International Conference on Learning Representations · 1682 papers
Partially Observed Trajectory Inference using Optimal Transport and a Dynamics Prior
Anming Gu (Boston University), Kristjan Greenewald (IBM Research)
CodeTime SeriesStochastic Differential Equation
🎯 What it does: The paper proposes a method for inferring partially observed trajectories based on optimal transport and dynamic priors, aimed at recovering the temporal dynamics of a population from unpaired temporal marginal data.
Liang CHEN, Kam-Fai Wong (Chinese University of Hong Kong)
CodeGenerationOptimizationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A framework named PEARL is proposed, which utilizes distributed robust optimization and a learnable permutation generation network (P-Net) to enhance the robustness of large language models in few-shot scenarios regarding demonstration order arrangements, and achieves adaptive learning of adversarial arrangements through adversarial training.
CodeObject DetectionObject TrackingPose EstimationTransformerDiffusion modelSimultaneous Localization and MappingVideoMultimodalityPoint CloudBenchmark
🎯 What it does: A large-scale pedestrian motion reconstruction dataset PMR based on a mixed reality platform has been proposed and released, and various SOTA methods have been benchmarked from third-person, first-person, and LiDAR perspectives.
🎯 What it does: This paper proposes a text-guided periodic material generation model TGDMat, which integrates diffusion atomic types, coordinates, and lattices, and introduces text context during the denoising process.
🎯 What it does: A full-scene waveform generation model called PeriodWave is proposed, based on periodic-aware flow matching, achieving multi-period estimation, discrete wavelet multi-band modeling, and FreeU denoising technology, enhancing high-frequency information and periodic reproduction.
Haoyu Wang (Renmin University of China), Ji-Rong Wen (Renmin University of China)
CodeRetrievalTransformerLarge Language ModelText
🎯 What it does: This paper explains and verifies the source bias caused by the over-scoring of low perplexity (PPL) documents by pre-trained language model (PLM) retrievers through causal graphs, and proposes a causal diagnosis and correction-based debiasing method (CDC) to mitigate this bias during inference.
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes the concept of 'personalized alignment', constructs a large-scale real user personality questionnaire dataset PAPI, and develops a parameter-free activation search (PAS) method, enabling large language models to quickly and cost-effectively align with individual personality traits such as the Big Five and the Dark Triad while maintaining their original capabilities.
Renjie Pi (Hong Kong University of Science and Technology), Tong Zhang (University of Illinois Urbana-Champaign)
CodeGenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Designed and implemented the Personalized Visual Instruction Tuning (PVIT) framework, which enhances the performance of multimodal large language models (MLLMs) in conversations tailored to specific individuals through an automated data generation pipeline.
PersonalLLM: Tailoring LLMs to Individual Preferences
Thomas P Zollo, Hongseok Namkoong (Columbia University)
CodeMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: A public benchmark PersonalLLM is proposed to evaluate the personalized alignment performance of LLMs;
Jun-Yu Ma (University of Science and Technology of China), Jia-Chen Gu (University of California, Los Angeles)
CodeLarge Language ModelText
🎯 What it does: The PRUNE framework is proposed, which reduces the disturbance to existing knowledge by constraining the condition number of the edit matrix during the continuous model editing process, thereby maintaining the general capabilities of large language models after multiple edits.
🎯 What it does: This paper proposes PETRA, a framework that utilizes reversible networks to achieve model parallel training, allowing for the decoupling of forward and backward propagation while maintaining single-version parameters, thus enabling the parallelization of forward and backward gradient computations.
🎯 What it does: This paper proposes a training-free sampling acceleration method called PFDiff, which can be combined with any ODE solver. It utilizes past and future score information to skip time steps, thereby reducing the number of function evaluations and correcting discretization errors.
🎯 What it does: This paper proposes PhiNet, a brain-inspired model based on non-contrastive self-supervised learning, which draws on the hippocampal temporal prediction hypothesis and uses StopGradient to simulate synaptic delay.
