Conference on Neural Information Processing Systems · 2283 papers
Online Portfolio Selection with ML Predictions
Ziliang Zhang (University of Sydney), Albert Zomaya
CodeRecommendation SystemOptimizationTime SeriesFinance Related
🎯 What it does: This paper proposes the Rebalanced Arithmetic Mean (RAM) online portfolio algorithm, which dynamically reallocates weights based on asset return rankings provided by machine learning, ensuring safe returns even in the worst-case scenario.
Hanshi Wang (Chinese Academy of Sciences), Zhipeng Zhang (Shanghai Jiao Tong University)
CodeObject TrackingSegmentationTransformerVision Language ModelPoint Cloud
🎯 What it does: This paper proposes an online real-time 3D instance segmentation framework called AutoSeg3D, which directly uses 2D masks obtained from Vision Foundation Models (such as SAM) as queries in each frame, and maintains long-term and short-term memory in the temporal dimension to achieve instance identity continuity and immediate context updates.
Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model
Jingcheng Hu (Tsinghua University), Heung-Yeung Shum
CodeReinforcement LearningText
🎯 What it does: Developed and publicly implemented Open-Reasoner-Zero, which directly applies large-scale reinforcement learning to enhance reasoning capabilities.
Open-Vocabulary Part Segmentation via Progressive and Boundary-Aware Strategy
Xinlong Li (Tianjin University), Qing Guo (Nankai University)
CodeSegmentationDiffusion modelImageBenchmark
🎯 What it does: This paper proposes an untrained open vocabulary part segmentation framework PBAPS, which achieves fine segmentation of the boundaries of structurally connected parts using a hierarchical structure graph and a boundary-aware refinement module.
🎯 What it does: This paper proposes an open-world drone active tracking benchmark DAT and a reinforcement learning method GC-VAT based on goal-centered rewards and curriculum learning.
OpenMMEgo: Enhancing Egocentric Understanding for LMMs with Open Weights and Data
Hao Luo (Peking University), Zongqing Lu (Peking University)
CodeRecognitionData SynthesisCompressionTransformerLarge Language ModelVision Language ModelVideoTextBenchmark
🎯 What it does: A large-scale self-supervised egocentric video QA dataset OME10M is constructed, and the OpenMMEgo framework is proposed to enhance LMM performance in first-person video understanding.
OpenOmni: Advancing Open-Source Omnimodal Large Language Models with Progressive Multimodal Alignment and Real-time Emotional Speech Synthesis
Run Luo (Shenzhen Key Laboratory for High Performance Data Mining, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Min Yang (Shenzhen Key Laboratory for High Performance Data Mining, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences)
CodeGenerationData SynthesisTransformerLarge Language ModelMixture of ExpertsTextMultimodalityAudio
🎯 What it does: OpenOmni has been constructed and trained, which is a two-stage open-source multimodal large language model. It first uses text as a hub to achieve speech-text alignment and image-text alignment, enabling zero-shot multimodal alignment. Subsequently, a lightweight parallel speech decoder is trained, combined with DPO for emotional speech generation, achieving real-time emotional expression.
OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles
Yihe Deng (University of California), Kai-Wei Chang (University of California)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a training framework based on an iterative SFT-RL cycle, creating an open-source LVLM—OpenVLThinker-7B, capable of performing complex Chain-of-Thought reasoning in visual tasks.
OpenWorldSAM: Extending SAM2 for Universal Image Segmentation with Language Prompts
Shiting Xiao (Yale University), Priyadarshini Panda (Yale University)
CodeSegmentationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes OpenWorldSAM, an open-world text prompt semantic segmentation framework built on SAM2, which can accurately generate multi-instance, semantic, and panoptic segmentation masks based on open vocabulary text descriptions (from words to complete sentences).
OPHR: Mastering Volatility Trading with Multi-Agent Deep Reinforcement Learning
Zeting Chen (Nanyang Technological University), Bo An (Nanyang Technological University)
CodeOptimizationReinforcement LearningAgentic AITime SeriesFinance Related
🎯 What it does: A multi-agent reinforcement learning framework for cryptocurrency options volatility trading, OPHR, is proposed, which includes options position decision-making (OP-Agent) and hedging strategy selection (HR-Agent), achieving a synergistic optimization of volatility timing and risk management.
