Bidirectional Bounded-Suboptimal Heuristic Search with Consistent Heuristics
Shahaf S. Shperberg (Ben-Gurion University of the Negev), Dor Atzmon (Bar-Ilan University)
CodeOptimization
π― What it does: Proposes a bidirectional heuristic search framework called WBAE* suitable for tolerable suboptimal search, which introduces a weight Ξ» to regulate the heuristic error term d based on BAE*, and provides a series of theoretical proofs and experimental analysis.
π― What it does: Propose a semi-supervised medical image segmentation framework BCSI based on weak-to-strong consistency learning, achieving effective interaction between labeled and unlabeled data through semantic-space perturbation, channel selection routing, and bidirectional channel interaction.
π― What it does: This paper proposes the BiCoD framework, which improves upon issues in comment recommendation caused by knowledge transfer pollution and homogenization of behavior features due to score distribution bias.
Binary Message Passing for Generalizable Semi-Supervised Graph Anomaly Detection
Jingyuan Zhang (Institute of Software Chinese Academy of Sciences), Fengjun Zhang (Institute of Software Chinese Academy of Sciences)
CodeAnomaly DetectionGraph Neural NetworkGraph
π― What it does: Proposed a Binary Message Passing (BMP) framework that enhances anomaly information propagation and improves semi-supervised graph anomaly detection by constructing binary trees based on anomaly probability for message routing.
Binary Split Categorical Feature with Mean Absolute Error Criteria in CART
Peng Yu (University of Electronic Science and Technology of China), Jesse Read (Ecole Polytechnique Institut Polytechnique de Paris)
CodeOptimizationComputational EfficiencyTabular
π― What it does: Studied the binary splitting of classification features under the MAE criterion, proved the infeasibility of unsupervised numerical encoding, and proposed an exact and efficient algorithm based on Unimodal Cost 2-Median.
BiO-HMC: Dynamic Human-Machine Collaboration for Consensus Decision-Making via Bilevel Optimization
Yinghui Pan (Shenzhen University), Mingwei Lin (Fujian Normal University)
CodeOptimizationImage
π― What it does: Propose the BiO-HMC framework, modeling the consensus decision-making task as a bi-level optimization to dynamically integrate answers and proactively select the most valuable questions in human-machine collaboration.
π― What it does: Propose a dynamic prompting strategy based on reinforcement learning, enabling the model to adaptively generate context-aware prompts for each medical image, thus achieving efficient VLM adaptation under few-shot scenarios.
π― What it does: Propose a training set search framework based on hierarchical data servers and bilateral mode matching (BMM) for unsupervised domain adaptation tasks (person/vehicle re-identification and object detection)
BIQ: Bisection Interval Quantization for Communication-efficient Federated Learning
Luyang Gai (Xi'an Jiaotong University), Zihao Zhou (Xi'an Jiaotong University)
CodeFederated LearningImageTabular
π― What it does: Proposes two non-uniform quantization methods, BIQ and WBIQ, based on binary interval quantization, specifically designed for federated learning to improve communication efficiency and reduce error.
BitDP: Ultra-low-bit Communication for Data Parallelism in LLM Training
Xiaozhe Ren (Hong Kong University of Science and Technology), Qiong Luo (Hong Kong University of Science and Technology)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose the BitDP system to achieve 1-bit/2-bit ultra-low precision gradient quantization and communication, significantly reducing communication overhead during LLM training while maintaining high accuracy;
BLADE: A Behavior-Level Data Augmentation Framework with Dual Fusion Modeling for Multi-Behavior Sequential Recommendation
Yupeng Li (University of Science and Technology of China), Shijin Wang (University of Science and Technology of China)
CodeRecommendation SystemTransformerMixture of ExpertsContrastive LearningVideoTextSequential
π― What it does: Propose the BLADE framework, combining early and intermediate dual-stage fusion with behavior-level data augmentation to address the issues of behavioral heterogeneity and sparsity in multi-behavior sequence recommendation.
π― What it does: An end-to-end dual-branch framework named JFD3 was designed, jointly performing feature domain deblurring and object detection. It leverages features from clear images to supervise the restoration of features from blurred images, and improves the detection accuracy of blurred infrared UAV targets through a frequency domain structure guidance module.
Boomda: Balanced Multi-objective Optimization for Multimodal Domain Adaptation
Jun Sun, Xiang Gao (Zhejinag Lab)
CodeDomain AdaptationMultimodality
π― What it does: This paper proposes an algorithm called Boomda for heterogeneous multi-modal domain adaptation, aiming to address the domain shift problem caused by the scarcity of labeled multi-modal data.
