Navigating Through Paper Flood: Advancing LLM-Based Paper Evaluation Through Domain-Aware Retrieval and Latent Reasoning
Wuqiang Zheng (University Of Science And Technology Of China), Fuli Feng (University Of Science And Technology Of China)
CodeRecommendation SystemTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposed and implemented the PaperEval framework, which utilizes large language models (LLMs) for automatic paper evaluation, achieving more accurate assessments through three modules: domain-aware retrieval, latent reasoning, and progressive ranking optimization.
NaVLA$^2$: A Vision-Language-Audio-Action Model for Multimodal Instruction Navigation
Jugang Fan (South China University of Technology), Mingkui Tan (South China University of Technology)
CodeExplainability and InterpretabilityRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelImageTextMultimodalityChain-of-ThoughtAudio
π― What it does: Designed a multi-modal instruction navigation task, MINav, and constructed a dataset containing 43.9K navigation samples with visual, linguistic, image, and audio prompts; proposed the NaVLA 2 model, integrating spatial semantic audio encoding and CoThinkAct reasoning decoding to achieve interpretable multi-step action planning.
π― What it does: Propose a robust cross-modal retrieval framework named NIRNL, integrating two modules: cross-modal margin preservation (CMP) and neighborhood-aware instance refinement (NIR). By leveraging neighborhood consistency, the training samples are finely divided into clean, easy-hard, and noisy subsets, with dedicated losses designed for each subset to enhance cross-modal retrieval performance under noisy label environments.
Xijia Tang (National University of Defense Technology), Chenping Hou (National University of Defense Technology)
CodeClassificationImage
π― What it does: Proposed the UIDPLL (Unreliable Instance Related Partial Labels) problem and designed the NLAP method to incrementally expand and prune candidate labels.
π― What it does: This paper proposes an algorithm called Nested Depth Search (NDS), a generalized extension of NMCS, which can explore the search space more extensively in higher-level simulations by setting depth d and step size s; it also provides time complexity analysis of NDS and an exact probability distribution calculation method for the Left Move Problem (LMP).
NeSTR: A Neuro-Symbolic Abductive Framework for Temporal Reasoning in Large Language Models
Feng Liang (China Academy of Launch Vehicle Technology), Xiang Zhao (National University of Defense Technology)
CodeExplainability and InterpretabilityLarge Language ModelTextBenchmark
π― What it does: Propose NeSTR, a neuro-symbolic temporal reasoning framework that combines structured symbolic representations with reflective reasoning to enhance the temporal reasoning capabilities of large language models.
π― What it does: In semi-supervised multi-view classification, we propose to utilize class prototypes generated by neural network convergence (Neural Collapse) as priors to guide feature learning for unlabeled samples, and combine attention mechanisms, cross-view contrastive learning, and evidence-based reliable fusion to achieve fine-grained calibration of class distributions and uncertainty modeling.
Neural Graph Navigation for Intelligent Subgraph Matching
Yuchen Ying (Zhejiang University), Mingli Song (Zhejiang University)
CodeGraph Neural NetworkTransformerGraph
π― What it does: Proposed Neural Graph Navigation (NeuGN), introducing neural networks into the subgraph matching enumeration phase to achieve structure-aware search navigation.
CodeComputational EfficiencyGraph Neural NetworkBiomedical DataPhysics Related
π― What it does: Developed a real-time soft tissue deformation prediction framework that combines Kelvinlet analytical solutions with neural networks, enhancing model accuracy through residual learning and regularization.
NeuralGS: Bridging Neural Fields and 3D Gaussian Splatting for Compact 3D Representations
Zhenyu Tang (Peking University), Li Yuan (Hong Kong University of Science and Technology)
CodeCompressionGaussian Splatting
π― What it does: Propose NeuralGS, a post-training compression method for 3D Gaussian Splatting, which re-encodes high-dimensional Gaussian attributes using multiple small MLPs after clustering, significantly reducing model size.
CodeGraph Neural NetworkTime SeriesPhysics Related
π― What it does: Proposed a neural operator framework named NEURALOM for high-accuracy, long-term ocean (Subseasonal-to-Seasonal) simulation and prediction.
NeuroBridge: Bio-Inspired Self-Supervised EEG-to-Image Decoding via Cognitive Priors and Bidirectional Semantic Alignment
Wenjiang Zhang (Beijing University of Posts and Telecommunications), Suyu Zhong (Beijing University of Posts and Telecommunications)
CodeRetrievalRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImageBiomedical Data
π― What it does: Propose the NeuroBridge framework, which leverages self-supervised learning combined with cognitive prior enhancement and shared semantic projection to achieve cross-modal alignment between EEG and images for zero-shot visual decoding.