PhyloLM: Inferring the Phylogeny of Large Language Models and Predicting their Performances in Benchmarks
Nicolas Yax (INSERM), Stefano Palminteri (INSERM)
CodeLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: The PhyloLM method is proposed, which uses evolutionary tree algorithms to infer the phylogenetic relationships between LLMs based on the token sequences output by LLMs and predict their benchmark performance.
🎯 What it does: A physical encoding message passing graph network (PhyMPGN) is designed and implemented for efficiently predicting the evolution of space-time PDE systems under coarse grids with only a small amount of training data.
Physics-aligned field reconstruction with diffusion bridge
Zeyu Li (Beihang University), Lijun Yang (Beihang University)
CodeDiffusion modelScore-based ModelTime SeriesPhysics Related
🎯 What it does: This paper proposes the Physics-aligned Schrödinger Bridge (PalSB) framework for reconstructing physical fields from sparse measurements.
CodeGenerationOptimizationDiffusion modelTabularPhysics Related
🎯 What it does: A physical information diffusion model (PIDM) is proposed, which incorporates PDE residual loss into the training of the diffusion model, ensuring that generated samples maintain diversity while satisfying physical constraints.
Physiome-ODE: A Benchmark for Irregularly Sampled Multivariate Time-Series Forecasting Based on Biological ODEs
Christian Klötergens (International School of Machine Learning and Learning Systems Volkswagen Financial Services Data Analytics Research Center University of Hildesheim), Lars Schmidt-Thieme (International School of Machine Learning and Learning Systems Volkswagen Financial Services Data Analytics Research Center University of Hildesheim)
🎯 What it does: A benchmark called Physiome-ODE is proposed and implemented, generating 50 irregular multivariate time series datasets using biological ODE models, and various IMTS prediction models are evaluated on this benchmark.
🎯 What it does: By introducing physical assumptions and prior knowledge, a new task framework for recovering and interpreting PDEs from experimental observation data (PDE interpretation) is proposed, and a mixed-integer programming solution based on decision forests is implemented.
🎯 What it does: A PianoMotion10M dataset containing 116 hours and 10 million frames of hand poses has been constructed, and a two-stage gesture generation baseline model based on position prediction and diffusion generation has been proposed.
PiCO: Peer Review in LLMs based on Consistency Optimization
Kun-Peng Ning (Peking University), Li Yuan (Peking University)
CodeOptimizationTransformerLarge Language ModelText
🎯 What it does: A completely unsupervised LLM evaluation framework called PiCO is designed, utilizing a peer review mechanism of LLM mutual evaluation to generate model rankings.
PIED: Physics-Informed Experimental Design for Inverse Problems
Apivich Hemachandra (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
CodeOptimizationMeta LearningReinforcement LearningTime SeriesPhysics Related
🎯 What it does: This paper proposes the PIED framework, which utilizes Physics-Informed Neural Networks (PINN) for both forward simulation and inverse problem solving, optimizing observation configurations (such as sensor locations) through gradient optimization for one-time experimental deployments.
PIORF: Physics-Informed Ollivier-Ricci Flow for Long–Range Interactions in Mesh Graph Neural Networks
Youn-Yeol Yu (Yonsei University), Noseong Park (KAIST)
CodeGraph Neural NetworkMeshGraphPhysics Related
🎯 What it does: A reconnection method that combines physical information with graph topology, called PIORF, is proposed to address the over-squashing problem in grid graph neural networks and significantly improve the accuracy of fluid dynamics simulations.
🎯 What it does: A generative adversarial imitation learning method named PN-GAIL is proposed, which can learn near-optimal policies in the presence of non-optimal examples by utilizing negative risk information and a small amount of demonstration data with confidence labels.
PnP-Flow: Plug-and-Play Image Restoration with Flow Matching
Ségolène Tiffany Martin (Technische Universität Berlin), Gabriele Steidl (Technische Universität Berlin)
CodeRestorationFlow-based ModelImage
🎯 What it does: An algorithm named PnP Flow Matching is proposed to solve inverse problems in image restoration, combining a pre-trained flow matching model with the PnP framework.