π― What it does: Proposed the ASR-TRA framework, inserting learnable prompts into the decoder of the Whisper model, generating multiple candidate transcriptions via temperature-controlled stochastic decoding, and using CLAP-based semantic reward-driven reinforcement learning (RL) updates to achieve unsupervised test-time adaptation.
Bot Meets Shortcut: How Can LLMs Aid in Handling Unknown Invariance OOD Scenarios?
Shiyan Zheng (Xi'an Jiaotong University), Junhang Huang (Beijing Institute of Technology)
CodeClassificationDomain AdaptationGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
π― What it does: This paper addresses the "shortcut" problem in social robot detection by systematically evaluating the impact of text feature shortcomings on model robustness, and proposes an adversarial data augmentation (CDA) strategy based on large language models to alleviate the shortcomings effect.
BrainHGT: A Hierarchical Graph Transformer for Interpretable Brain Network Analysis
Jiajun Ma (Anhui University), Shengbing Pei (Anhui University)
CodeAnomaly DetectionExplainability and InterpretabilityGraph Neural NetworkTransformerBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
π― What it does: Propose a BrainHGT model based on hierarchical graph Transformers for interpretable brain network analysis from local brain regions to global functional communities.
BrainLMM: A Label-Free Framework for Mapping Multi-Semantic Representation in the Human Visual Cortex
Tan Gao (Beijing Institute of Technology), Guoyuan Yang (Beijing Institute of Technology)
CodeExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerDiffusion modelBiomedical DataMagnetic Resonance Imaging
π― What it does: Proposed the BrainLMM framework, achieving label-free, voxel-based multi-semantic mapping to explore how neurons in the human visual cortex respond to multiple semantic concepts.
Break the Tie: Learning Cluster-Customized Category Relationships for Categorical Data Clustering
Mingjie Zhao (Hong Kong Baptist University), Yiu-ming Cheung (BNU-HKBU United International College)
CodeTabular
π― What it does: This paper proposes a method called DISC, specifically designed for clustering tasks involving discrete (categorical) data, by improving distance metrics through learning subspace class relationships for each cluster, thereby achieving more accurate clustering.
π― What it does: Propose an asymmetric Kronecker compressive sensing (AKCS) model and a measurement-aware cross-attention (MACA) module, integrating them into MEUNet to achieve high-quality image compressive sensing reconstruction.
Breaking Model Lock-in: Cost-Efficient Zero-Shot LLM Routing via a Universal Latent Space
Cheng Yan, Yanyong Zhang (University Of Science And Technology Of China)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: Propose the ZeroRouter framework to address the model lock problem in LLM routing, enabling zero-shot rapid integration of new models and intelligent query allocation across multiple model pools.
π― What it does: Proposed a unified graph OOD detection framework called UniGOD, which can adaptively select geometric spaces and achieve efficient discrete detection by leveraging neural SDE dynamic uncertainty.
π― What it does: This paper addresses the fairness issue in link prediction by proposing an evaluation framework based on exposure fairness and a post-processing method.
BREPS: Bounding-Box Robustness Evaluation of Promptable Segmentation
Andrey Moskalenko (Lomonosov Moscow State University), Vlad Shakhuro (Lomonosov Moscow State University)
CodeSegmentationAdversarial AttackPrompt EngineeringImageBiomedical Data
π― What it does: This paper evaluates the robustness of prompt-based segmentation models using real user-drawn bounding box annotations and proposes a white-box bounding box adversarial attack method called BREPS to measure the model's sensitivity to variations in box inputs.
π― What it does: High-quality completion from incomplete 3D shapes to complete shapes is achieved by constructing a Schrodinger bridge model in the latent space; during this process, Depth-Enhanced VQ-VAE is first used to compress 3D shapes into structured latent vectors, followed by optimal transport via potential diffusion Schrodinger bridge in the latent space; finally, the latent codes are decoded back into complete 3D surfaces.
π― What it does: Propose a multimodal test-time adaptation framework called BriMPR based on prompt adjustment, which first calibrates monomodal feature distributions using modality-specific prompts, and then enhances modality alignment through cross-modal mask recombination and instance-level contrastive learning.
π― What it does: Propose a lightweight and interpretable BONE framework that utilizes optimization-guided low-level feature extraction and high-level feature learning with a small number of learnable parameters to accomplish multi-view clustering tasks.
