🎯 What it does: Proposes the SRP-KGC method with Soft Reasoning Paths (SRP) and hierarchical ranking, using learnable embeddings to fill missing paths and improve the stability and accuracy of knowledge graph completion.
Solving QNP and FOND+ with Generating, Testing and Forbidding
Zheyuan Shi (Sun Yat-sen University), Yongmei Liu (Sun Yat-sen University)
CodeOptimizationBenchmark
🎯 What it does: Propose a Generate-Test-and-Forbid framework based on PRP to solve QNP and FOND+ planning problems, implementing SIEVE* termination testing and various optimization techniques.
🎯 What it does: Designed and implemented a multi-branch and similarity-constrained spatial-spectral fusion network (S3FNet) to fuse low-resolution multispectral images with high-resolution panchromatic images to generate high-resolution multispectral images;
SPoRt - Safe Policy Ratio: Certified Training and Deployment of Task Policies in Model-Free RL
Jacques Cloete (University of Oxford), Alessandro Abate (University of Oxford)
CodeSafty and PrivacyReinforcement LearningSequential
🎯 What it does: This paper proposes the SPoRt framework, which provides an upper bound on the probability of safety property violations in model-free RL environments through the maximum policy ratio, and trains task-specific policies based on this upper bound.
Squeezing Context into Patches: Towards Memory-Efficient Ultra-High Resolution Semantic Segmentation
Wang Liu (Hunan University), Shutao Li (Hunan University)
CodeSegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a single-branch UHR semantic segmentation method called SCPSeg, which utilizes a Context Compression Module (CSM) to compress global information into local patches, while incorporating an auxiliary super-resolution decoder and a Local Feature Alignment loss (LFA) to achieve a balance between high accuracy and low memory consumption;
ST-USleepNet: A Spatial-Temporal Coupling Prominence Network for Multi-Channel Sleep Staging
Jingying Ma (Chinese Academy of Sciences), Mengling Feng (Chinese Academy of Sciences)
CodeClassificationConvolutional Neural NetworkGraph Neural NetworkTime SeriesBiomedical Data
🎯 What it does: Proposed ST-USleepNet, a method that constructs multi-channel raw sleep signals into a spatiotemporal graph and utilizes a U-shaped network to simultaneously extract significant temporal features and spatial brain networks for sleep staging.
StarFT: Robust Fine-tuning of Zero-shot Models via Spuriosity Alignment
Younghyun Kim (Samsung), Jinwoo Shin (KAIST)
CodeClassificationDomain AdaptationRepresentation LearningLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Introduce 'Spurious Textual Alignment Regularization (StarFT)' during fine-tuning of large zero-shot vision-language models (e.g., CLIP) to suppress the learning of spurious features, thereby enhancing out-of-distribution (OOD) robustness.
State Feedback Enhanced Graph Differential Equations for Multivariate Time Series Forecasting
Jiaxu Cui (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education), Bo Yang (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education)
🎯 What it does: This paper proposes a state feedback-based graph differential equation (SF-GDE) for multivariate time series prediction, aiming to alleviate the over-smoothing problem in traditional graph networks.
Mingkai Tang (Hong Kong University of Science and Technology), Kaichen Zhang (Hong Kong University of Science and Technology)
CodeOptimizationGraphBenchmark
🎯 What it does: This paper proposes the stream-based multi-agent path finding (S-MAPF) problem for assembly lines, and presents the ASCBS algorithm based on conflict-driven search to solve this problem;
🎯 What it does: Propose a multi-scale node pair grouping strategy and a corresponding multi-scale knowledge distillation optimization loss to achieve efficient knowledge distillation from GNN to MLP;
T-T: Table Transformer for Tagging-based Aspect Sentiment Triplet Extraction
Kun Peng (Chinese Academy of Sciences), Philip S. Yu (University of Illinois at Chicago)
CodeRecognitionTransformerTextBenchmark
🎯 What it does: This paper proposes Table-Transformer (T-T), which encodes sentences into 2D tables and employs an improved Transformer layer (stripe attention + loop-shift) for relation learning, achieving superior performance in the Aspect Sentiment Triplet Extraction task.
T2S: High-resolution Time Series Generation with Text-to-Series Diffusion Models
Yunfeng Ge (Xidian University), Shirui Pan (Griffith University)
CodeGenerationTransformerDiffusion modelFlow-based ModelAuto EncoderTextTime Series
🎯 What it does: This paper proposes the T2S framework to achieve text-to-time series generation and constructs a fragment-level high-resolution text-time series pair dataset at the 600K scale.
