scheduling
Sirui Li (Massachusetts Institute of Technology), Cathy Wu (Massachusetts Institute of Technology)
CodeOptimizationAuto EncoderTabular
π― What it does: A learning-based rolling horizon optimization framework L-RHO is proposed to accelerate and improve the solution of long-slot combinatorial optimization problems, such as flexible job shop scheduling.
π― What it does: A lesion-focused diffusion model named LeFusion has been developed, capable of synthesizing images/annotations with lesions from normal medical images. It achieves fine control over lesion size, location, texture, and category through texture histogram control, multi-channel decomposition, and lesion mask diffusion.
π― What it does: This paper proposes a graph feature selection method called GFS, which uses a new metric TFI to distinguish between beneficial and detrimental feature dimensions for graph convolution, significantly improving the node classification performance of various GNNs.
π― What it does: This paper proposes the technique of 'omniscient trajectory' to derive an explicit mutual information theoretical generalization bound that utilizes the flatness of SGD, and based on this, obtains a lower bound of O(1/βn) for GD on the CLB problem;
Leveraging Submodule Linearity Enhances Task Arithmetic Performance in LLMs
Rui Dai (National Engineering Laboratory for Brain Inspired Intelligence Technology and Application, University of Science and Technology of China), Jieping Ye (Independent Researcher)
CodeTransformerLarge Language ModelText
π― What it does: This paper finds through statistical analysis that, although the overall model lacks linearity, its sub-modules (layers, attention, MLP, etc.) exhibit high linearity. It then proposes a training-free task arithmetic model merging method: first, the model is split into sub-modules, and the closed-form optimal merging weights are derived using the linear characteristics of the sub-modules. Subsequently, the sub-modules are linearly merged to enhance the performance of multi-task models.
Leveraging Variable Sparsity to Refine Pareto Stationarity in Multi-Objective Optimization
Zeou Hu (University of Waterloo), Yaoliang Yu (University of Waterloo)
CodeOptimizationFederated LearningImageTabular
π― What it does: This paper proposes the concept of Refined Pareto Stability (RPS) using a function-variable sparse structure and designs the RP-MGDA algorithm based on this to solve multi-objective optimization problems.
Zhuorui Ye (Tsinghua University), Fei Fang (Carnegie Mellon University)
CodeExplainability and InterpretabilityReinforcement LearningVision Language ModelImage
π― What it does: Under a limited budget for concept annotation, a new training framework called LICORICE is proposed, enabling reinforcement learning agents to learn interpretable concept bottleneck policies from a small amount of annotated data, achieving performance that is comparable to or even better than traditional baselines.
π― What it does: This paper proposes the Synthetic Coordinate Embedding (SYCO) framework, which maps two-dimensional molecular graphs to three-dimensional Euclidean point clouds, and uses diffusion models to generate molecular graphs in this latent space, resulting in a new method for molecular graph generation.
π― What it does: The Latent Optimal Linear combinations (LOL) method is proposed to construct linear combinations that satisfy the prior distribution in generative models, thereby achieving more reliable latent space interpolation and subspace definition.
π― What it does: A linear multi-step solver distillation framework is proposed, allowing the student solver to approximate the teacher solver's sampling trajectory with very few function evaluations (NFE), thus achieving fast and high-quality diffusion model sampling.
π― What it does: A linear spherical slice optimal transport (LSSOT) framework is proposed and implemented for rapid comparison of spherical probability distributions, applied to brain surface registration and point cloud interpolation.
RaphaΓ«l Sarfati (Cornell University), Christopher Earls (Cornell University)
CodeTransformerLarge Language ModelTextStochastic Differential Equation
π― What it does: By studying the internal token trajectories of large language models, it was found that they evolve along low-dimensional manifolds and can be approximated using a linear + noise model.
LiveBench: A Challenging, Contamination-Limited LLM Benchmark
Colin White (Abacus.AI), Micah Goldblum (Columbia)
CodeLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: A sustainable and pollution-resistant evaluation benchmark for LLMs, named LiveBench, has been constructed and released. It includes multiple categories (mathematics, coding, reasoning, language, instruction following, data analysis) tasks and achieves evaluation without LLM/human judgment through an automated, objective truth-based scoring system.