CodeGenerationOptimizationRepresentation LearningGaussian SplattingVideoPhysics Related
π― What it does: Reconstruct and simulate the geometry, appearance, and physical properties of deformable objects using video data to build a physical digital twin.
New Synthetic Goldmine: Hand Joint Angle-Driven EMG Data Generation Framework for Micro-Gesture Recognition
Nana Wang (BeiHang University), Hao Su (BeiHang University)
CodeRecognitionGenerationData SynthesisConvolutional Neural NetworkRecurrent Neural NetworkGenerative Adversarial NetworkTime SeriesSequentialBiomedical Data
π― What it does: Built a conditional generation framework based on SeqEMG-GAN, which uses hand joint angle sequences to drive the generation of high-fidelity, temporally consistent EMG signals;
Next Generation Active Learning: Mixture of LLMs in the Loop
Yuanyuan Qi (Monash University), Lan Du (Monash University)
CodeClassificationTransformerLarge Language ModelText
π― What it does: Designed a fully manual-annotation-free active learning framework called MoLLIA, which generates high-quality labels by aggregating outputs from multiple lightweight LLMs, and enhances model robustness through negative learning and annotation discrepancy mechanisms;
NICE: Neural Implicit Craniofacial Model for Orthognathic Surgery Prediction
Jiawen Yang (ShanghaiTech University), Hongjiang Wei (ShanghaiTech University)
CodePoint CloudBiomedical DataComputed Tomography
π― What it does: Proposed a craniofacial model called NICE based on neural implicit functions for predicting facial appearance after orthognathic surgery.
Nighttime Flare Removal via Wavelet-Guided and Gated-Enhanced Spatial-Frequency Fusion Network
Yun Liu (Southwest University), Weisi Lin (Southwest University)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: Propose a spatial-frequency fusion network (WGSF-Net) based on multi-level wavelet enhancement and gated attention, specifically designed for nighttime lens flare removal.
π― What it does: Propose the Embedding CFR algorithm, which utilizes a pre-trained low-dimensional embedding space to solve strategies in incomplete information games (e.g., poker).
NODiff: Neural Operator Diffusion for Multispectral Image Fusion
Junming Hou (Southeast University), Liang-Jian Deng (University Of Electronic Science And Technology Of China)
CodeRestorationDiffusion modelImage
π― What it does: Proposed a diffusion model called NODiff based on neural operators (FNO) to efficiently perform multispectral image fusion (pansharpening), and achieved parameter-efficient fine-tuning through two-stage pre-training combined with a lightweight adapter.
Noisy Correspondence Learning with Modality Gap Direction Correction
Wuyuqing Wang (Xidian University), Erkun Yang (Xidian University)
CodeRetrievalConvolutional Neural NetworkRecurrent Neural NetworkVision Language ModelContrastive LearningMultimodality
π― What it does: Propose a robust learning framework named MGCS for cross-modal retrieval, which corrects cross-modal similarity by leveraging sample-level alignment drift (SAD), and adaptively segments noisy samples through dynamic regularization.
Piotr Gorczyca (TUD Dresden University of Technology), Hannes Strass (TUD Dresden University of Technology)
Code
π― What it does: Proposed a new framework combining non-monotonic S4F logic with stance logic, termed S4F Stance Logic, and provided its semantics, minimal model determination, and complexity analysis;
CodeLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextGraphBenchmark
π― What it does: Proposes the NoReGeo benchmark to evaluate whether large language models and vision-language models possess innate geometric intuition without relying on reasoning or algebraic operations;
Not Just for Archiving: Provable Benefits of Reusing the Archive in Evolutionary Multi-objective Optimization
Shengjie Ren (Nanjing University), Chao Qian (University of Birmingham)
CodeOptimizationBenchmark
π― What it does: This paper investigates reusing the archive in multi-objective evolutionary algorithms (MOEA) to accelerate the search process, demonstrating that it provides polynomial speedup for SMS-EMOA on OneJumpZeroJump and its variants, and verifying its superiority through experiments.
Not Just Whatβs There: Enabling CLIP to Comprehend Negated Visual Descriptions Without Fine-Tuning
Junhao Xiao (Central China Normal University), Zejiang He (National University Of Defense Technology)
CodeTransformerVision Language ModelContrastive LearningMultimodality
π― What it does: Propose the CLIPGLASSES framework, enabling CLIP to understand visual descriptions with negations without modifying the original parameters.