Template-based Uncertainty Multimodal Fusion Network for RGBT Tracking
Zhaodong Ding (National Key Laboratory Of Opto Electronic Information Acquisition And Protection Technology), Jin Tang (National Key Laboratory Of Opto Electronic Information Acquisition And Protection Technology)
🎯 What it does: Propose a transferable adversarial attack method BMTC based on background mixing and temporal consistency for video action recognition models.
The Devil is in Fine-tuning and Long-tailed Problems: A New Benchmark for Scene Text Detection
Tianjiao Cao (Chinese Academy of Sciences), Yu Zhou (Nankai University)
CodeObject DetectionVision Language ModelAuto EncoderImageBenchmark
🎯 What it does: This paper addresses the performance gap between scene text detection in academic benchmarks and real-world applications by analyzing the fine-tuning gap and long-tailed distribution issues, proposing the Joint Dataset Learning (JDL) evaluation protocol and Long-Tailed Benchmark (LTB), and providing the MAEDet self-supervised baseline.
Toward Reliable Scientific Hypothesis Generation: Evaluating Truthfulness and Hallucination in Large Language Models
Guangzhi Xiong (University of Virginia), Aidong Zhang (University of Virginia)
CodeTransformerLarge Language ModelBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the TruthHypo benchmark and KnowHD framework to evaluate the authenticity and hallucination issues of large language models when generating scientific hypotheses.
Towards Anytime Retrieval: A Benchmark for Anytime Person Re-Identification
Xulin Li (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
CodeRetrievalTransformerMixture of ExpertsImageBenchmark
🎯 What it does: Proposed the AT-ReID task, constructed the AT-USTC large-scale dataset, and designed the Uni-AT model to achieve real-time person re-identification across day/night, seasonal, and long/short-term scenarios.
🎯 What it does: Propose the AutoSAM automatic sampling framework, which uses reinforcement learning to perform non-uniform sampling on historical behaviors in sequence recommendation, thereby enhancing the model's generalization ability.
🎯 What it does: This paper proposes an instant detection-rebalancing multi-modal learning framework (IPRM), which first instantly evaluates modality strength through two forward passes, and then instantly recalibrates modality weights based on the evaluation results, achieving real-time balanced training for modality imbalance.
🎯 What it does: Propose a neural simulator named CoPoNDP that can simultaneously address the effects of environmental and temporal distribution shifts on dynamical systems;
Towards Micro-Action Recognition with Limited Annotations: An Asynchronous Pseudo Labeling and Training Approach
Yan Zhang (Hefei University of Technology), Meng Wang (Hefei University of Technology)
CodeRecognitionConvolutional Neural NetworkVideo
🎯 What it does: Proposed an asynchronous pseudo-labeling and training framework APLT for semi-supervised learning in micro-action recognition, significantly reducing annotation requirements.
🎯 What it does: Propose a Regional Adaptive Feature Decoupling and Enhancement (RAFDE) strategy for UAV small target detection, which includes Boundary Transitional Region-enhanced Downsampling (BTRD) and Regional Adaptive Feature Fusion (RAFF) modules;
Towards Safer Pretraining: Analyzing and Filtering Harmful Content in Webscale Datasets for Responsible LLMs
Sai Krishna Mendu (Microsoft), Parag Agrawal (Microsoft)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper conducts systematic audits of large-scale web datasets such as Common Crawl, C4, and FineWeb for harmful content, constructing a three-dimensional (Safe, Topical, Toxic) harm classification framework. It generates TTP prompts using GPT-4 Omni Prompt, trains a Longformer-based HarmFormer model to achieve high-accuracy filtering of long texts, and releases the TTP-Eval evaluation set and HAVOC multi-harm generation benchmark.
🎯 What it does: Propose a training-free Fourier phase diffusion model that uses the Fourier phase spectrum of the content image as a condition to modulate intermediate samples during the diffusion process, achieving high-quality style transfer while preserving the content structure.
TrajCogn: Leveraging LLMs for Cognizing Movement Patterns and Travel Purposes from Trajectories
Zeyu Zhou (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
CodeClassificationRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTime SeriesSequential
🎯 What it does: Proposed the TrajCogn model, which leverages LLM (GPT-2) to learn spatiotemporal trajectories through trajectory prompts and a trajectory semantic embedder, enabling the model to recognize motion patterns and travel purposes from trajectories, and directly apply to multiple tasks (trip duration prediction, destination prediction, similar trajectory retrieval, trajectory classification).