Optical Coherence Tomography Harmonization with Anatomy-Guided Latent Metric Schrödinger Bridges
Shuwen Wei (Johns Hopkins University), Jerry L Prince
CodeImage HarmonizationSegmentationDiffusion modelImageBiomedical Data
🎯 What it does: This paper proposes a method for unpaired optical coherence tomography (OCT) image harmonization based on reversible networks and latent Euclidean space;
Optimal Adjustment Sets for Nonparametric Estimation of Weighted Controlled Direct Effect
Ruiyang Lin (University of Science and Technology of China), Kyra Gan (Cornell University)
CodeTabularBiomedical DataAgriculture Related
🎯 What it does: This paper proposes the identifiability conditions, influence function, and optimal adjustment set for weighted control direct effects (WCDE) in observational studies, and validates the AIPW estimator.
Optimal Graph Clustering without Edge Density Signals
Maximilien Dreveton (École Polytechnique Fédérale de Lausanne), Patrick Thiran (École Polytechnique Fédérale de Lausanne)
CodeOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies the optimal error rate of graph clustering under the Popularity Adjusted Block Model (PABM), revealing the possibility of achieving clustering even when traditional edge density signals disappear.
Optimal Neural Compressors for the Rate-Distortion-Perception Tradeoff
Eric Lei (JPMorganChase Global Technology Applied Research), Shirin Saeedi Bidokhti (University of Pennsylvania)
CodeCompressionImagePhysics RelatedAudio
🎯 What it does: This paper designs and implements a low-complexity neural compressor that achieves optimal or near-optimal compression for the R-D-P trade-off using lattice coding and various shared randomness designs.
Optimistic Query Routing in Clustering-based Approximate Maximum Inner Product Search
Sebastian Bruch (Northeastern University), Franco Maria Nardini (ISTI-CNR)
CodeRetrievalOptimizationTextBenchmark
🎯 What it does: In clustering-based approximate maximum inner product search, a router called OPTIMIST is proposed based on the principle of 'optimism towards uncertainty' to more accurately select the shards required for queries.
🎯 What it does: A topology optimization framework OAT based on a foundational model is proposed, which can directly generate approximately optimal material distributions under arbitrary shapes, aspect ratios, and resolutions.
Optimizing Anytime Reasoning via Budget Relative Policy Optimization
Penghui Qi (Sea AI Lab), Min Lin (National University of Singapore)
CodeOptimizationLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: A framework named AnytimeReasoner is proposed, which utilizes budget sampling, dense verifiable rewards, and Budget Relative Policy Optimization (BRPO) to achieve anytime reasoning for LLMs and enhance their performance.
Option-aware Temporally Abstracted Value for Offline Goal-Conditioned Reinforcement Learning
Hongjoon Ahn (Seoul National University), Taesup Moon (Seoul National University)
CodeReinforcement LearningTabular
🎯 What it does: This paper proposes an option-based temporal abstraction value learning method called OTA, aimed at enhancing the performance of high-level policies in offline goal-oriented reinforcement learning.
OptiTree: Hierarchical Thoughts Generation with Tree Search for LLM Optimization Modeling
Haoyang Liu (University of Science and Technology of China), Jianye HAO
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By constructing a hierarchical 'model tree' and utilizing tree search, the complex operations research modeling task is decomposed into a series of subproblems, thereby guiding large language models to generate more accurate optimization models.
Jung-hun Kim (ENSAE Paris), Min-hwan Oh (Seoul National University)
CodeOptimizationReinforcement LearningTabular
🎯 What it does: Two sparse oracle query frameworks (adaptive and planned) are proposed, reducing oracle calls from linear to double logarithmic in the combinatorial semi-bandit problem, while maintaining an almost optimal no-gap O(√T) scheduling reward.
Order-Level Attention Similarity Across Language Models: A Latent Commonality
Jinglin Liang (South China University of Technology), Hanlin Gu (Hong Kong University of Science and Technology)
CodeTransformerLarge Language ModelText
🎯 What it does: Exploring the common patterns of context aggregation in different language models, we propose Order-Level Attention (OLA) and implement a training-free cross-model adapter transfer (TOA) based on it.