🎯 What it does: Constructed and released the POGEMA benchmark platform, which includes a fast learning environment, problem instance generator, visualization tools, and automatic evaluation tools for the unified assessment of multi-agent pathfinding (MAPF and LMAPF) learning, planning, and hybrid methods.
🎯 What it does: PointOBB-v2 is proposed, which generates pseudo-rotating boxes using only single-point supervision to accomplish directional object detection.
🎯 What it does: This paper proposes a training data poisoning attack called Poison-splat, which leverages the adjustable complexity characteristics of the 3D Gaussian Splatting system, resulting in extremely high GPU memory usage and time consumption during model training, potentially triggering service interruptions.
🎯 What it does: A polar-aware linear attention mechanism (Polafomer) is proposed to address the issues of insufficient representation caused by negative value loss and high attention entropy in traditional linear attention, significantly enhancing the performance and efficiency of visual Transformers.
Polynomial Composition Activations: Unleashing the Dynamics of Large Language Models
Zhijian Zhuo (Peking University), Jinwen Ma (Peking University)
CodeTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: A class of polynomial combination activation functions named PolyCom (PolyReLU, PolyNorm) is proposed and implemented, and its effectiveness is validated in large language models (1B dense models and 1B/7B Mixture of Experts models).
Population Transformer: Learning Population-level Representations of Neural Activity
Geeling Chau (California Institute of Technology), Andrei Barbu (Massachusetts Institute of Technology)
CodeAnomaly DetectionRepresentation LearningTransformerContrastive LearningTime SeriesBiomedical Data
🎯 What it does: This paper proposes a self-supervised Population Transformer (PopT) that aggregates multi-channel neural data based on pre-trained time series embeddings, capable of handling arbitrary channel combinations.
Port-Hamiltonian Architectural Bias for Long-Range Propagation in Deep Graph Networks
Simon Heilig (Ruhr University Bochum), Davide Bacciu (University of Pisa)
CodeGraph Neural NetworkGraph
🎯 What it does: A deep graph network based on port-Hamiltonian dynamics (PH-DGN) is proposed, which achieves long-distance information propagation by introducing energy conservation and non-conservative dynamics into graph neural networks, enhancing performance without the need for global encoding or rearrangement.
CodeGenerationData SynthesisSafty and PrivacyDiffusion modelImage
🎯 What it does: A positive-negative unlabeled diffusion model is proposed, which trains the diffusion model using a small number of sensitive samples and unlabeled data to avoid generating sensitive content.
PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling
Junchao Gong (Shanghai Jiao Tong University), LEI BAI
CodeRestorationGenerationDiffusion modelTime Series
🎯 What it does: This paper proposes a post-processing method named PostCast, aimed at eliminating the ambiguity in precipitation nowcasting and improving the prediction accuracy of extreme precipitation events.
🎯 What it does: A novel non-reverse, non-training image editing method called PostEdit is proposed, which incorporates measurement terms and Langevin dynamics into the diffusion process using posterior sampling theory to achieve efficient zero-shot image editing while maintaining background consistency.
🎯 What it does: This paper proposes the Patch Order-Aware Pretext Task (PPT), which enhances temporal classification performance by learning patch order information through weak/strong permutation of time series patches along the channel dimension and utilizing consistency and contrastive loss.
PQMass: Probabilistic Assessment of the Quality of Generative Models using Probability Mass Estimation
Pablo Lemos (Sandbox Quantum), Yashar Hezaveh (Trottier Space Institute)
CodeGenerationAnomaly DetectionImageMultimodalityTabularTime Series
🎯 What it does: A likelihood-free, model-free training method for probability mass estimation, PQMass, is proposed to determine whether two sets of samples come from the same distribution.
Precedence-Constrained Winter Value for Effective Graph Data Valuation
Hongliang Chi (Rensselaer Polytechnic Institute), Yao Ma (IBM)
CodeGraph Neural NetworkGraph
🎯 What it does: A value assessment method for graph data called PC-Winter is proposed, which quantifies the contribution of nodes (including unlabeled nodes) and edges in the graph to the performance of graph neural networks.
