IPSI: Enhancing Structural Inference with Automatically Learned Structural Priors
Zhongben Gong (Shenzhen University), Mingyang Zhou (Shenzhen University)
CodeRecurrent Neural NetworkGraph Neural NetworkAuto EncoderGraphTime SeriesPhysics Related
🎯 What it does: An iterative pre-training structure inference framework IPSI is proposed, which alternately updates a pre-trained structure estimator and a VAE-based joint inference module to improve the accuracy of structure inference in interactive dynamic systems.
🎯 What it does: Proposes π-Quant, a complex rotation algorithm that utilizes the properties of the irrational number π to compress the optimizer state to a low bit-width (as low as 3.32-bit) while maintaining full precision;
Is Limited Participant Diversity Impeding EEG-based Machine Learning?
Philipp Bomatter (University of Edinburgh), Henry Gouk (University of Edinburgh)
CodeTime SeriesBiomedical DataAlzheimer's Disease
🎯 What it does: A systematic quantitative study on the impact of sample partitioning and participant diversity in EEG machine learning, and an evaluation of the effectiveness of data augmentation and self-supervised pre-training under different data scales.
It’s Hard to Be Normal: The Impact of Noise on Structure-agnostic Estimation
Jikai Jin (Stanford University), Vasilis Syrgkanis (Stanford University)
Code
🎯 What it does: This paper studies the impact of treatment noise distribution on the performance of Structural Agnostic Estimation (SAE) in certain linear models. It proves that under Gaussian noise, Double Machine Learning (DML) has reached the lower bound, while under non-Gaussian noise, higher-order orthogonal moment functions can be constructed to achieve higher-order robustness, and proposes the ACE (Agnostic Cumulant-based Estimation) estimator.
Jacobian-Based Interpretation of Nonlinear Neural Encoding Model
Xiaohui Gao (Northwestern Polytechnical University), Xintao Hu (Northwestern Polytechnical University)
CodeExplainability and InterpretabilityImageMagnetic Resonance Imaging
🎯 What it does: This paper proposes and validates a nonlinear explanation metric JNE based on Jacobian dispersion to quantify the degree of nonlinearity in neural encoding models predicting BOLD signals, revealing hierarchical and stimulus-specific nonlinear distributions in the visual cortex.
JADE: Joint Alignment and Deep Embedding for Multi-Slice Spatial Transcriptomics
Yuanchuan Guo (Harvard University), Ying Ma (Brown University)
CodeRepresentation LearningGraph Neural NetworkAuto EncoderContrastive LearningBiomedical Data
🎯 What it does: This paper proposes the JADE model, which can simultaneously perform spatial alignment and low-dimensional embedding learning for multi-slice spatial transcriptomic data.
🎯 What it does: This paper studies the semantic correspondence problem and finds that supervised methods have poor generalization ability on unseen keypoints. To address this issue, the authors utilize monocular depth estimation to elevate 2D keypoints to 3D space, learning a category-level continuous 3D canonical manifold. They achieve an extension of sparse supervision through geometric consistency and, based on this, construct a new SPair-U evaluation benchmark to test the model's performance on unseen keypoints.
🎯 What it does: The JAMUN model is constructed to generate conformational ensembles of peptide fragments by performing Langevin dynamics (walk) and denoising (jump) in a noise-added space.
JanusDNA: A Powerful Bi-directional Hybrid DNA Foundation Model
Qihao Duan (Berlin Institute of Health), Benjamin Wild (Berlin Institute of Health)
CodeMixture of ExpertsBiomedical DataBenchmark
🎯 What it does: A bidirectional foundational model for DNA, JanusDNA, has been developed, capable of processing sequences of up to 1 Mbp at single nucleotide resolution and achieving efficient global modeling.
Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search
Yuxian Gu (NVIDIA), Han Cai (NVIDIA)
CodeOptimizationComputational EfficiencyKnowledge DistillationNeural Architecture SearchTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes Jet-Nemotron, a hybrid architecture language model that combines full attention and linear attention, and quickly transforms the pre-trained full attention model through PostNAS (Post Neural Architecture Search);
🎯 What it does: An end-to-end protein surface and structure joint generation framework called PepBridge is designed, which directly drives the generation of ligand surfaces and complete scaffolds using receptor surface point clouds.