🎯 What it does: A framework named Orientation-anchored Gaussian Splatting (OriGS) is proposed for high-quality reconstruction of four-dimensional (time + space) scenes from casually captured monocular videos.
ORIGAMISPACE: Benchmarking Multimodal LLMs in Multi-Step Spatial Reasoning with Mathematical Constraints
Rui Xu, Xu Yinghui
CodeAI Code AssistantTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: This paper designs and releases the ORIGAMISPACE dataset along with four multimodal LLM evaluation tasks (pattern prediction, multi-step spatial reasoning, spatial relationship prediction, and CP code generation), and conducts experiments and evaluations based on an improved origami compiler.
🎯 What it does: This paper proposes an Orthogonal Contrastive Learning (OCL) framework for aligning and feature learning of multi-subject and cross-site task fMRI data without the need for temporal alignment and uniform sequence lengths.
Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data
Dennis Frauen (Ludwig Maximilian University of Munich), Stefan Feuerriegel (Ludwig Maximilian University of Munich)
CodeOptimizationDrug DiscoveryTime SeriesBiomedical Data
🎯 What it does: A set of Neyman-orthogonal survival learners is proposed to estimate heterogeneous treatment effects (HTE) in the presence of censored time-to-event data.
Oryx: a Scalable Sequence Model for Many-Agent Coordination in Offline MARL
Juan Claude Formanek, Arnu Pretorius (Stellenbosch University)
CodeReinforcement LearningSequential
🎯 What it does: This paper presents Oryx, an autoregressive sequence model for offline multi-agent reinforcement learning, focusing on large-scale multi-agent collaboration and long-term temporal dependencies.
🎯 What it does: A first-order diffusion code (OSCAR) is proposed, which achieves multi-bitrate image compression in a single-step inference and requires training only one unified model.
OSTAR: Optimized Statistical Text-classifier with Adversarial Resistance
Yuhan Yao (Beijing University of Posts and Telecommunications), LI Haisheng
CodeClassificationAdversarial AttackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: The OSTAR framework is proposed, which achieves robust detection of machine-generated text through statistical feature profiling and multi-faceted contrastive learning.
🎯 What it does: Proposes the OSVI-WM framework, which utilizes single-segment expert videos and agent initial observations to generate potential state trajectories through a world model and decode them into physical waypoints, achieving one-shot visual imitation for unseen tasks.
Overcoming Long Context Limitations of State Space Models via Context Dependent Sparse Attention
Zhihao Zhan (Mila - Québec AI Institute), Jian Tang (Mila - Québec AI Institute)
CodeTransformerLarge Language ModelTextBenchmark
🎯 What it does: Designed and validated a Joint Recall task to assess long context modeling capabilities, and proposed the HAX architecture that combines SSM with context-aware sparse attention.
CodeObject DetectionSegmentationSupervised Fine-TuningMixture of ExpertsImage
🎯 What it does: We propose ConOVS, an incremental learning framework for Open Vocabulary Segmentation (OVS) that can gradually expand the model's recognition capabilities in an environment where new data is continuously added.
OWL: Optimized Workforce Learning for General Multi-Agent Assistance in Real-World Task Automation
Mengkang Hu (University of Hong Kong), Guohao Li (Eigent.AI)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodality
🎯 What it does: This paper presents WORKFORCE, a hierarchical multi-agent framework that separates task planning, coordination, and domain-specific execution; it also introduces the OPTIMIZED WORKFORCE LEARNING (OWL) training paradigm, specifically designed to enhance the cross-domain generalization ability of general planners.
P-Law: Predicting Quantitative Scaling Law with Entropy Guidance in Large Recommendation Models
Tingjia Shen (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeRecommendation SystemTransformerSequential
🎯 What it does: This paper proposes the Performance Law, which evaluates data quality using minimum encoding length and true entropy, and introduces a decay term to quantitatively predict the performance of large sequence recommendation models.
🎯 What it does: This paper proposes PairEdit, a continuous image editing method that learns complex semantic transformations based on a small number of image pairs without relying on text prompts.
🎯 What it does: This paper proposes an All-to-All Flow-based Transfer Model (A2A-FM) based on flow matching, which can learn approximate optimal transport mappings between all pairs of conditions under any conditional distribution, achieving conditional transfer.