Precise Localization of Memories: A Fine-grained Neuron-level Knowledge Editing Technique for LLMs
Haowen Pan (University of Science and Technology of China), Meng Wang (Hefei University of Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: A knowledge editing method based on neuron-level fine-grained localization, FiNE, is proposed, which directly fine-tunes specific neurons in the FFN to achieve precise updates of knowledge in large language models.
🎯 What it does: This paper proposes an end-to-end Predictive Inverse Dynamics Model (PIDM) called Seer, which combines conditional visual foresight with inverse dynamics prediction to learn actions in a closed-loop manner for robotic manipulation tasks. By pre-training on a large-scale robotic dataset and fine-tuning on downstream tasks, it achieves efficient and scalable control strategies.
Predictive Uncertainty Quantification for Bird's Eye View Segmentation: A Benchmark and Novel Loss Function
Linlin Yu (University of Texas at Dallas), Feng Chen (University of Texas at Dallas)
CodeSegmentationAutonomous DrivingImageBenchmark
🎯 What it does: A benchmark for uncertainty quantification was established in the bird's-eye view semantic segmentation task, and the UFCE loss and its related regularization framework were proposed to enhance model calibration and uncertainty prediction.
🎯 What it does: Proposes the PreferDiff optimization objective to improve the training of diffusion models in sequential recommendation, balancing generation and personalized ranking.
🎯 What it does: A one-shot post-training pruning framework called SNOWS is proposed, aiming to significantly improve the performance of pruned models by adjusting the weights of existing sparse masks without retraining, thereby preserving deep network representations.
Preserving Diversity in Supervised Fine-Tuning of Large Language Models
Ziniu Li (Chinese University of Hong Kong), Ruoyu Sun (Chinese University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The GEM algorithm is proposed, which is a distribution matching framework based on game theory. It achieves sparse updates and adaptive termination in the supervised fine-tuning of LLMs through auxiliary variables to maintain output diversity.
🎯 What it does: This study focuses on the problem of negative transfer in continual reinforcement learning and proposes a simple baseline method called R&D based on reset and distillation.
Ozan Ozdenizci (Montanuniversität Leoben), Robert Legenstein (Graz University of Technology)
CodeSafty and PrivacyKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A privacy-aware lifelong learning (PALL) framework is proposed, achieving non-catastrophic forgetting, selective task forgetting (complete task forgetting) in task incremental learning while maintaining knowledge transfer and keeping low model memory usage in fixed capacity networks.
Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance
Yaxi Lu (Tsinghua University), Maosong Sun (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: A ProactiveBench dataset was constructed and an LLM agent was trained to proactively predict and provide tasks that users may need without explicit instructions.
Probabilistic Conformal Prediction with Approximate Conditional Validity
Vincent Plassier (Lagrange Mathematics and Computing Research Center), Eric Moulines (École Polytechnique)
CodeTabularTime Series
🎯 What it does: The CP2 framework is proposed, which combines split conformal prediction with conditional distribution estimation to generate an approximately conditionally valid prediction set.
Sanghyuk Chun (NAVER AI Lab), Sangdoo Yun (NAVER AI Lab)
CodeClassificationRecognitionTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A full-probability visual-language pre-training model ProLIP is proposed, which maps images and text to Gaussian distributions and effectively estimates variance through uncertainty tokens to address the ambiguity of many-to-many matching.
Probabilistic Learning to Defer: Handling Missing Expert Annotations and Controlling Workload Distribution
Cuong C. Nguyen (University of Surrey), Gustavo Carneiro (University of Surrey)
CodeClassificationOptimizationMixture of ExpertsBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a probabilistic Learning to Delay (L2D) model that can use only a portion of expert annotations during training and infer missing annotations through the EM algorithm, while incorporating constraints in the E-step to control the workload of human experts and AI classifiers.