NOTAM-Evolve: A Knowledge-Guided Self-Evolving Optimization Framework with LLMs for NOTAM Interpretation
Maoqi Liu, Kaiquan Cai (Beihang University)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation
π― What it does: Proposed and implemented a self-evolving framework called NOTAM-Evolve for deeply parsing highly concise, industry-specific NOTAM text into structured information.
Note2Chat: Improving LLMs for Multi-Turn Clinical History Taking Using Medical Notes
Yang Zhou (Institute of High Performance Computing, Agency for Science, Technology and Research), Yong Liu (Institute of High Performance Computing, Agency for Science, Technology and Research)
CodeTransformerLarge Language ModelSupervised Fine-TuningBiomedical DataElectronic Health Records
π― What it does: Proposed the Note2Chat framework based on medical notes, utilizing notes to drive LLMs for multi-round medical history collection and differential diagnosis.
π― What it does: Designed and implemented a multi-interest learning framework based on neural processes, NP-MiSR, for session recommendation, which can adaptively learn multi-interest representations of sessions and integrate information from similar sessions.
NTSFormer: A Self-Teaching Graph Transformer for Multimodal Isolated Cold-Start Node Classification
Jun Hu (National University of Singapore), Bingsheng He (National University of Singapore)
CodeClassificationGraph Neural NetworkTransformerMixture of ExpertsMultimodalityGraph
π― What it does: This paper proposes a self-teaching framework called NTSFormer based on graph Transformer, aimed at solving the cold start node classification problem in multimodal graphs, capable of simultaneously handling node isolation and modality missing;
NucEL: Single-Nucleotide ELECTRA-Style Genomic Pre-training for Efficient and Interpretable Representations
Ke Ding (Australian National University), Jiayu Wen (Australian National University)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningBiomedical Data
π― What it does: Proposed and implemented NucEL, a genome pre-training framework based on ELECTRA, which employs single-nucleotide tokenization, is pre-trained specifically on the human genome, and is fine-tuned on multiple genomic functional prediction tasks.
NURBGen: High-Fidelity Text-to-CAD Generation Through LLM-Driven NURBS Modeling
Muhammad Usama (German Research Center for Artificial Intelligence), Muhammad Zeshan Afzal (German Research Center for Artificial Intelligence)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelTextMesh
π― What it does: Propose the NURBGen framework, which directly translates natural language prompts into editable NURBS parameters via large language models and generates high-precision 3D CAD models.
π― What it does: Proposes an object-centric latent action learning framework that leverages object slot decomposition from unlabeled videos to learn agents' latent actions;
Oblivionis: A Lightweight Learning and Unlearning Framework for Federated Large Language Models
Fuyao Zhang (Nanyang Technological University), Qiang Yang (Hong Kong Polytechnic University)
CodeFederated LearningSafty and PrivacyTransformerLarge Language ModelTextBenchmark
π― What it does: Developed a lightweight federated learning and machine learning forgetting framework named Oblivionis for training and targeted forgetting in federated LLMs.
OccamVTS: Distilling Vision Models to 1% Parameters for Time Series Forecasting
Sisuo Lyu (Hong Kong University of Science and Technology Guangzhou), Yuxuan Liang (Hong Kong University of Science and Technology Guangzhou)
CodeKnowledge DistillationRepresentation LearningTransformerVision Language ModelAuto EncoderTime Series
π― What it does: Transfer 1% of the parameters from a large visual model to a lightweight network through knowledge distillation, achieving efficient time series forecasting.
Offline Multi-Objective Bandits: From Logged Data to Pareto-Optimal Policies
Ji Cheng (City University of Hong Kong), Bo Xue (City University of Hong Kong)
CodeOptimizationReinforcement Learning
π― What it does: Proposed an offline multi-objective contextual bandit algorithm called OffMOB, aiming to learn Pareto optimal strategies and generate the complete Pareto frontier from static log data.
π― What it does: Proposed the OFL-SAM2 framework, leveraging online few-shot learning and adaptive fusion to achieve zero-shot medical image segmentation
OmniDPO: A Preference Optimization Framework to Address Omni-Modal Hallucination
Junzhe Chen (Tsinghua University), Lijie Wen (Hong Kong University of Science and Technology)
CodeOptimizationReinforcement Learning from Human FeedbackLarge Language ModelMultimodality
π― What it does: Designed and implemented the OMNIDPO framework, leveraging multimodal preference learning (text, video, audio) to reduce multimodal hallucinations and enhance reasoning capabilities.