PANDA: Towards Generalist Video Anomaly Detection via Agentic AI Engineer
Zhiwei Yang (Xidian University), Mike Zheng Shou (National University of Singapore)
CodeAnomaly DetectionLarge Language ModelAgentic AIVision Language ModelVideoMultimodalityRetrieval-Augmented Generation
🎯 What it does: A universal video anomaly detection framework called PANDA is proposed, which is training-free and requires no human intervention, and can adapt to different scenes and types of anomalies.
PANTHER: Generative Pretraining Beyond Language for Sequential User Behavior Modeling
Guilin Li (Tencent), Matthias Hwai Yong Tan (City University of Hong Kong)
CodeRecommendation SystemAnomaly DetectionTransformerContrastive LearningTime SeriesSequentialFinance Related
🎯 What it does: Proposes the PANTHER framework, which combines generative pre-training and real-time discrimination to achieve multi-task applications such as fraud detection and next transaction prediction in payment scenarios.
Mouxiang Chen (Zhejiang University), Zhongxin Liu (Zhejiang University)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A scale expansion method named PAR-SCALE is proposed, which enhances model performance and improves inference efficiency by executing multiple streams in parallel on the same set of model parameters (each stream uses learnable prefix transformations and dynamically weighted aggregation) without increasing the number of parameters.
🎯 What it does: This paper proposes a parallel MCMC method based on the parallel Newton iteration (DEER) framework, which enables parallelization along the sequence length dimension. The framework is applied to three commonly used samplers: Gibbs, MALA, and HMC, while also providing memory-friendly techniques such as random approximate Jacobian, sliding window, and block quantum Newton.
CodeGenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes the ParamMute framework to suppress the activation of specific FFNs to reduce the LLM's reliance on internal memory and enhance the credibility of retrieval-augmented generation.
Partial Information Decomposition via Normalizing Flows in Latent Gaussian Distributions
Wenyuan Zhao (Texas A&M University), Paul Pu Liang (Massachusetts Institute of Technology)
CodeOptimizationFlow-based ModelMultimodalityBiomedical Data
🎯 What it does: A framework for partial information decomposition (PID) based on normalized flows under a latent Gaussian distribution is proposed, which includes an efficient Thin-PID algorithm and a scalable Flow-PID encoder.
Partition to Evolve: Niching-enhanced Evolution with LLMs for Automated Algorithm Discovery
Qinglong Hu (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelPrompt Engineering
🎯 What it does: A framework for evolutionary search (LES) that combines large language models is proposed, utilizing feature-assisted partitioning to construct niches and achieving algorithmic auto-discovery (PartEvo) based on this.
🎯 What it does: To address the bias prediction problem of multimodal data under temporal shifts during testing, we propose the Partition-Then-Adapt (PTA) method, which first divides samples into reliable and unreliable subsets based on predicted bias, and then adapts through quantized reweighting and attention alignment.
PARTONOMY: Large Multimodal Models with Part-Level Visual Understanding
Ansel Blume (University of Illinois Urbana-Champaign), Heng Ji (University of Illinois Urbana-Champaign)
CodeRecognitionObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This study investigates the capabilities of multimodal models in pixel-level part recognition and reasoning, proposing the Explanatory Part Segmentation task and the PARTONOMY benchmark.
🎯 What it does: The PASS framework is proposed for object/action recognition in event cameras, supporting event lengths from 10⁶ to 10⁹ while maintaining good generalization at different inference frequencies.
PaTH Attention: Position Encoding via Accumulating Householder Transformations
Songlin Yang (Massachusetts Institute of Technology), Yoon Kim (Massachusetts Institute of Technology)
CodeTransformerTextSequential
🎯 What it does: A data-driven cumulative Householder transformation-based polynomial position encoding (PaTH) is proposed for the attention mechanism of Transformers;
🎯 What it does: This paper proposes a zero-shot PBR texture super-resolution method called PBR-SR, which can enhance low-resolution PBR textures to high resolution without requiring any PBR-specific training data, while maintaining material consistency and real-time relighting capabilities.