🎯 What it does: This paper proposes a probability neural pruning framework based on the Sparse Evolving Fokker-Planck-Kolmogorov equation (SFPK) and implements the corresponding particle simulation pruning algorithm SFPK-pruner.
Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models
Biao Yi (Nankai University), Yiming Li (Nanyang Technological University)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a black-box input-level defense method named BEAT, which can detect and suppress security failures caused by backdoor unalignment during LLM inference.
Probe Pruning: Accelerating LLMs through Dynamic Pruning via Model-Probing
Qi Le (University of Minnesota), Ali Anwar (University of Minnesota)
CodeGenerationOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A novel online dynamic structured pruning framework called Probe Pruning is proposed, which guides weight channel pruning through small-scale probing for each batch while maintaining efficiency during the full inference phase.
🎯 What it does: A dual-layer program synthesis framework is designed to generate synthesizable molecules and their analogs, optimizing molecular structures under given property targets.
Wendi Li (Huazhong University of Science and Technology), Yixuan Li (University of Wisconsin Madison)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the Process Q-value Model (PQM), which transforms the modeling of process rewards into a Q-value ranking problem and designs a margin-based comparison loss to train the model.
Progressive Compression with Universally Quantized Diffusion Models
Yibo Yang (University of California), Stephan Mandt (University of California)
CodeCompressionDiffusion modelImage
🎯 What it does: This paper proposes an unconditional progressive compression framework based on diffusion models called Universally Quantized Diffusion Model (UQDM), which enables stepwise decoding from low bit rates to lossless compression.
Progressive Parameter Efficient Transfer Learning for Semantic Segmentation
Nan Zhou (Beihang University), Di Huang (Beihang University)
CodeSegmentationSupervised Fine-TuningImage
🎯 What it does: This paper proposes a staged parameter-efficient transfer learning framework called ProPETL, designed to quickly adapt large-scale pre-trained models to semantic segmentation tasks.
🎯 What it does: PRO-SCALE is proposed, an efficient implementation of hierarchical token length extension for the Transformer encoder within the Mask2Former universal segmentation framework, and introduces a lightweight pixel embedding (LPE) module.
🎯 What it does: This paper conducts an in-depth theoretical analysis of the projection head in contrastive learning from an information theory perspective. Based on the conclusion that the projection head should act as an information bottleneck, it further proposes two types of methods to improve the projection head: training regularization (mutual information constraint) and structural regularization (discretized projection head, sparse autoencoder).
Prompt as Knowledge Bank: Boost Vision-language model via Structural Representation for zero-shot medical detection
Yuguang Yang (Beihang University), Baochang Zhang (Beihang University)
CodeObject DetectionTransformerPrompt EngineeringVision Language ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes the StructuralGLIP framework, which utilizes a structured prompt knowledge base to achieve zero-shot object detection in medical images, supporting category-level prompts and instance-level dynamic selection;
Promptriever: Instruction-Trained Retrievers Can Be Prompted Like Language Models
Orion Weller (Johns Hopkins University), Jack Hessel (Samaya AI)
CodeRetrievalTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study proposes a model called Promptriever that can perform retrieval based on natural language prompts, and constructs nearly 500k training data instances with instruction-level guidance.
🎯 What it does: ProtComposer has been developed, a protein structure generation method based on 3D ellipsoid layout, which can control the spatial layout and secondary structure of proteins through manually specified or statistically modeled ellipsoids, and implements conditional generation on the MultiFlow generation framework; it also introduces techniques such as Invariant Cross Attention, classifier-independent guidance, and self-conditioning to enhance generation quality, controllability, and diversity.
🎯 What it does: This paper presents ProteinBench, a unified multi-dimensional evaluation framework for systematically assessing the performance of protein foundation models on eight sub-tasks related to design and conformation prediction.
🎯 What it does: This paper reduces Large-Neighborhood Local Search (LNLS) to Block Langevin Diffusion (BLD) and provides a proof of non-asymptotic convergence for random and cyclic block selection strategies under non-ideal devices. Based on this, it derives the Wasserstein error bounds for devices with finite variations.