π― What it does: Proposes a unified event camera data representation learning framework called OmniEvent, which can be directly applied to multiple tasks such as classification, optical flow estimation, and point cloud registration.
OmniScale: Scaling Any Modality Model Training with Model-Centric Distributed Recipe Zoo
Qianli Ma (ByteDance Seed), Xin Liu (ByteDance Seed)
CodeComputational EfficiencyLarge Language ModelMixture of ExpertsImageVideoTextMultimodalityAudio
π― What it does: Proposed the OmniScale framework for efficient training of large language models (LLMs) across any modality, supporting modular configuration and n-dimensional parallelism.
On Logical Extrapolation for Mazes with Recurrent and Implicit Networks
Brandon Knutson (Colorado School of Mines), Daniel McKenzie (Colorado School of Mines)
CodeExplainability and InterpretabilityRecurrent Neural Network
π― What it does: Investigated the logical extrapolation capabilities of recursive and implicit networks in maze-solving tasks, examined whether they truly learned scalable algorithms, and analyzed their potential convergence dynamics.
On Modality Weighting and Specificity for Multi-Modal Entity Alignment
Yu Xing (Nanjing University), Tieke He (Nanjing University)
CodeKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerMixture of ExpertsContrastive LearningMultimodality
π― What it does: Propose a multi-modal entity alignment framework HUMEA based on hierarchical Mixture of Experts (MoE) and unimodal distillation, which can adaptively weight different modalities while preserving modality-specific information.
Marcin Podhajski (Institute of Fundamental Technological Research Polish Academy of Sciences), Tomasz PaweΕ Michalak (Institute of Fundamental Technological Research Polish Academy of Sciences)
CodeSafty and PrivacyAdversarial AttackGraph Neural NetworkGraph
π― What it does: Propose a strategy in GNN model stealing attacks under extreme query budgets, first locally acquiring the encoder and then utilizing query selection to maximize information extraction.
On the Approximation Ratio of Optimal Fixed-Price Mechanisms for Single and Multi-Unit Bilateral Trade
Giordano Giambartolomei (King's College London), Bart de Keijzer (King's College London)
CodeOptimization
π― What it does: This paper improves mathematical programming techniques to re-estimate the approximate ratio of optimal social welfare under fixed-price mechanisms in single-item and multi-item bilateral trade, enhancing the lower and upper bounds for single items, and providing better lower bounds for multi-items (especially two items) and improvements in symmetric cases.
On the Calibration of Image Semi-Supervised Learning Models
Mehrab Mustafy Rahman (University of Illinois Chicago), Cornelia Caragea (University of Illinois Chicago)
CodeClassificationImage
π― What it does: Proposes CalibrateMix, a targeted mixup strategy based on training dynamics and sample difficulty, to enhance confidence calibration and classification accuracy in semi-supervised learning models.
π― What it does: This paper studies the existence of core stability in approval-based multi-elections and proposes a method based on mixed integer linear programming (MILP) to determine whether the core is empty.
On the Evaluation of Capability Estimation Methods for Large Language Models
Qiang Hu (Tianjin University), Yongqiang Lyu (University of Luxembourg)
CodeLarge Language ModelTextBenchmark
π― What it does: This paper proposes AEBench, an unlabeled capability estimation benchmark for large language models (LLMs), covering 12 AutoEval methods.
π― What it does: Studied the learning dynamics of stochastic gradient descent (SGD) with label noise in two-layer linear networks, revealing a two-phase transition from lazy to rich regimes, and generalized this idea to Sharpness-Aware Minimization (SAM).
On the Probabilistic Learnability of Compact Neural Network Preimage Bounds
Luca Marzari (University of Verona), Alessandro Farinelli (University of Verona)
CodeSafty and PrivacyReinforcement Learning from Human FeedbackBenchmark
π― What it does: Propose a probabilistic method called RF-ProVe based on random forests and active resampling to approximate the preimage boundary of neural networks with high confidence.
OncoCoT: A Temporal-causal Chain-of-Thought Dataset for Oncologic Decision-Making
Peiru Yang (Tsinghua University), Yongfeng Huang (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataElectronic Health RecordsBenchmarkChain-of-Thought
π― What it does: Constructed the OncoCoT dataset, a long-chain-of-thought dataset tailored for tumor diagnosis and treatment, along with its benchmark OncoEval.