🎯 What it does: A non-autoregressive relational database generation method based on a graph conditional diffusion model (GRDM) is proposed, which can jointly generate row attributes and primary-foreign key structures for multiple tables.
🎯 What it does: This paper proposes a Velocity-Growth Flow Matching (VGFM) framework that utilizes semi-relaxed optimal transport theory to jointly learn the velocity field of single-cell state transitions and the growth function of cell mass, thereby reconstructing the time evolution dynamics of single cells.
Jury-and-Judge Chain-of-Thought for Uncovering Toxic Data in 3D Visual Grounding
Kaixiang Huang (Zhejiang University), Shengfeng He (Singapore Management University)
CodeAnomaly DetectionTransformerLarge Language ModelPoint CloudChain-of-Thought
🎯 What it does: A Refer-Judge framework is constructed, utilizing multi-perspective and multi-role Chain-of-Thought to evaluate 3D visual localization data, identifying and filtering 'poisonous samples'.
Just One Layer Norm Guarantees Stable Extrapolation
Juliusz Ziomek (University of Oxford), Michael A Osborne
CodeProtein Structure PredictionTabularBiomedical Data
🎯 What it does: This study investigates the impact of adding a single layer of LayerNorm on the out-of-distribution performance of models in infinitely wide networks, providing both theoretical proof and experimental validation.
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The K-DECORE framework is proposed, achieving continuous learning for heterogeneous structured knowledge reasoning tasks through knowledge decoupling, dual-view memory, and structure-guided pseudo-query synthesis.
🎯 What it does: This paper proposes KaRF, a weakly supervised Kolmogorov-Arnold Network (KAN) driven two-stage radiance field for achieving high-quality local color editing.
🎯 What it does: An improved A*-based search algorithm called KeeA* is proposed, which enhances search efficiency and solution quality by introducing 'cluster scouting' and 'path-aware' sampling to calibrate the selection of candidate nodes.
Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning
Yaxin Hou (Southeast University), Junhui Hou (City University of Hong Kong)
CodeClassificationContrastive LearningImage
🎯 What it does: A controllable pseudo-label generation framework (CPG) is proposed to address the challenge of unknown distribution of unlabeled data in long-tail semi-supervised learning; it dynamically filters controllable labels, iteratively constructs a label set with known distribution, and uses logit adjustment to obtain a Bayesian optimal classifier.
KGGen: Extracting Knowledge Graphs from Plain Text with Language Models
Belinda Mo (Stanford University), Sanmi Koyejo (Stanford University)
CodeGenerationRetrievalTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: A method has been developed to automatically extract knowledge graphs (KGGen) from raw text using large language models (LLM), and a new evaluation benchmark MINE (MINE-1 and MINE-2) has been proposed to measure the quality of KG generation.
🎯 What it does: A memory model called Kinaema based on recursive transformers is proposed to integrate visual observations and perform relative pose estimation in continuous robotic operations, which is then embedded into the DEBiT navigation agent to achieve the Mem-Nav task.
🎯 What it does: Proposed KL-Adaptive Stability Sampling (KLASS), a training-free, token-level KL divergence and confidence-based parallel decoding method.
Knee-Deep in C-RASP: A Transformer Depth Hierarchy
Andy Yang (University of Notre Dame), David Chiang (University of Notre Dame)
CodeTransformerSequential
🎯 What it does: Investigate the impact of Transformer depth on expressive power, proving that fixed-precision Transformers are equivalent to C-RASP/TL[#] and establishing a strict depth hierarchy;
Know Thyself by Knowing Others: Learning Neuron Identity from Population Context
Vinam Arora (University of Pennsylvania), Eva L Dyer
CodeClassificationRepresentation LearningTransformerContrastive LearningBiomedical Data
🎯 What it does: This work proposes a self-supervised framework called NuCLR to learn neuron identity representations from neuronal population activity, supporting zero-shot classification.
Knowledge Distillation Detection for Open-weights Models
Qin Shi (Purdue University), Raymond A. Yeh (Purdue University)
CodeKnowledge DistillationImageText
🎯 What it does: This paper proposes a knowledge distillation detection task, aiming to determine whether a student model originates from a certain teacher based solely on the student model weights and the teacher API.
🎯 What it does: A Gaussian distribution distillation method based on the Deep Latent Factor model (DLF) is proposed to compress deep ensembles while preserving uncertainty.