LLaMA-Omni: Seamless Speech Interaction with Large Language Models
Qingkai Fang (Institute of Computing Technology Chinese Academy of Sciences), Yang Feng (Institute of Computing Technology Chinese Academy of Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextAudio
π― What it does: LLaMA-Omni is proposed, an end-to-end low-latency speech-to-text interaction model that can generate both text and speech responses simultaneously without relying on ASR.
LLaRA: Supercharging Robot Learning Data for Vision-Language Policy
Xiang Li (Stony Brook University), Michael S Ryoo
CodeRobotic IntelligenceTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
π― 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-Mini: Efficient Image and Video Large Multimodal Models with One Vision Token
Shaolei Zhang (Chinese Academy of Sciences), Yang Feng (Chinese Academy of Sciences)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageVideoMultimodality
π― 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.
LLM Unlearning via Loss Adjustment with Only Forget Data
Yaxuan Wang (University of California), Wei Wei (Accenture)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― 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)
CodeTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― 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)
CodeLarge Language ModelPrompt EngineeringPhysics Related
π― 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)
CodeRecognitionObject DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodality
π― 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)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A learning-based framework LLMOPT is constructed to automatically define and solve multi-type optimization problems from natural language descriptions.
CodeGenerationRepresentation LearningTransformerLarge Language ModelText
π― 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.
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)
CodeAnomaly DetectionPrompt EngineeringVision Language ModelContrastive LearningImage
π― 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.
π― 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.
Locally Connected Echo State Networks for Time Series Forecasting
Filip Matzner (Charles University), FrantiΕ‘ek MrΓ‘z (Charles University)
CodeTime Series
π― 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.
π― 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.
Logical Consistency of Large Language Models in Fact-Checking
Bishwamittra Ghosh (Max Planck Institute for Software Systems), Arijit Khan (Aalborg University)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
π― 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)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― 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.
LOKI: A Comprehensive Synthetic Data Detection Benchmark using Large Multimodal Models
Junyan Ye (Sun Yat-sen University), Weijia Li (Sun Yat-sen University)
CodeData SynthesisAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageVideoTextMultimodalityBenchmarkChain-of-ThoughtAudio
π― 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.
Long-horizon Visual Instruction Generation with Logic and Attribute Self-reflection
Yucheng Suo (Zhejiang University), Yi Yang (Zhejiang University)
CodeGenerationTransformerLarge Language ModelDiffusion modelTextMultimodality
π― 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)
CodeRecommendation SystemSequential
π― 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.
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)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― 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.
LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs
Yushi Bai (Tsinghua University), Juanzi Li (Tsinghua University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmark
π― 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.
π― 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.
Ying Fan (University of Wisconsin Madison), Kangwook Lee (UC Berkeley)
CodeTransformerSequential
π― 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.
π― 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.
π― 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.
π― 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.
π― 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)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― 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.
π― 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.
MACPO: Weak-to-Strong Alignment via Multi-Agent Contrastive Preference Optimization
Yougang Lyu (University of Amsterdam), Zhaochun Ren (Leiden University)
CodeOptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningContrastive LearningText
π― 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.
π― 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)
CodeKnowledge DistillationRepresentation LearningTransformerBiomedical Data
π― 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.
MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation
Zhaoning Yu (Iowa State University), Hongyang Gao (Iowa State University)
CodeExplainability and InterpretabilityDrug DiscoveryGraph Neural NetworkAuto EncoderGraphBiomedical Data
π― 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.
MAI: A Multi-turn Aggregation-Iteration Model for Composed Image Retrieval
Yanzhe Chen (Peking University), Yuxin Peng (University of Science and Technology Beijing)
CodeRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― 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.
CodeRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
π― 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)
CodeRetrievalOptimizationTransformerLarge Language ModelText
π― 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.
π― 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.
π― 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.