Provable Robust Overfitting Mitigation in Wasserstein Distributionally Robust Optimization
Shuang Liu (Academy of Mathematics and Systems Science Chinese Academy of Sciences), Xiao-Shan Gao (Academy of Mathematics and Systems Science Chinese Academy of Sciences)
🎯 What it does: A Statistically Robust Wasserstein Distributionally Robust Optimization (SR-WDRO) framework is proposed, which incorporates KL divergence on top of WDRO to simultaneously control statistical errors and adversarial noise, thereby alleviating robust overfitting.
Provably Robust Explainable Graph Neural Networks against Graph Perturbation Attacks
Jiate Li (Illinois Institute of Technology), Binghui Wang (Illinois Institute of Technology)
CodeExplainability and InterpretabilityAdversarial AttackDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A provably robust and interpretable graph neural network (XGNN) defense framework called XGNNCert is proposed, which ensures the consistency and accuracy of explanation results when facing graph structure perturbation attacks.
🎯 What it does: Transforming the trained classifier into a self-rejecting classifier that can refuse to make predictions when encountering OOD (out-of-distribution) or adversarial samples.
CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: A context pruner called Provence is proposed, which is based on sentence-level sequence labeling and can automatically detect and remove irrelevant sentences in retrieval-augmented generation (RAG), compatible with multi-domain and varying lengths of retrieval results.
Proving Olympiad Inequalities by Synergizing LLMs and Symbolic Reasoning
Zenan Li (Nanjing University), Xiaoxing Ma (Nanjing University)
CodeLarge Language ModelPrompt EngineeringChain-of-Thought
🎯 What it does: A neural symbolic framework called LIPS has been designed and implemented for the automatic derivation of inequality proofs at the level of mathematical Olympiads, ultimately generating formal proofs that are both readable and verifiable in Lean 4.
Proximal Mapping Loss: Understanding Loss Functions in Crowd Counting & Localization
Wei Lin (Harbin Institute of Technology), Antoni B. Chan (City University of Hong Kong)
CodeObject DetectionOptimizationImage
🎯 What it does: A new loss function called Proximal Mapping Loss (PML) is proposed for training crowd counting models without the assumption of intersection hypothesis;
PseDet: Revisiting the Power of Pseudo Label in Incremental Object Detection
Qiuchen Wang (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
CodeObject DetectionImage
🎯 What it does: This study explores the use of pseudo-labels in incremental object detection and proposes the PseDet framework to enhance the quality of pseudo-labels from old models and mitigate forgetting.
🎯 What it does: This paper proposes utilizing the property of Neural Collapse to achieve dimensional separation of ID and OOD features through a feature separation loss during the model fine-tuning phase, thereby enhancing OOD detection performance.
Q-Adapter: Customizing Pre-trained LLMs to New Preferences with Forgetting Mitigation
Yi-Chen Li (Nanjing University), Bo An (Nanyang Technological University)
CodeRecommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes Q-Adapter, which utilizes residual Q learning to directly learn adaptive modules from new human preference data, customizing pre-trained LLMs.
Qinco2: Vector Compression and Search with Improved Implicit Neural Codebooks
Théophane Vallaeys (Meta), Matthijs Douze (Meta)
CodeRetrievalCompressionTabular
🎯 What it does: Proposes QINCO2, an improved implicit neural residual quantization method that balances vector compression and trillion-scale approximate nearest neighbor search;
QMP: Q-switch Mixture of Policies for Multi-Task Behavior Sharing
Grace Zhang (University of Southern California), Joseph J Lim
CodeRobotic IntelligenceReinforcement LearningMixture of Experts
🎯 What it does: Proposes the Q-switch Mixture of Policies (QMP) framework, which utilizes task-specific Q-functions to select optimal behaviors, thereby achieving unbiased behavior sharing in multi-task reinforcement learning and improving sample efficiency.