🎯 What it does: This paper proposes a new Knowledge Insulation training method to transform large-scale pre-trained Vision-Language Models (VLM) into real-time continuous action control Vision-Language-Action (VLA) models, balancing fast training, real-time inference, and better generalization capabilities.
Knowledge Starts with Practice: Knowledge-Aware Exercise Generative Recommendation with Adaptive Multi-Agent Cooperation
Yangtao Zhou (Xidian University), Qingshan Li (Xidian University)
CodeRecommendation SystemTransformerLarge Language ModelGenerative Adversarial NetworkTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the Knowledge-Aware Exercise Generation Recommendation (KEGR) task and constructs a multi-agent collaborative framework based on LLM called ExeGen, which dynamically perceives students' knowledge states and generates personalized exercises.
KOALA++: Efficient Kalman-Based Optimization with Gradient-Covariance Products
Zixuan Xia (University of Bern), Paolo Favaro (University of Bern)
CodeOptimizationImageText
🎯 What it does: KOALA++, a scalable optimization algorithm based on Kalman filtering, is proposed to explicitly model the structured gradient uncertainty in neural network training.
KScope: A Framework for Characterizing the Knowledge Status of Language Models
Yuxin Xiao (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)
CodeLarge Language ModelTextBiomedical Data
🎯 What it does: Five types of knowledge states based on knowledge consistency and correctness are proposed, and a hierarchical verification framework KScope is constructed to provide a refined description and classification of the knowledge states of large language models in question-answering tasks.
KTAE: A Model-Free Algorithm to Key-Tokens Advantage Estimation in Mathematical Reasoning
Wei Sun (Institute of Automation Chinese Academy of Sciences), Jiajun Zhang (Institute of Automation Chinese Academy of Sciences)
CodeLarge Language ModelTextBenchmark
🎯 What it does: A model-free Keyword Advantage Estimation (KTAE) algorithm is proposed, which generates fine-grained token-level advantage signals by statistically associating correct and incorrect answers in the sampled rounds.
🎯 What it does: A periodic domain score-based generative model based on Kuramoto coupling dynamics is proposed, achieving directionally rich image generation through a synchronization-desynchronization process.
KVCOMM: Online Cross-context KV-cache Communication for Efficient LLM-based Multi-agent Systems
Hancheng Ye (Duke University), Yiran Chen (Duke University)
CodeTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes the KVCOMM framework, which dynamically reuses the KV cache of multiple agent LLMs using an anchor pool, significantly reducing pre-filling latency;
KVLink: Accelerating Large Language Models via Efficient KV Cache Reuse
Jingbo Yang (University of California Santa Barbara), Shiyu Chang (University of California Santa Barbara)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the KVLINK method, which precomputes the KV cache of documents in large language models and reuses it during inference to avoid redundant encoding.
KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction
Jang-Hyun Kim (Seoul National University), Hyun Oh Song (Seoul National University)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes KVzip, a query-independent KV cache compression method that evaluates the importance of KV by utilizing LLM to reconstruct the context, thereby compressing the KV cache while maintaining model performance.
L-MTP: Leap Multi-Token Prediction Beyond Adjacent Context for Large Language Models
Xiaohao Liu (National University of Singapore), Tat-Seng Chua (National University of Singapore)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a Leap Multi-Token Prediction (L-MTP) method, which extends Multi-Token Prediction (MTP) to simultaneously enhance inference speed and model performance.
🎯 What it does: A modeling method for spatial dynamic systems based on latent space encoding is proposed—LAM-SLIDE, which tracks individual entities in a fixed-size latent system representation using assignable identifiers (IDs) and predicts future trajectories in the latent space through flow models.
Language Model Behavioral Phases are Consistent Across Architecture, Training Data, and Scale
James A. Michaelov (Massachusetts Institute of Technology), Ben Bergen
CodeTransformerLarge Language ModelText
🎯 What it does: A systematic analysis of the behavior of autoregressive language models across multiple architectures (Transformer, Mamba, RWKV), scales (1.4M–12B parameters), and datasets (OpenWebText, The Pile) during the pre-training process is conducted, proposing and validating three simple heuristics—word frequency, n-gram probability, and semantic similarity—that can explain their prediction variance.