TsCA: On the Semantic Consistency Alignment via Conditional Transport for Compositional Zero-Shot Learning
Miaoge Li (Hong Kong Polytechnic University), Song Guo (Hong Kong University of Science and Technology)
CodeClassificationRecognitionRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: By constructing three groups of distributions of image patches, combinations, and primitives, and leveraging conditional transport with cyclic consistency regularization, the TsCA framework achieves fine-grained alignment and reasoning for Compositional Zero-Shot Learning (CZSL).
Unleashing the Semantic Adaptability of Controlled Diffusion Model for Image Colorization
Xiangcheng Du (Fudan University), Cheng Jin (Fudan University)
CodeRestorationDiffusion modelImage
🎯 What it does: This paper proposes a semantics-adaptive controlled diffusion model called SeAda, which automatically restores grayscale images into colorful, semantically consistent color images.
🎯 What it does: Propose a low-bit quantization framework (including Connected Quantized Block and Shifted Logarithmic Redistribution Quantizer) for efficiently performing deepfake detection on resource-constrained edge devices.
Unsupervised Feature Transformation via In-context Generation, Generator-critic LLM Agents, and Duet-play Teaming
Nanxu Gong (Arizona State University), Yanjie Fu (Arizona State University)
CodeRepresentation LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTabular
🎯 What it does: Propose an unsupervised feature transformation framework that leverages two LLMs, a generator and a critic, to generate feature transformations through dialogue and iteratively optimize them;
Unveiling Maternity and Infant Care Conversations: A Chinese Dialogue Dataset for Enhanced Parenting Support
Bo Xu (Dalian University of Technology), Hongfei Lin (Dalian University of Technology)
CodeGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: Construct a Chinese maternal and infant care dialogue dataset called MicDialogue and propose a knowledge-driven dialogue generation model called Kng-BART
🎯 What it does: This paper proposes a framework (VO-MASD) for automatically learning multi-agent skills from offline multi-task data, which constructs 3D or hierarchical codebooks through dynamic grouping and VQ-VAE, enabling simultaneous capture of subgroup and temporal hierarchical collaborative abstractions.
🎯 What it does: Propose the VeRecycle framework, which automatically recovers safety guarantees from existing probabilistic neural Lyapunov certificates after local changes in system dynamics, avoiding the need to re-verify the entire system.
🎯 What it does: In the cross-view geolocation task, the VimGeo network is proposed, combining Vision Mamba backbone, Channel Group Pooling (CGP), and Dynamic Weighted Batch-tuple Loss (DWBL) to achieve efficient and accurate localization.
Visual Perturbation and Adaptive Hard Negative Contrastive Learning for Compositional Reasoning in Vision-Language Models
Xin Huang (Nanyang Normal University), Ya Wang (Nanyang Normal University)
CodeRepresentation LearningSupervised Fine-TuningVision Language ModelContrastive LearningMultimodality
🎯 What it does: By mapping the hard negative semantic shifts in the text layer to the visual space, generating image-level hard negatives, and employing adaptive hard negative contrastive learning to enhance VLM performance on compositional reasoning tasks.
Volumetric Axial Disentanglement Enabling Advancing in Medical Image Segmentation
Xingru Huang (Hangzhou Dianzi University), Xiaoshuai Zhang (Ocean University of China)
CodeSegmentationConvolutional Neural NetworkTransformerBiomedical Data
🎯 What it does: The paper proposes a post-refinement module based on volume axial decoupling (PaR) to improve the results of 3D medical image segmentation.
Andreas Roth (TU Dortmund University), Thomas Liebig (TU Dortmund University)
CodeGraph Neural NetworkGraph
🎯 What it does: This paper derives the essence of multi-channel (MIMO) graph convolution (MIMO-GC) and proposes a localized MIMO graph convolution (LMGC) framework based on this derivation; subsequently, it theoretically proves that LMGC is injective under a single computational graph and generates linearly independent representations when using multiple computational graphs, and experimentally verifies its performance on graph-level and node-level tasks.
X-KAN: Optimizing Local Kolmogorov-Arnold Networks via Evolutionary Rule-Based Machine Learning
Hiroki Shiraishi (Yokohama National University), Masaya Nakata (Yokohama National University)
CodeOptimizationBenchmark
🎯 What it does: Propose X-KAN, a partitioned function approximation method that integrates the Kolmogorov-Arnold network (KAN) with the XCSF rule system, using local KAN models in rule consequents for approximation and employing evolutionary algorithms to adaptively partition rule antecedents;
🎯 What it does: Proposes UNPrompt, a zero-shot general graph anomaly detection framework that can detect abnormal nodes in any other arbitrary graph without any fine-tuning or labels, trained only on a single graph dataset.