QPM: Discrete Optimization for Globally Interpretable Image Classification
Thomas Norrenbrock (Institute for Information Processing), Bodo Rosenhahn (Institute for Information Processing)
CodeClassificationOptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: A discrete optimization method based on quadratic programming, QPM, is proposed, which achieves a globally interpretable image classification model using optimal binary feature allocation with only 5 common features for each category.
🎯 What it does: A continuous-time dynamic graph generation model evaluation metric based on the Johnson-Lindenstrauss transform is proposed, which directly performs random projection on event sequences without the need for discrete snapshots.
Quality over Quantity in Attention Layers: When Adding More Heads Hurts
Noah Amsel (New York University), Joan Bruna (Flatiron Institute)
CodeTransformer
🎯 What it does: This study investigates the impact of the rank of attention layers and the number of heads on representational capacity, proving that low-rank attention cannot effectively approximate tasks such as nearest neighbor search, and experimentally validating the theoretical conclusions.
Quamba: A Post-Training Quantization Recipe for Selective State Space Models
Hung-Yueh Chiang (University of Texas at Austin), Diana Marculescu (University of Texas at Austin)
CodeLarge Language ModelText
🎯 What it does: Designed and implemented a post-training 8-bit quantization scheme for selective state space models (SSM) - Quamba, supporting the deployment of large models like Mamba and Jamba on low-resource devices.
Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model
Chaochen Gao (Institute of Information Engineering, Chinese Academy of Sciences), Songlin Hu (Xiaohongshu Inc)
CodeGenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: A query-based long text synthesis method called Quest has been developed to generate diverse and semantically relevant long-context training data.
R-Sparse: Rank-Aware Activation Sparsity for Efficient LLM Inference
Zhenyu Zhang (University of Texas at Austin), Steven Li (Meta AI)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a training-free, input activation sparse-based Rank-Aware Activation Sparsity (R-Sparse) method for efficient inference of large language models (LLMs).
🎯 What it does: This paper proposes a Relaxed Rotation-Equivariant Group (R_n) and its corresponding R2GConv operation, and constructs a lightweight backbone network R2Net and an improved 2D object detector R2Det, aiming to address the issue of rotation symmetry disruption in practical scenarios.
RA-TTA: Retrieval-Augmented Test-Time Adaptation for Vision-Language Models
Youngjun Lee (KAIST), Jae-Gil Lee (KAIST)
CodeRetrievalDomain AdaptationTransformerLarge Language ModelVision Language ModelImageTextRetrieval-Augmented Generation
🎯 What it does: A retrieval-enhanced test-time adaptive method RA-TTA is designed to improve zero-shot inference of VLM under distribution shift by utilizing an external image database.
RAG-DDR: Optimizing Retrieval-Augmented Generation Using Differentiable Data Rewards
Xinze Li (Northeastern University), Chenyan Xiong (Carnegie Mellon University)
CodeRetrievalOptimizationTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Differentiable Data Rewards (DDR) method, which optimizes retrieval-augmented generation (RAG) systems end-to-end by collecting system rewards and aligning the data preferences of various modules, enabling the retrieval and generation modules to work together to produce more accurate answers.
CodeRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A unified preference optimization framework RAINBOWPO is proposed, which disassembles and combines seven key components of the DPO series algorithms to enhance the alignment effect of LLM.
RandLoRA: Full rank parameter-efficient fine-tuning of large models
Paul Albert (Australian Institute for Machine Learning), Ehsan Abbasnejad (Australian Institute for Machine Learning)
CodeSupervised Fine-TuningImageText
🎯 What it does: This paper proposes RandLoRA, a parameter-efficient and memory-friendly full-rank fine-tuning method that achieves this by learning a linear combination of random low-rank bases and only training diagonal scaling matrices.