Bridging Synthetic and Real Routing Problems via LLM-Guided Instance Generation and Progressive Adaptation
Jianghan Zhu (Singapore Management University), Xiaoli Li (Nanyang Technological University)
CodeData SynthesisDomain AdaptationOptimizationLarge Language ModelSupervised Fine-TuningPrompt EngineeringGraph
π― What it does: This paper proposes the EvoReal framework, which generates synthetic data with structures similar to real VRP instances using LLM-driven evolutionary algorithms, and transfers pre-trained neural combinatorial optimization (NCO) models from uniformly distributed synthetic data to real TSPLib and CVRPLib instances through a phased stepwise fine-tuning approach;
Bridging the Modality Reliability Gap in Drug-Target Interaction Prediction via a Confidence-aware Multimodal Fusion Framework
Jie Yang (ShanghaiTech University), Zhen Cheng (Shanghai Institute of Materia Medica)
CodeDrug DiscoveryTransformerMultimodalityBiomedical Data
π― What it does: Designed a confidence-aware multimodal fusion framework named DrugCMF to address the modality reliability gap in drug-target interaction prediction;
Bridging the Tokenizer Gap: Semantics and Distribution-aware Knowledge Transfer for Unbiased Cross-Tokenizer Distillation
Huazheng Wang (Beijing University of Posts and Telecommunications), Dacheng Tao (AGH University of Krakow)
CodeKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: This paper proposes a cross-tokenizer knowledge distillation framework called SEDI, which can efficiently transfer knowledge between teacher and student models even when they use different tokenizers.
π― What it does: Propose the BulletTime4D framework, which integrates spiking cameras with stereo RGB cameras to achieve dynamic scene rendering at high spatiotemporal resolution;
C3RL: Rethinking the Combination of Channel-independence and Channel-mixing from Representation Learning
Shusen Ma, Yu Kang (University Of Science And Technology Of China)
CodeRepresentation LearningContrastive LearningTime Series
π― What it does: This paper proposes a representation learning framework called C3RL, which jointly utilizes channel mixing (CM) and channel independent (CI) strategies, aligning positive and negative samples in time series data through a SimSiam-style two-branch network;
CABTO: Context-Aware Behavior Tree Grounding for Robot Manipulation
Yishuai Cai (National University of Defense Technology), Minglong Li (National University of Defense Technology)
CodeRobotic IntelligenceLarge Language ModelVision Language ModelMultimodality
π― What it does: This paper proposes a framework for automatically constructing a complete and consistent behavior tree (BT) system, formally defines the BT grounding problem, and designs the CABTO method to achieve automatic matching and verification of high-level planning and low-level control strategies for robotic manipulation tasks.
π― What it does: Proposes CAD-VAE, a method within the variational autoencoder framework that achieves fair disentanglement between target attributes and sensitive attributes by introducing a co-related latent variable z_R and directly minimizing conditional mutual information, while supporting the generation of fair counterfactuals and fine-grained image editing.
CADTrack: Learning Contextual Aggregation with Deformable Alignment for Robust RGBT Tracking
Hao Li (Army Engineering University Of Pla), Huchuan Lu (Dalian University Of Technology)
CodeObject TrackingTransformerMixture of ExpertsMultimodality
π― What it does: Propose a novel RGBT object tracking framework CADTrack to address issues of modal differences, feature fusion, and spatial mismatch, thereby enhancing all-weather tracking performance.
π― What it does: Propose a unified weight adjustment framework, and based on this framework, introduce two parameter-efficient fine-tuning methods: Pre-Diag and SORA.
CAMA: Enhancing Mathematical Reasoning in Large Language Models with Causal Knowledge
Lei Zan (Huawei Noah's Ark Lab), Lujia Pan (Huawei Noah's Ark Lab)
CodeExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelTextBenchmarkChain-of-Thought
π― What it does: By constructing a mathematical causal graph (MCG) and embedding it into LLM, the accuracy of LLMs in complex mathematical reasoning tasks is enhanced.
π― What it does: Proposed the CAMAR environment, a multi-agent path planning benchmark that supports continuous actions and runs efficiently on GPUs;
CAMERA: Multi-Matrix Joint Compression for MoE Models via Micro-Expert Redundancy Analysis
Yuzhuang Xu (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
CodeCompressionMixture of ExpertsText
π― What it does: This paper proposes three compression methods for Mixture-of-Experts (MoE) large language models: CAMERA, CAMERA-P, and CAMERA-Q, which can achieve structured pruning and mixed-precision quantization at the micro-expert level without additional training.
Canyu Chen (Illinois Institute of Technology), Kai Shu (Emory University)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Propose the concept of 'Editing Attack', viewing knowledge editing as a security threat to large language models (LLMs), and construct the EditAttack dataset to systematically evaluate misinformation injection and bias injection.
Can Molecular Evolution Mechanism Enhance Molecular Representation?