Language Models (Mostly) Know When to Stop Reading
Roy Xie (Duke University), Bhuwan Dhingra (Duke University)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study proposes a dynamic context truncation method that allows large language models to self-stop reading after obtaining sufficient information, thereby improving reasoning efficiency.
Dhananjay Ashok (University of Southern California), Jonathan May (University of Southern California)
CodeClassificationGenerationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Training a linear probe utilizes the hidden states of the input layer of the language model to predict the overall behavior of the model before generating any tokens, and combines Conformal Prediction to build an early warning system;
Language Models can Self-Improve at State-Value Estimation for Better Search
Ethan Mendes (Georgia Institute of Technology), Alan Ritter (Georgia Institute of Technology)
CodeTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: A self-supervised framework called STL is proposed, which utilizes language models to generate actions, outcomes, and reasoning in a single-step lookahead to improve state value estimation and enhance search performance without the need for rewards or demonstrations.
Language Ranker: A Lightweight Ranking framework for LLM Decoding
Chenheng Zhang (Peking University), Zhouchen Lin (Peking University)
CodeRecommendation SystemTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The study proposes Language Ranker, a lightweight ranking framework that reorders candidate responses based on features extracted from a base model, thereby enhancing the decoding performance of LLMs.
Language‑Bias‑Resilient Visual Question Answering via Adaptive Multi‑Margin Collaborative Debiasing
Huanjia Zhu (Beijing Institute of Technology), Bingzhi Chen (Beijing Institute of Technology)
CodeClassificationRecognitionTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: This study investigates the formation mechanism of language bias in Visual Question Answering (VQA) and proposes an Adaptive Multi-Angle Boundary (MMCD) framework that combines frequency, confidence, and difficulty as three adaptive angle boundaries with difficulty-aware contrastive learning to dynamically reshape decision boundaries to suppress language bias.
LaRes: Evolutionary Reinforcement Learning with LLM-based Adaptive Reward Search
Pengyi Li (Tianjin University), Jianye HAO
CodeRobotic IntelligenceLarge Language ModelReinforcement LearningSequential
🎯 What it does: This paper proposes the LaRes framework, which generates and evolves reward functions through LLM and combines evolutionary reinforcement learning to achieve efficient policy learning.
Large Language Models as End-to-end Combinatorial Optimization Solvers
Xia Jiang (Eindhoven University of Technology), Yingqian Zhang (Eindhoven University of Technology)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: An end-to-end combinatorial optimization solving framework based on LLM is proposed, enabling LLMs to directly generate feasible high-quality solutions from natural language descriptions through supervised fine-tuning and constraint- and optimality-friendly reinforcement learning.
Large Language Models Think Too Fast To Explore Effectively
Lan Pan (Georgia Institute of Technology), Robert Wilson
CodeTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study investigates the performance of LLMs in open-ended exploration tasks (Little Alchemy 2) and compares it with human performance.
LARGO: Latent Adversarial Reflection through Gradient Optimization for Jailbreaking LLMs
Ran Li (Columbia University), Chengzhi Mao (Rutgers University)
CodeOptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The LARGO framework is proposed, utilizing gradient optimization in the continuous latent space of LLMs, and then generating smooth jailbreak suffixes through self-reflective decoding to achieve more efficient and natural attacks.
LASeR: Learning to Adaptively Select Reward Models with Multi-Arm Bandits
Duy Nguyen (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The LASER method is proposed, which utilizes a multi-armed bandit dynamic selection of the reward model (Reward Model) to choose the most suitable RM for each training batch or instance, generating preference pairs for iterative fine-tuning of LLM;
Guohao Sun (Rochester Institute of Technology), Zhiqiang Tao (Rochester Institute of Technology)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a latent chain reasoning method (LaCoT) based on variational inference, which samples diverse visual reasoning chains and generates answers through an autoregressive strategy.
🎯 What it does: The FVAE-LoRA method is proposed, which combines Variational Autoencoders (VAE) with LoRA, using two separate latent spaces (z1 responsible for task-related information and z2 responsible for remaining information) to achieve low-rank fine-tuning.
LaViDa: A Large Diffusion Model for Vision-Language Understanding
Shufan Li (UCLA), Aditya Grover (Adobe Research)
CodeRecognitionGenerationTransformerVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: LaViDa is proposed, a visual language model based on discrete diffusion models, capable of parallel decoding, bidirectional reasoning, and controllable text generation.