🎯 What it does: This paper proposes a weakly supervised compositional moment retrieval task and designs PC-Net to achieve retrieval using only video-text pairs without timestamp annotations.
PCA++: How Uniformity Induces Robustness to Background Noise in Contrastive Learning
Mingqi Wu (McGill University), Archer Y. Yang (McGill University)
CodeOptimizationRepresentation LearningContrastive LearningImageBiomedical Data
🎯 What it does: A new contrastive learning method PCA++ is proposed, which achieves robust signal subspace recovery through hard unification constraints.
Sebastian Gutierrez Hernandez (Georgia Institute of Technology), Hao-Min Zhou
CodeOptimizationOrdinary Differential Equation
🎯 What it does: This paper proposes the Parametric Density Path Optimization (PDPO) method, which uses a parametric pushforward mapping (Neural ODE) and cubic Hermite spline approximation to solve for the path that minimizes action in probability density space, applicable to various constraints such as obstacles, interactions, and stochastic control.
Per-Architecture Training-Free Metric Optimization for Neural Architecture Search
Mingzhuo Lin (Shenzhen University), Jianping Luo (Shenzhen University)
CodeOptimizationNeural Architecture SearchImage
🎯 What it does: This paper proposes a Per-Architecture Training-Free Metric Optimization NAS (PO-NAS), which achieves efficient and competitive performance in neural architecture search by dynamically assigning training-independent metric weights for each candidate architecture and combining proxy models with evolutionary search.
Perception Encoder: The best visual embeddings are not at the output of the network
Daniel Bolya (Meta), Christoph Feichtenhofer (Meta)
CodeClassificationObject DetectionSegmentationDepth EstimationRetrievalTransformerLarge Language ModelContrastive LearningImageVideoTextMultimodality
🎯 What it does: This paper proposes the Perception Encoder (PE) family, which combines contrastive vision-language (CLIP) pre-training with video data engines and language and spatial alignment techniques to generate a unified image and video encoder, achieving multi-task performance through short-term fine-tuning.
Perception-R1: Pioneering Perception Policy with Reinforcement Learning
En Yu (Huazhong University of Science and Technology), Wenbing Tao (Huazhong University of Science and Technology)
CodeRecognitionObject DetectionTransformerReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Explores the learning of perceptual strategies in multimodal LLM post-training with regularized RL, and proposes the Perception-R1 framework, achieving performance improvements across various visual perception tasks.
PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding
Jang Hyun Cho (Meta), Christoph Feichtenhofer (Meta)
CodeRecognitionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageVideoTextBenchmark
🎯 What it does: A fully open and reproducible visual language model, PerceptionLM (PLM), has been constructed, along with the release of a large-scale human-annotated video question-answering and spatiotemporal subtitle dataset, as well as a dedicated benchmark for fine-grained video understanding, PLM-VideoBench.
Performative Risk Control: Calibrating Models for Reliable Deployment under Performativity
Victor Li (New York University), Zhun Deng (University of California, Los Angeles)
CodeTabularFinance Related
🎯 What it does: A Performative Risk Control (PRC) framework is proposed for risk control of black-box machine learning models in environments with performativity.
PeRL: Permutation-Enhanced Reinforcement Learning for Interleaved Vision-Language Reasoning
Yizhen Zhang (Tsinghua University), Yujiu Yang (Tsinghua University)
CodeTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Proposes the PeRL method, utilizing a reinforcement learning framework with image sequence permutation and rollout filtering, trained on multi-image interactive visual-language reasoning tasks.
PermLLM: Learnable Channel Permutation for N:M Sparse Large Language Models
Lancheng Zou (Chinese University of Hong Kong), Bei Yu (Chinese University of Hong Kong)
CodeTransformerLarge Language ModelText
🎯 What it does: This study proposes the PermLLM framework, which utilizes learnable channel permutation (LCP) for post-training pruning of N:M semi-structured sparse large language models, directly minimizing the output error between the sparse model and the original model;
🎯 What it does: Proposed and implemented Permutation Equivariant Neural Graph Controlled Differential Equations (PENG-CDE) for representation learning of dynamic graphs.