π― What it does: This paper proposes a synthesis-free open-source method for attributing the origin of synthetic images, which uses a random encoder-decoder to simulate fingerprints of trillions of generative models, and learns frequency-domain fingerprint features to verify unknown models.
π― What it does: Introduce a first-order generation strategy, achieving a one-step direct mapping from noise to actions through residual reconstruction of MeanFlow, compatible with Q-learning in offline reinforcement learning.
One2Seq: One-Token Wise Decoder for Efficient Scene Text Recognition
Zhibin Ma (Sun Yat-sen University), Xugong Qin (Shenzhen University)
CodeRecognitionTransformerImageText
π― What it does: Propose a one-token-wise decoder called One2Seq, which performs autoregressive decoding using a single context token to address issues such as attention drift, slow decoding speed, and lack of global context in traditional AR decoders.
Ruoyu Wu (University of Sydney), Hequn Wang (University of Sydney)
CodeOptimizationGraph
π― What it does: This paper proposes an online capacity general matching with knapsack (OCGMK) problem and designs a new online capacity-knapsack allocation (OCKA) algorithm to maximize rewards without knowing future arrival orders and reward distributions.
Online Conformal Selection with Accept-to-Reject Changes
Kangdao Liu (University of Macau), Hongxin Wei (Southern University of Science and Technology)
CodeTextTabularBiomedical Data
π― What it does: Propose an online acceptable-reject modification conformal selection method OCS-ARC, satisfying the ARC constraint of irrevocably selected candidates and controlling FDR
CodeAnomaly DetectionAuto EncoderTime SeriesPhysics Related
π― What it does: Studies unsupervised anomaly detection in spectral data streams, first evaluating existing methods on multi-dimensional table streams, then proposing and implementing a framework named OnlineBootKNN specifically for real-time detection of spectral anomalies.
OPERA: A Reinforcement Learning--Enhanced Orchestrated Planner-Executor Architecture for Reasoning-Oriented Multi-Hop Retrieval
Yu Liu (Chinese Academy of Sciences), Zhiyuan Ma (Huazhong University of Science and Technology)
CodeRetrievalLarge Language ModelReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose a multi-agent planning-execution architecture called OPERA for multi-hop retrieval tasks, which can dynamically decompose problems, rewrite queries, and filter retrieval results.
OptiHive: Ensemble Selection for LLM-Based Optimization via Statistical Modeling
Maxime Bouscary (Massachusetts Institute of Technology), Saurabh Amin (Massachusetts Institute of Technology)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language Model
π― What it does: The OptiHive framework leverages LLM to simultaneously parallelly generate candidate solvers, problem instances, and verification tests, filtering out interpretable components and then selecting the optimal optimization solver through statistical inference.
Sanmay Das (Virginia Polytechnic Institute and State University), Yuang Zhang (George Mason University)
CodeOptimizationReinforcement Learning
π― What it does: This study designs and solves an optimal audit strategy within a master-slave game framework to maximize the leader's utility or social welfare when adversarial agents select the worst equilibrium;
Optimized Algorithms for Text Clustering with LLM-Generated Constraints
Chaoqi Jia (RMIT University), Kok-Leong Ong (Western Sydney University)
CodeOptimizationTransformerLarge Language ModelText
π― What it does: Propose a text clustering framework based on LLM automatically generating set-based must-link and cannot-link constraints, and design a penalty-based local search clustering algorithm to utilize these constraints.
Luise Ge (Washington University in St. Louis), Yevgeniy Vorobeychik (Washington University in St. Louis)
CodeOptimizationTabular
π― What it does: The study investigates distortion in social choice under a linear utility model, proposes new voting rules (such as Maximum Coordinate Majority (MCP), Linear Stable Lottery (LSLR), and Pure Stable Lottery (PSLR)), designs instance-optimal deterministic and randomized algorithms, and evaluates them on real-world data.
OptScale: Probabilistic Optimality for Inference-time Scaling
Youkang Wang (PolySmart Group), Xiao-Yong Wei (Sichuan University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper proposes the OPTSCALE framework for inference time scaling based on probabilistic optimality, which dynamically determines the optimal number of samples to enhance the inference performance of large language models.
OR-R1: Automating Modeling and Solving of Operations Research Optimization Problem via Test-Time Reinforcement Learning
Zezhen Ding (Hong Kong University of Science and Technology), Tianlong Chen (University of North Carolina)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
π― What it does: Propose the OR-R1 framework, which uses a small amount of labeled data for Supervised Fine-Tuning (SFT) and further trains on unlabeled data through Test-Time Group Relative Policy Optimization (TGRPO), automatically completing the modeling and solving of operations research optimization problems.