Layer as Puzzle Pieces: Compressing Large Language Models through Layer Concatenation
Fei Wang (South China University of Technology), Changxing Ding (South China University of Technology)
CodeCompressionKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposes the CoMe framework for hierarchical pruning and post-training recovery of large language models, reducing the number of parameters while maintaining performance;
Layer-wise Update Aggregation with Recycling for Communication-Efficient Federated Learning
Jisoo Kim (Inha University), Sunwoo Lee (Inha University)
CodeFederated LearningText
🎯 What it does: A hierarchical update recycling mechanism called FedLUAR is proposed, which uses the gradient-weight ratio to select less important layers and reuses old gradients to significantly reduce communication costs.
LayerCraft: Enhancing Text-to-Image Generation with CoT Reasoning and Layered Object Integration
Yuyao Zhang, Yu-Wing Tai
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageTextChain-of-Thought
🎯 What it does: Proposes the LayerCraft framework, using large language models (LLM) as autonomous agents to achieve structured generation from text to image and layered object integration, supporting batch synthesis and multi-round editing.
LayerNavigator: Finding Promising Intervention Layers for Efficient Activation Steering in Large Language Models
Hao Sun (Institute of Information Engineering), Yanan Cao (Institute of Information Engineering)
CodeTransformerLarge Language ModelText
🎯 What it does: This paper proposes LayerNavigator, a method for selecting activation-guided layers based on layer guidability scores to enhance the behavioral alignment of large language models.
🎯 What it does: An adaptive training strategy called learn2mix is proposed in resource-constrained environments, which accelerates the convergence of neural networks by dynamically adjusting the class ratio in the batch.
🎯 What it does: A Burst-Encodable Time-of-Flight (BE-ToF) system is proposed and implemented, achieving long-distance high-precision depth perception with a single-frequency pulse bundle mode.
Learned Prefix Caching for Efficient LLM Inference
Dongsheng Yang (Princeton University), Wyatt Lloyd (Princeton University)
CodeComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper studies a learning-based prefix caching strategy (LPC) that enhances the inference efficiency of LLMs by predicting whether a session will continue.
Learning “Partner-Aware” Collaborators in Multi-Party Collaboration
Abhijnan Nath (Colorado State University), Nikhil Krishnaswamy (Colorado State University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: The Interruptible Collaborative Roleplayer (ICR) model is proposed, utilizing counterfactual KL regularization with adversarial prefixes to enable large language models to distinguish between useful and misleading interventions in multi-party collaboration, thereby achieving 'partner-aware' collaboration.
🎯 What it does: Proposes the AniDS framework, which utilizes a structure-aware full covariance noise generator for 3D molecular coordinate denoising, thereby learning molecular force fields.
🎯 What it does: A Plug-and-Play method for the Poisson inverse problem based on the cocoercive conservative (CoCo) assumption is proposed, constructing a CoCo denoiser and proving it to be a proximal operator of a weakly convex prior. Subsequently, two convergent PnP algorithms, CoCo-ADMM and CoCo-PEGD, are presented.
Learning Crossmodal Interaction Patterns via Attributed Bipartite Graphs for Single-Cell Omics
Xiaotang Wang (Hong Kong University of Science and Technology), Yongqi Zhang (Hong Kong University of Science and Technology)
CodeGraph Neural NetworkTransformerGraphBiomedical Data
🎯 What it does: This paper proposes a method to reformulate the single-cell cross-modal matching problem as a graph classification task on an Attributed Bipartite Graph (ABG), and uses the Bi 2 Former model to learn ATAC-RNA interactions.
Learning Dense Hand Contact Estimation from Imbalanced Data
Daniel Sungho Jung (Seoul National University), Kyoung Mu Lee (Seoul National University)
CodePose EstimationTransformerImage
🎯 What it does: A framework named HACO is proposed for learning dense hand contact estimation from hand contact data with category and spatial imbalance.
Learning Equilibria from Data: Provably Efficient Multi-Agent Imitation Learning
Till Freihaut (University of Zurich), Giorgia Ramponi
CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: The study focuses on learning Nash equilibria using expert data in zero-sum Markov games, providing a sample complexity analysis for behavior cloning (BC) and two interactive multi-agent imitation learning methods (MAIL-BRO, MURMAIL).