Personalized Bayesian Federated Learning with Wasserstein Barycenter Aggregation
Ting Wei (Renmin University of China), Yifan Sun (Renmin University of China)
CodeFederated LearningImage
🎯 What it does: Proposes the FedWBA framework, which implements personalized Bayesian inference and Wasserstein barycenter aggregation in federated learning.
CodeRecommendation SystemRecurrent Neural NetworkTransformerLarge Language ModelReinforcement LearningContrastive LearningTextSequential
🎯 What it does: This paper proposes the ExRec framework for personalized exercise recommendation. The framework automatically annotates knowledge concepts (KC) using LLM, utilizes contrastive learning to obtain semantically rich exercises, problem-solving steps, and KC embeddings, trains a calibrated knowledge tracing (KT) model for instant knowledge state estimation, and uses the calibrated KT model as a reinforcement learning (RL) environment, enhancing the learning effect of continuous action RL with model-based value estimation (MVE).
Personalized Federated Conformal Prediction with Localization
Yinjie Min (Nankai University), Changliang Zou (Nankai University)
CodeFederated LearningTabular
🎯 What it does: A personalized federated quantile prediction framework (PFCP) is proposed, which combines personalized federated learning and local compliant prediction to provide a statistically valid prediction set for the target agent and achieve instance localization.
Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections
Wei Zhuo (Nanyang Technological University), Han Yu (Nanyang Technological University)
CodeFederated LearningSafty and PrivacyGraph Neural NetworkGraph
🎯 What it does: Designed and implemented FedAux, a personalized framework in the subgraph federated learning scenario, which captures the heterogeneity of local subgraphs by learning auxiliary projection vectors (APV) at each client, thereby achieving privacy-preserving client similarity assessment and personalized model aggregation.
Perturb a Model, Not an Image: Towards Robust Privacy Protection via Anti-Personalized Diffusion Models
Tae-Young Lee (Korea University), Gyeong-Moon Park (Korea University)
CodeGenerationSafty and PrivacyDiffusion modelImage
🎯 What it does: A model-level protection framework called APDM has been designed and implemented to prevent diffusion models from generating personalized outputs for specified subjects while maintaining generation quality.
PhySense: Sensor Placement Optimization for Accurate Physics Sensing
Yuezhou Ma (Tsinghua University), Mingsheng Long (Tsinghua University)
CodeOptimizationComputational EfficiencyTransformerFlow-based ModelTime SeriesPhysics Related
🎯 What it does: A two-stage framework called PhySense has been developed, which utilizes a flow matching reconstructor and projection gradient descent to optimize sensor placement, achieving a synergy between sparse observation reconstruction of physical fields and sensor deployment optimization.
🎯 What it does: The PCFM (Physics-Constrained Flow Matching) framework is proposed, which enforces any nonlinear equality constraints through constrained projection of intermediate states during inference based on a pre-trained flow matching generative model, achieving zero-shot hard constraints.
Physics-Driven Spatiotemporal Modeling for AI-Generated Video Detection
Shuhai Zhang (University of Science and Technology of China), Mingkui Tan (South China University of Technology)
CodeGenerationAnomaly DetectionDiffusion modelScore-based ModelVideoBenchmarkPhysics Related
🎯 What it does: A spatiotemporal modeling method based on the principle of physical conservation is proposed, defining and utilizing the Normalized Spatiotemporal Gradient (NSG) to detect AI-generated videos.
Physics-informed Neural Operator for Pansharpening
Xinyang Liu (Southeast University), Bo Liu (University of Electronic Science and Technology of China)
CodeRestorationTransformerImagePhysics Related
🎯 What it does: A physics-informed neural operator framework (PINO) is proposed to fuse high-resolution panchromatic images and low-resolution multispectral images for spatial-spectral joint reconstruction.
Physics-informed Value Learner for Offline Goal-Conditioned Reinforcement Learning
Vittorio Giammarino (Purdue University), Ahmed H Qureshi
CodeReinforcement LearningTabularPhysics Related
🎯 What it does: A physical information regularization based on Eikonal PDE is proposed for value function learning in offline goal-conditioned reinforcement learning, and it is integrated into the HIQL framework.
🎯 What it does: A PID control-based Langevin dynamics algorithm (PIDLD) was designed and validated to accelerate the sampling process of unsupervised generative models.