Range, not Independence, Drives Modularity in Biologically Inspired Representations
Will Dorrell (University College London), James C. R. Whittington (Stanford University)
CodeOptimizationRepresentation LearningReinforcement Learning from Human FeedbackRecurrent Neural NetworkAuto EncoderImageSequential
🎯 What it does: A modular theory for linear autoencoders with biological constraints (non-negativity, energy efficiency) is proposed, providing necessary and sufficient conditions. This theory is extended to nonlinear feedforward networks, recurrent networks, and neuroscience experiments, explaining how neurons exhibit modular or mixed selection under different conditions.
Ranking-aware adapter for text-driven image ordering with CLIP
Wei-Hsiang Yu (National Yang Ming Chiao Tung University), Yi-Hsuan Tsai (Atmanity Inc.)
CodeRetrievalRecommendation SystemTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: A lightweight Ranking-Aware Adapter is proposed, reconstructing the pre-trained CLIP for learning to rank (LTR) tasks, utilizing text-driven visual differences to achieve multi-image ranking.
RAPID: Retrieval Augmented Training of Differentially Private Diffusion Models
Tanqiu Jiang (Stony Brook University), Ting Wang (Stony Brook University)
CodeGenerationData SynthesisSafty and PrivacyComputational EfficiencyDiffusion modelContrastive LearningImageRetrieval-Augmented Generation
🎯 What it does: By integrating Retrieval-Augmented Generation (RAG) technology into the training of differential privacy diffusion models, a trajectory knowledge base is constructed using public data. During training on private data, only the later sampling steps are subjected to differential privacy processing, significantly improving generation quality, reducing memory usage, and lowering inference costs.
Zhiwei He (Shanghai Jiao Tong University), Rui Wang (Tencent AI Lab)
CodeOptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the RaSA (Rank‑Sharing Low‑Rank Adaptation) method, which enhances the model's expressive capability by sharing a portion of the rank across layers based on LoRA's low-rank parameter updates, while maintaining the same number of trainable parameters.
Rational Decision-Making Agent with Learning Internal Utility Judgment
Yining Ye (Tsinghua University), Maosong Sun (Tsinghua University)
CodeLarge Language ModelReinforcement LearningAgentic AITabular
🎯 What it does: Proposes RaDAgent, which constructs an internal utility judgment mechanism based on LLM, enabling the agent to make autonomous decisions without external evaluation.
Rationalizing and Augmenting Dynamic Graph Neural Networks
Guibin Zhang (Tongji University), Jian Guo (International Digital Economy Academy)
CodeDomain AdaptationAnomaly DetectionGraph Neural NetworkContrastive LearningGraphTime Series
🎯 What it does: A graph data augmentation framework for dynamic graphs, DyAug, is proposed, which maintains temporal consistency through temporal conditional graph rationalization to enhance the performance, robustness, and generalization ability of dynamic GNNs.
Re-Aligning Language to Visual Objects with an Agentic Workflow
Yuming Chen (Nankai University), Yibing Song
CodeObject DetectionTransformerLarge Language ModelAgentic AIVision Language ModelImageTextMultimodality
🎯 What it does: A workflow based on LLM agents (Real-LOD) has been designed and implemented to automatically correct visual object descriptions generated by VLM, aligning language expressions with target objects and generating high-quality language-visual paired data for training language-based object detection models.
🎯 What it does: Reassess the impact of unseen classes in semi-supervised learning on unlabeled data, and propose the RE-SSL evaluation framework along with a set of global and local robustness metrics.
Re-Imagining Multimodal Instruction Tuning: A Representation View
Yiyang Liu (University of Missouri Kansas City), Cheng Han (University of Missouri Kansas City)
CodeRepresentation LearningTransformerVision Language ModelMultimodality
🎯 What it does: For instruction fine-tuning of large-scale multimodal models, a parameter-efficient representation tuning method (MRT) is proposed, which only edits the representations in the visual encoder, cross-modal projection layer, and language model while keeping the model parameters frozen.
🎯 What it does: The HeartLang framework is proposed, utilizing QRS-Tokenizer to split electrocardiograms into heartbeat words and rhythm sentences, constructing an ECG vocabulary and learning multi-level representations of unlabeled ECG through vector quantization of heartbeat reconstruction and masked sentence pre-training.