Learning Expandable and Adaptable Representations for Continual Learning
Ruilong Yu (University of Electronic Science and Technology of China), shijie zhou
CodeRepresentation LearningTransformerMixture of ExpertsImage
🎯 What it does: A scalable and adaptive representation learning framework called LEAR is designed to achieve multi-domain continual learning using a collaborative backbone and expert networks.
Learning from Demonstrations via Capability-Aware Goal Sampling
Yuanlin Duan (Rutgers University), He Zhu (Rutgers University)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningWorld ModelSequential
🎯 What it does: This paper proposes a learning method called Cago, which dynamically tracks the capability boundaries of agents on demonstration trajectories to sample intermediate sub-goals that guide exploration, thereby achieving more efficient learning in long-horizon tasks with sparse rewards.
🎯 What it does: For regression tasks with only a given target interval, this paper proposes projection loss and worst-case loss, and provides non-asymptotic generalization bounds.
Learning Grouped Lattice Vector Quantizers for Low-Bit LLM Compression
Xi Zhang (Nanyang Technological University), Weisi Lin (Nanyang Technological University)
CodeCompressionLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A Group Lattice Vector Quantization (GLVQ) framework is designed to achieve post-training quantization of large language models at extremely low bit widths (≤3bit) while maintaining or improving performance.
Jiajun Hong (Zhejiang University), Wenguan Wang (Zhejiang University)
CodeRecognitionObject DetectionTransformerVision Language ModelImage
🎯 What it does: The GroupHOI framework is proposed, treating human-object interactions as groups, constructing groups using geometric proximity and semantic similarity, and achieving relational reasoning through local context aggregation and global transformers.
Learning Individual Behavior in Agent-Based Models with Graph Diffusion Networks
Francesco Cozzi (Sapienza University), Corrado Monti (CENTAI)
CodeGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A differentiable surrogate model based on Graph Diffusion Network (GDN) is proposed, which can learn individual-level stochastic behaviors from the data of the original Agent-Based Model (ABM) and generate future states.
🎯 What it does: The Diversity-regularized Actor Critic (DrAC) algorithm is proposed, which utilizes reparameterization and distance-based multimodal policy diversity regularization to train intractable multimodal decision policies.
🎯 What it does: Proposes an EM method based on the Jarzynski-adjusted Langevin algorithm (JALA-EM) to estimate the maximization of marginal likelihood for latent variable models.
🎯 What it does: A forward learning algorithm based on Hamiltonian systems, RHEL, is proposed to train recursive state space models (SSM), achieving gradient computation without explicit Jacobian matrices and using only forward propagation.
Learning Multi-Source and Robust Representations for Continual Learning
Fei Ye (University of Electronic Science and Technology of China), shijie zhou
CodeRepresentation LearningTransformerImage
🎯 What it does: A continuous learning framework LMSRR based on multi-source pre-trained ViT is proposed, which integrates multi-source features and enhances the model's adaptability to new tasks through dynamic fusion.
🎯 What it does: Train a network that can directly output normalized energy (log-likelihood) by learning the probability distribution of image data through dual score matching.
Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift
Yanru Sun (Tianjin University), Min Wu (Astar Research Institute)
CodeDomain AdaptationTransformerMixture of ExpertsTime Series
🎯 What it does: A TFPS (Time-Frequency Pattern Specific) framework is proposed, utilizing dual-domain encoding, subspace clustering pattern recognition, and a mixture of experts network to achieve dynamic expert allocation for time series patches, resulting in more accurate long-term predictions.
Learning Provably Improves the Convergence of Gradient Descent
Qingyu Song (Xiamen University), Hong Xu (Chinese University of Hong Kong)
CodeOptimizationTabular
🎯 What it does: This paper proposes a Learning to Optimize (L2O) framework that accelerates the optimization process by learning the hyperparameters of gradient descent (GD) and proves the training convergence of this framework.
🎯 What it does: In the multimodal federated learning scenario, the PEPSY framework is proposed, which constructs data missing profiles by learning the missing pattern features of each client, to reconfigure local models so that they can produce complete and aligned representations even in environments where missing modalities and missing features coexist.
Learning Relative Gene Expression Trends from Pathology Images in Spatial Transcriptomics
Kazuya Nishimura (National Cancer Center Japan), Yasuhiro Kojima (National Cancer Center Japan)
CodeConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: The paper proposes a gene expression relative trend learning method based on pathological images called STRank, which improves the robustness of traditional absolute value estimation.