🎯 What it does: A secure model repair framework called SURE is proposed, allowing service providers and third-party repair institutions to collaboratively fix errors and biases in text-to-image diffusion models without disclosing user feedback, model weights, or repair knowledge.
Sourav Pal (University of Wisconsin Madison), Vikas Singh (University of Wisconsin Madison)
CodeTabularPhysics Related
🎯 What it does: A learnable efficient numerical integration rule (LearnQuad) is proposed in PINNs (Physics-Informed Neural Networks), achieving adaptive sampling and weak/strong form integration for low-dimensional PDE solving by parameterizing the adjustable weight function with neural network and utilizing the asymptotic expansion of orthogonal polynomials.
PIPE: Physics-Informed Position Encoding for Alignment of Satellite Images and Time Series in Typhoon Forecasting
Haobo Li (Hong Kong University of Science and Technology), Alexis Kai Hon Lau (Hong Kong University of Science and Technology)
CodeTransformerSupervised Fine-TuningVision Language ModelImageMultimodalityTime SeriesPhysics Related
🎯 What it does: A multimodal weather forecasting task based on satellite images and numerical time series is proposed, along with a lightweight Physical Information Position Encoding (PIPE) method.
PipeFusion: Patch-level Pipeline Parallelism for Diffusion Transformers Inference
Jiarui Fang (Tencent), WANG Jiannan
CodeGenerationTransformerDiffusion modelImage
🎯 What it does: This paper proposes PipeFusion, a patch-level pipeline parallel inference scheme for Diffusion Transformers, which significantly reduces inference latency and memory usage.
🎯 What it does: A particle image velocimetry (PIV) framework based on neural operators (PIVNO) is proposed, which directly maps particle image pairs to flow fields, avoiding the traditional issues of cost volume construction and local matching.
🎯 What it does: This paper proposes modeling the hippocampal cell population as a non-negative positional embedding, achieved through spectral decomposition of the multi-step random walk transition kernel, and utilizes the inner product of the embedding to approximate transition probabilities, representing a multi-scale cognitive map of the environment. Subsequently, a path planning algorithm based on gradient ascent and adaptive scale selection is designed and validated in simulated open fields and multi-obstacle mazes.
🎯 What it does: This paper proposes a neural network called PLEIADES, which uses orthogonal polynomial expansion for adaptive temporal convolution kernels, suitable for low-latency spatiotemporal classification and detection with event cameras.
PLMTrajRec: A Scalable and Generalizable Trajectory Recovery Method with Pre-trained Language Models
Tonglong Wei (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
CodeRestorationDomain AdaptationTransformerLarge Language ModelPrompt EngineeringTime Series
🎯 What it does: A scalable general trajectory recovery method based on pre-trained language models (PLMTrajRec) has been developed, capable of recovering missing points in sparse trajectories and achieving cross-domain generalization across different sampling intervals.
🎯 What it does: A parallelizable linear source-transition-labeling network (pLSTM) is proposed for long-range information propagation in multidimensional data (images, graphs).
🎯 What it does: A pluggable ReCAP module is proposed, which alternates between full evaluation and lightweight evaluation during the inference process of Masked Generative Models (MGMs) to reduce the computational load at each step and improve generation efficiency.
POCO: Scalable Neural Forecasting through Population Conditioning
Yu Duan (Massachusetts Institute of Technology), Kanaka Rajan (Harvard Medical School)
CodeTime SeriesBiomedical Data
🎯 What it does: A POCO model that combines a single neuron predictor and a global population encoder is proposed for cell-level time series prediction on multi-animal multi-session calcium imaging data.
Point or Line? Using Line-based Representation for Panoptic Symbol Spotting in CAD Drawings
Xingguang Wei (University of Science and Technology of China), Hongjie Zhang (Shanghai AI Laboratory)
CodeObject DetectionSegmentationTransformerImage
🎯 What it does: For the task of panoramic symbol localization in CAD drawings, VecFormer is proposed to construct type-independent line segment representations through line segment sampling, completing symbol detection and segmentation with joint instance and semantic branches.
🎯 What it does: A dual diffusion mask autoencoder called Point-MaDi is proposed for point cloud pre-training, which eliminates position encoding leakage and learns global semantics and local geometry.