π― What it does: Designed and implemented the ORACLE framework, generating and verifying high-quality training data for multi-step reasoning through templated prompts and a symbolic reasoning engine to enhance the reasoning capabilities of large language models.
Organ-Aware Routing Mixture-of-Retrieval Augmented Generation for Fetal Ultrasound Reporting
Bin Pu (Hunan University), Kenli Li (Southern Medical University)
CodeGenerationAnomaly DetectionTransformerMixture of ExpertsVision Language ModelContrastive LearningBiomedical DataUltrasoundRetrieval-Augmented Generation
π― What it does: Proposed the FetusR dataset and the ORM-RAG model for multi-organ fetal ultrasound report generation;
ORTCL: Towards Continual Learning of Time Series Foundation Models on Streaming Data via Orthogonal Rotation
Li Lin (Southeast University), Kaiwen Xia (Southeast University)
CodeRepresentation LearningTime Series
π― What it does: Propose the ORTCL method, which achieves lossless continual learning on time series base models using an orthogonal rotation matrix;
π― What it does: Investigate the oscillatory behavior in large-scale flow models during the inversion process, proposing the Oscillation Inversion method to achieve training-agnostic image and video enhancement and editing.
OscuFit: Learning to Fit Osculating Implicit Quadrics for Point Clouds
Rao Fu (Nanyang Technological University), Jianmin Zheng (Nanyang Technological University)
CodeOptimizationGraph Neural NetworkPoint Cloud
π― What it does: This paper proposes a learning-based framework for estimating local surface differential properties of point clouds, utilizing osculating implicit quadrics to simultaneously output normals and curvature.
OSVBench: Benchmarking LLMs on Specification Generation Tasks for Operating System Verification
Shangyu Li, Jiasi Shen (Georgia Institute of Technology)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Propose the OSVBENCH benchmark to evaluate large language models in generating complete formal specifications required for correctness verification of operating system kernel functions;
OTI: A Model-free and Visually Interpretable Measure of Image Attackability
Jiaming Liang (University of Macau), Chi-Man Pun (University of Macau)
CodeSegmentationExplainability and InterpretabilityAdversarial AttackImageBiomedical Data
π― What it does: Propose a model-free and visually interpretable image adversarial susceptibility metric called OTI, which directly evaluates the degree to which an image is vulnerable to adversarial perturbations based on the texture intensity and area ratio of semantic objects.
π― What it does: Investigate the background interference problem in wide-view videos for few-shot action recognition, and propose the Otter model, which enhances the subject through the Compound Segmentation Module (CSM) and reconstructs temporal relationships via the Temporal Reconstruction Module (TRM) to improve performance.
Out-of-Context Misinformation Detection via Variational Domain-Invariant Learning with Test-Time Training
Xi Yang (Xidian University), Hong Han (Xidian University)
CodeDomain AdaptationAnomaly DetectionTransformerLarge Language ModelVision Language ModelAuto EncoderContrastive LearningMultimodality
π― What it does: Proposes a framework combining variational domain-invariant learning with test-time training (VDT) for cross-domain detection of out-of-context (OOC) misinformation in the news domain.
π― What it does: Propose a single-stage adaptive reinforcement learning framework, SPARC, for achieving robust control in out-of-distribution (OOD) environments.
OwlCap: Harmonizing Motion-Detail for Video Captioning via HMD-270K and Caption Set Equivalence Reward
Chunlin Zhong (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
CodeGenerationLarge Language ModelReinforcement LearningVision Language ModelVideoTextMultimodality
π― What it does: Propose OwlCap through two-phase HMD-270K dataset construction with CSER reward-based reinforcement learning to address the imbalance between motion and detail in video captioning.
OX-MABSR: A Benchmark for Open-domain Explainable Multimodal Aspect-Based Sentiment Reasoning
Xinjing Liu (Beijing University of Posts and Telecommunications), Ruifan Li (Beijing University of Posts and Telecommunications)
CodeExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Propose an open-domain explainable multi-modal aspect-based sentiment reasoning task (OX-MABSR) and construct a corresponding high-quality dataset (OX-MABSR-Bench), while designing a multi-modal LLM framework (MABSR-LLM) to accomplish open-vocabulary aspect-sentiment pair prediction, dual-layer (perceptual-cognitive) explanation generation, and reasoning path generation.