Learning Robust Spectral Dynamics for Temporal Domain Generalization
En Yu (Australian Artificial Intelligence Institute University of Technology Sydney), Zhen Fang (Australian Artificial Intelligence Institute University of Technology Sydney)
CodeDomain AdaptationAuto EncoderTime Series
🎯 What it does: This paper studies the concept drift problem in Time Domain Generalization (TDG) and proposes the FreKoo framework, which achieves robust generalization to future unseen domains by decomposing model parameter trajectories in the frequency domain, using Koopman operators to predict low-frequency trends, and regularizing high-frequency noise.
🎯 What it does: A source-agnostic domain adaptation framework SVIP is proposed for visible-infrared person recognition, which transfers a pre-trained source model to an unlabeled, unpaired target domain.
Andrew Francesco Ilersich, Prasanth B. Nair (Institute for Aerospace Studies University of Toronto)
CodeMixture of ExpertsPhysics RelatedStochastic Differential Equation
🎯 What it does: By constructing macro and micro layers of hidden states and modeling them in the form of coupled stochastic differential equations (SDEs), a probabilistic model capable of directly predicting multi-scale dynamics from observational data has been developed.
🎯 What it does: This paper proposes a task-agnostic multi-teacher distillation framework that learns general representations by maximizing the mutual information between the student and the embeddings of multiple teachers.
Learning the Wrong Lessons: Syntactic-Domain Spurious Correlations in Language Models
Chantal Shaib (Northeastern University), Marzyeh Ghassemi (Massachusetts Institute of Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This study investigates the erroneous associations between grammatical templates and knowledge domains in the instruction-following process of large language models, revealing that models often rely on shallow grammatical-domain associations rather than true semantic understanding.
Fares El Khoury (Universite Grenoble Alpes), Michael Arbel (Universite Grenoble Alpes)
CodeOptimizationTabular
🎯 What it does: This paper studies the learning theory of Kernel Bilevel Optimization (KBO), providing a unified generalization error bound for value functions and gradients under finite samples, and verifying the convergence of gradient descent on synthetic instrumental variable regression tasks;
🎯 What it does: A Guided-ReST algorithm is proposed, which guides the search by inserting sub-goals of the optimal solution into the self-generated search trajectory, and fine-tunes using PPO based on this.
Nina S. Nellen (University Göttingen), Alexander S. Ecker (Max Planck Institute for Dynamics and Self-Organization)
CodeRepresentation LearningBiomedical Data
🎯 What it does: This study proposes a new deep embedding clustering method called DECEMber, aimed at improving the clustering consistency of neuronal functional embeddings by introducing clustering bias, thereby better identifying cell types in the mouse visual cortex.
Learning to Condition: A Neural Heuristic for Scalable MPE Inference
Brij Malhotra (University of Texas at Dallas), Vibhav Giridhar Gogate
CodeOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies an adaptive variable conditioning method based on neural networks, L2C, to accelerate MPE inference in probabilistic graphical models.
🎯 What it does: A new generative pre-training method FGP is proposed to enable ordinary neural network architecture encoders to learn the 'information flow' features of neural networks, thereby enhancing performance prediction and NAS effectiveness.
Learning to Focus: Causal Attention Distillation via Gradient‐Guided Token Pruning
Yiju Guo (Renmin University of China), Yankai Lin (Renmin University of China)
CodeKnowledge DistillationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A two-stage framework named Learning to Focus (LeaF) is proposed, which automatically identifies and removes distracting words through gradient comparison, achieving causal attention alignment during the distillation process, thereby enhancing the ability of large language models to focus on key information and improve reasoning accuracy in long-context reasoning, code generation, and multi-hop question answering.
🎯 What it does: A neural combinatorial optimization framework L2C-Insert based on the insertion paradigm is proposed to construct high-quality solutions for the vehicle routing problem (TSP and CVRP).
Learning to Instruct for Visual Instruction Tuning
Zhihan Zhou (Cooperative Medianet Innovation Center, Shanghai Jiao Tong University), Yanfeng Wang (School of Artificial Intelligence, Shanghai Jiao Tong University)
CodeRecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: By allowing the multimodal LLM to simultaneously learn to generate instructions corresponding to images and responses to regularize the Visual Instruction Tuning (VIT) process.