🎯 What it does: An online streaming 3D reconstruction framework called Point3R is proposed, which utilizes explicit spatial pointer memory to achieve global point cloud generation from continuous images.
Polar Sparsity: High Throughput Batched LLM Inferencing with Scalable Contextual Sparsity
Susav Shrestha (Texas A&M University), A. L. Narasimha Reddy (Texas A&M University)
CodeTransformerLarge Language ModelText
🎯 What it does: For large-scale batch inference of LLMs, Polar Sparsity is proposed, achieving contextual sparsity by dynamically activating sub-networks and heads in the MLP and attention layers.
PolarQuant: Leveraging Polar Transformation for Key Cache Quantization and Decoding Acceleration
Songhao Wu (Renmin University of China), Rui Yan (Renmin University of China)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: The PolarQuant method is proposed to perform polar coordinate quantization on the key vectors in the KV cache of large language models, thereby reducing memory usage and accelerating decoding.
Policy Compatible Skill Incremental Learning via Lazy Learning Interface
Daehee Lee (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningMultimodality
🎯 What it does: A framework named SIL-C is proposed to maintain compatibility between policies and skills during the Skill Incremental Learning (SIL) process, allowing improvements in learned lower-level skills to seamlessly enhance the performance of higher-level policies without the need for retraining or structural adjustments.
Pool Me Wisely: On the Effect of Pooling in Transformer-Based Models
Sofiane ENNADIR, Lele Cao (Microsoft Gaming)
CodeClassificationRestorationSegmentationData SynthesisTransformerImageTextTime Series
🎯 What it does: This paper systematically studies the role of pooling layers in the Transformer model, providing a theoretical framework and closed-form upper bounds for pooling expressiveness, and empirically validating it in visual, natural language, and time series tasks.
Pose Splatter: A 3D Gaussian Splatting Model for Quantifying Animal Pose and Appearance
Jack Goffinet (Duke University), David Carlson
CodePose EstimationGaussian SplattingVideo
🎯 What it does: Proposes the Pose Splatter framework, which utilizes shape sculpting and 3D Gaussian splatting to achieve unlabelled, non-frame-by-frame optimization of 3D pose and appearance reconstruction for animals.
🎯 What it does: To address the problem of extreme perspective changes (where images have little or no overlap), the PoseCrafter framework is proposed, which assists camera pose estimation by synthesizing intermediate views.
Post Hoc Regression Refinement via Pairwise Rankings
Kevin Tirta Wijaya (Max Planck Institute for Informatics), Vahid Babaei (Max Planck Institute for Informatics)
CodeOptimizationDrug DiscoveryLarge Language ModelTabularAgriculture Related
🎯 What it does: A post-hoc correction framework called RankRefine is proposed, which enhances the accuracy of continuous attribute predictions in low-data environments by combining the predictions of a benchmark regressor with pairwise ranking information generated by an external ranker (such as LLMs or experts).
🎯 What it does: A dynamic stereo matching framework called PPMStereo based on Pick-and-Play Memory is proposed to achieve spatiotemporal consistency in disparity estimation for long-sequence videos.
Practical and Effective Code Watermarking for Large Language Models
Zhimeng Guo (Pennsylvania State University), Minhao Cheng (Pennsylvania State University)
CodeGenerationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: A practical watermarking framework ACW suitable for large language model code generation is proposed, which can embed and detect watermarks without exposing the LLM or prompt information.
Marcus Lassila (Chalmers University of Technology), Alexandre Graell i Amat (Chalmers University of Technology)
CodeSafty and PrivacyAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper studies and implements Bayesian optimal membership inference attacks (MIA) for i.i.d. and graph-structured data, particularly focusing on node-level MIA for Graph Neural Networks (GNN).
Practical Kernel Selection for Kernel-based Conditional Independence Test
Wenjie Wang (University of Melbourne), Feng Liu (University of Melbourne)
CodeOptimizationHyperparameter SearchTabular
🎯 What it does: This paper studies and implements a kernel parameter selection method based on power ratio (Power) to improve the kernel parameter setting in the Kernel-based Conditional Independence Test (KCI), thereby enhancing the effectiveness and robustness of CI testing.