Pedro IlΓdio (KU Leuven), Celine Vens (Universidade de SΓ£O Paulo)
CodeDrug DiscoveryBiomedical Data
π― What it does: Proposed the Oxytrees model tree, which uses a proxy matrix to compress the interaction matrix, significantly accelerating the training and inference of bilateral learning.
π― What it does: This paper proposes an unsupervised point cloud semantic segmentation framework called P-SLCR, which achieves consistent structural learning and inference through the construction of a learnable prototype library.
PA-FAS: Towards Interpretable and Generalizable Multimodal Face Anti-Spoofing via Path-Augmented Reinforcement Learning
Yingjie Ma (Shenzhen University), Zitong Yu (Great Bay University)
CodeAnomaly DetectionExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality
π― What it does: Propose the PA-FAS framework, combining reasoning path enhancement and answer shuffling, to achieve interpretable and generalizable learning for multi-modal face anti-spoofing under limited annotations.
Ju Jia (Southeast University), Guang Cheng (Southeast University)
CodeSafty and PrivacyRepresentation LearningGraph Neural NetworkPrompt EngineeringContrastive LearningGraph
π― What it does: A privacy-aware graph prompt learning framework named PAGPL is developed for efficient graph prompt learning on differentially private perturbed graphs.
Pairing-free Group-level Knowledge Distillation for Robust Gastrointestinal Lesion Classification in White-Light Endoscopy
Qiang Hu (Wuhan National Laboratory for Optoelectronics, Huazhong University of Science and Technology), Zhiwei Wang (Wuhan National Laboratory for Optoelectronics, Huazhong University of Science and Technology)
CodeClassificationKnowledge DistillationConvolutional Neural NetworkTransformerContrastive LearningBiomedical Data
π― What it does: This paper proposes PaGKD, a group-level knowledge distillation framework that does not require paired images, to transfer diagnostic information from narrow-band imaging (NBI) to white-light imaging (WLI) models, thereby improving the performance of gastrointestinal lesion classification.
Palimpsest: Reconciling the CISS Trilemma for Incremental Nuclei Segmentation
Jiajia Li (Shenzhen University), Huisi Wu (Shenzhen University)
CodeSegmentationBiomedical Data
π― What it does: Proposes the Palimpsest framework for incremental nucleus segmentation without requiring examples, while maintaining the model size unchanged.
Panda: Test-Time Adaptation with Negative Data Augmentation
Ruxi Deng (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
CodeDomain AdaptationImage
π― What it does: Proposed a test-time adaptation method called Panda, which suppresses prediction bias caused by image distortion through negative data augmentation (NDA), and can be easily integrated into existing TTA frameworks.
π― What it does: This study investigates high-resolution fusion (pansharpening) of remote sensing images under thin cloud contamination, proposing a unified end-to-end framework called Pan-TCR to simultaneously address cloud removal and resolution enhancement.
Paper Folding Puzzles: Can Multimodal Large Language Models Perform Spatial Reasoning?
Dibin Zhou (Hangzhou Normal University), Fuchang Liu (University of Nottingham Ningbo China)
CodeTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
π― What it does: Proposed the Paper Folding Puzzles (PFP) benchmark to evaluate the ability of multimodal large language models in reasoning from 2D to 3D space.
ParaDySe: A Parallel Strategy Switching Framework for Dynamic Sequences in Transformer-based Large Language Models
Zhixin Ou (National University of Defense Technology), Baihui Liu (National University of Defense Technology)
CodeComputational EfficiencyTransformerLarge Language ModelTextBiomedical Data
π― What it does: Propose the ParaDySe framework to achieve dynamic parallelism strategy hot switching at the Transformer level, addressing out-of-memory (OOM) issues and communication bottlenecks for variable-length sequences;
Parallel Training Time-to-First-Spike Spiking Neural Networks
Kaiwei Che (Peking University), Yonghong Tian (Peking University)
CodeClassificationSpiking Neural NetworkImage
π― What it does: Designed a fully parallel Time-to-First-Spike (TTFS) neural network training framework and proposed a temporal selection decoder based on membrane potential.
CodePose EstimationOptimizationSimultaneous Localization and MappingGraphBenchmark
π― What it does: Propose a fully parallel Riemannian ADMM algorithm called PRADMM with closed-form solutions for subproblems, designed for large-scale pose graph optimization (PGO)
ParaMETA: Towards Learning Disentangled Paralinguistic Speaking Styles Representations from Speech
Haowei Lou (University of New South Wales), Lina Yao (University of New South Wales)
CodeClassificationRecognitionGenerationRepresentation LearningConvolutional Neural NetworkRecurrent Neural NetworkTransformerVision Language ModelTextMultimodalityAudio
π― What it does: Propose the ParaMETA framework to learn and control disentangled speaking style embeddings in speech, supporting multi-task recognition and TTS control with text/speech prompts.