π― 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;
Ziyue Li (University of Maryland), Tianyi Zhou (University of Maryland)
CodeOptimizationFederated LearningText
π― What it does: The MosT framework is proposed, which simultaneously trains m models in multi-objective optimization where nβ«m, allowing them to form diverse and complementary solutions on the Pareto front, thereby covering all objectives.
Xinran Wang (University of Minnesota), Ali Anwar (University of Minnesota)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: A Multi-Human-Value Alignment Palette (MAP) method is proposed, allowing for a one-time alignment of generative AI according to user-specified multi-dimensional value objectives while maintaining the model's original distribution.
MAPS: Advancing Multi-Modal Reasoning in Expert-Level Physical Science
Erle Zhu (Tsinghua University), Hongning Wang (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodalityPhysics RelatedChain-of-Thought
π― What it does: The MAPS framework is proposed, which combines multimodal large language models with physical perception models and simulators, using chain simulation to improve the reasoning accuracy of circuit analysis problems.
π― What it does: A fast sampler MRSampler based on semi-analytical solutions is proposed to accelerate the sampling process of Mean Reverting Diffusion.
MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model
Junjie Li (Microsoft Research Asia), Jiang Bian (Microsoft Research Asia)
CodeGenerationTransformerLarge Language ModelReinforcement LearningTime SeriesSequentialFinance Related
π― What it does: This paper proposes a large-scale order-level financial market simulation foundational model (LMM) and a financial market simulation engine (MarS) based on LMM, capable of generating realistic order flows across three dimensions: high resolution, controllability, and interactivity. It supports various downstream financial tasks such as forecasting, risk detection, impact analysis, and reinforcement learning.
CodeOptimizationTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes and implements Mask-DPO, a direct preference optimization method that combines sentence-level factual masking to reduce hallucinations in large language models.
π― What it does: A program planning framework based on diffusion models, MTID, is proposed, which utilizes intermediate latent visual features for temporal interpolation and restricts the action space through a masking mechanism, ultimately generating a coherent action sequence that aligns with task objectives from initial and final visual observations.
π― What it does: A task arithmetic regularization method based on ΟβJacobian product (Ο Jp) is proposed and validated, significantly reducing task interference and enhancing model editing effects;
π― What it does: The study investigates Test-Time Adaptation (TTA) of graph neural networks in environments with structural shifts and proposes a new framework called Matcha.
MatExpert: Decomposing Materials Discovery By Mimicking Human Experts
Qianggang Ding (Universite de Montreal), Bang Liu (Universite de Montreal)
CodeGenerationRetrievalOptimizationExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningTextChain-of-Thought
π― What it does: The MatExpert framework is proposed, breaking down material discovery into three stages: retrieval, transformation, and generation, using LLM and contrastive learning to generate solid materials that meet user attributes.
MathGAP: Out-of-Distribution Evaluation on Problems with Arbitrarily Complex Proofs
Andreas Opedal (ETH Zurich), Mrinmaya Sachan (ETH Zurich)
CodeTransformerLarge Language ModelTextChain-of-Thought
π― What it does: The MathGAP framework is proposed, utilizing an automatic generation method for controllable proof tree structures to evaluate the OOD generalization ability of LLMs on arbitrarily complex arithmetic proofs.
MatryoshkaKV: Adaptive KV Compression via Trainable Orthogonal Projection
Bokai Lin (Shanghai Jiao Tong University), Zhijie Deng (Shanghai Jiao Tong University)
CodeCompressionKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper compresses the feature dimensions of the KV cache of large language models by training a learnable orthogonal projection matrix, utilizing PCA initialization followed by knowledge distillation and a Matryoshka training strategy, ultimately achieving an adaptive heterogeneous compression rate that significantly reduces KV cache usage.
MAVIS: Mathematical Visual Instruction Tuning with an Automatic Data Engine
Renrui Zhang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
CodeData SynthesisOptimizationTransformerLarge Language ModelContrastive LearningImageTextMultimodalityChain-of-Thought
π― What it does: This paper presents an automated data engine and a four-stage training pipeline named MAVIS, specifically designed to enhance the visual mathematical reasoning capabilities of multimodal large language models (MLLMs);