Parameter-Free Clustering via Self-Supervised Consensus Maximization
Lijun Zhang (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)
CodeAuto EncoderContrastive LearningImage
π― What it does: Proposes a completely parameter-free hierarchical clustering framework SCMax, integrating self-supervised consensus maximization to automatically generate and evaluate clustering structures.
Parametric Pareto Set Learning for Expensive Multi-Objective Optimization
Ji Cheng (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
CodeOptimizationBenchmark
π― What it does: Propose a Parametric Pareto Set Learning with Multi-Objective Bayesian Optimization (PPSL-MOBO) framework that can learn the entire Pareto set varying with parameters in one go, infer Pareto solutions in real-time for any parameter, and significantly reduce the demand for expensive function evaluations.
Pareto-Grid-Guided Large Language Models for Fast and High-Quality Heuristics Design in Multi-Objective Combinatorial Optimization
Ha Minh Hieu (Hanoi University of Science and Technology), Huynh Thi Thanh Binh (FPT Software AI Center)
CodeOptimizationLarge Language ModelPrompt EngineeringBenchmark
π― What it does: Propose an automatic heuristic design framework named MPaGE based on large language models (LLM) for multi-objective combinatorial optimization problems (MOCOP), guiding heuristic evolution through Pareto front grid and semantic clustering to balance solution quality and runtime efficiency.
PARS: Partial-Label-Learning-inspired Recommender Systems
Shanshan Ye (University of Technology Sydney), Jie Lu (University of Technology Sydney)
CodeRecommendation SystemTransformerSequential
π― What it does: This paper proposes a framework called PARS that constructs a recommendation system using only user browsing history, addressing the problem of recommendation without explicit purchase labels.
Delong Zhao (Harbin Institute of Technology), Jun Yu (Harbin Institute of Technology)
CodeClassificationExplainability and InterpretabilityLarge Language ModelVision Language ModelContrastive LearningImage
π― What it does: Developed a partially shared concept bottleneck model (PS-CBM) that achieves interpretable image classification through multimodal concept generation, concept sharing strategies, and concept efficiency metrics.
π― What it does: Propose PartialNet, which achieves fewer parameters and lower FLOPs without accuracy degradation by proportionally splitting feature maps across channels and separately performing convolution and attention.
π― What it does: Proposes Progressive-Adaptive Spectral Augmentation (PASA), a reinforcement learning framework that treats data augmentation as a Markov Decision Process (MDP), for adaptive augmentation of automated auscultation data;
π― What it does: Propose a generative speech enhancement framework called PASE based on the pre-trained WavLM model, leveraging its inherent phonetic prior to alleviate hallucination issues under low signal-to-noise ratio (SNR) conditions;
PASS: Probabilistic Agentic Supernet Sampling for Interpretable and Adaptive Chest X-Ray Reasoning
Yushi Feng (University of Hong Kong), Lequan Yu (University of Hong Kong)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerReinforcement LearningAgentic AIContrastive LearningMultimodalityBiomedical Data
π― What it does: Designed and implemented the PASS framework, utilizing probabilistic proxy supersampling to achieve interpretable, adaptive, and computationally cost-controllable multimodal chest X-ray reasoning systems.
PatchET: Learning Enzyme Temperature Properties Through Patch-Based Neural Architectures
Ziqi Zhang (Jiangnan University), Zhaohong Deng (Macquarie University)
CodeProtein Structure PredictionConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningSequentialBiomedical DataBenchmark
π― What it does: Propose the PatchET model, which employs a patch-based two-stage architecture to directly predict the optimal temperature, stability, and temperature range of enzymes from amino acid sequences, and release newly constructed benchmark datasets for optimal temperature (10,371 samples) and temperature range (1,818 samples).
PathFLIP: Fine-grained Language-Image Pretraining for Versatile Computational Pathology
Fengchun Liu (Harbin Institute of Technology), Yongbing Zhang (Harbin Institute of Technology)
CodeClassificationRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityBiomedical Data
π― What it does: Propose PathFLIP, a fine-grained language-image pre-training framework specifically designed for whole slide images (WSI), which achieves precise alignment between images and text without requiring region-level annotations.