Kun Li (Wuhan University), Jia Wu (Macquarie University)
CodeDrug DiscoveryGraph Neural NetworkTransformerBiomedical Data
π― What it does: Construct a molecular evolutionary network MEvoN, and enhance molecular representations using its evolutionary paths and label information for molecular property prediction.
π― What it does: Proposed a topology-aware self-training method called TA-GST, which selects reliable pseudo-labels by leveraging node classification scores and neighbor label distributions.
π― What it does: Propose the CANDI framework, integrating false positive mining and spatiotemporal adaptive modules to achieve test-time adaptation in multivariate time series anomaly detection.
CANVAS: A Benchmark for Vision-Language Models on Tool-Based User Interface Design
Daeheon Jeong (KAIST), Juho Kim (KAIST)
CodeVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmark
π― What it does: Proposed the CANVAS benchmark to evaluate the multi-round interaction capabilities of vision-language models (VLMs) in tool-driven UI design, covering two task categories: design replication and design modification.
CapeNext: Rethinking and Refining Dynamic Support Information for Category-Agnostic Pose Estimation
Yu Zhu (Sun Yat-sen University), Bo Tang (Guilin University of Technology)
CodePose EstimationGraph Neural NetworkTransformerVision Language ModelImageText
π― What it does: Proposes CapeNext, which dynamically improves keypoint embeddings by leveraging query images and category descriptions to enhance category-agnostic pose estimation.
CaPro: Curvilinear-aware Prompt Learning with Single Unlabeled Image for Cost-effective Curvilinear Structure Segmentation
Zhuangzhuang Chen (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
CodeSegmentationTransformerPrompt EngineeringAuto EncoderImageBiomedical Data
π― What it does: Propose a two-stage self-supervised Curvilinear-aware Prompt Learning (CaPro) framework without fine-tuning, which adapts the Segment Anything Model (SAM) to curve structure segmentation tasks under the condition of only being given a single unlabeled image.
Careful Queries, Credible Results: Teaching RAG Models Advanced Web Search Tools with Reinforcement Learning
Yuqin Dai (Tsinghua University), Shuai Lu (Tsinghua University)
CodeRetrievalTransformerLarge Language ModelReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
π― What it does: Proposed the WebFilter framework, modeling the retrieval process as a Markov Decision Process (MDP), enabling large language models (LLMs) to autonomously generate precise queries with advanced search operators (e.g., site:, after:, AND, OR), filtering online rumors and improving retrieval accuracy.
π― What it does: Proposed a curvature-enhanced self-supervised learning framework named CASL for 3D point cloud anomaly detection and general point cloud tasks;
CastX: Cohort-Level Causal Inference Meets Statistical Testing for Faithful and Reliable GNN Explanations
Guanyuan Yu, Gang Kou (Southwestern University Of Finance And Economics)
CodeExplainability and InterpretabilityGraph Neural NetworkReinforcement LearningGraph
π― What it does: Propose the CastX framework, combining collective layer causal inference with non-parametric permutation tests, using reinforcement learning to iteratively remove edges and generate credible and concise GNN explanation subgraphs.
CAT-Net: A Cross-Attention Tone Network for Cross-Subject EEG-EMG Fusion Tone Decoding
Yifan Zhuang (Sony Interactive Entertainment), Jiawei Ju (Shanghai Center for Brain Science and Brain-Inspired Technology)
CodeClassificationRecognitionDomain AdaptationRecurrent Neural NetworkTransformerMultimodalityBiomedical Data
π― What it does: This paper proposes a Chinese four-tone recognition framework based on EEG and EMG bimodal fusion (CAT-Net), achieving tone classification under both audible and silent speech conditions;
Catastrophic Forgetting in Kolmogorov-Arnold Networks
Mohammad Marufur Rahman (Wake Forest University), Fan Yang (Wake Forest University)
CodeTransformerLarge Language ModelImageTextBenchmark
π― What it does: Systematic theoretical analysis and experimental validation of catastrophic forgetting in Kolmogorov-Arnold networks (KAN) for continual learning, and the proposal of the KAN-LoRA adapter for continual editing of language models.
CATCH: A Controllable Theme Detection Framework with Contextualized Clustering and Hierarchical Generation
Rui Ke (Chinese University of Hong Kong), Haizhou Li (Chinese University of Hong Kong)
CodeRecommendation SystemTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextSequentialFinance Related
π― What it does: Proposes the controllable topic detection framework CATCH, achieving cross-dialogue consistent and personalized topic identification through context-aware topic segmentation, user preference-enhanced clustering, and hierarchical topic generation.
CaTFormer: Causal Temporal Transformer with Dynamic Contextual Fusion for Driving Intention Prediction
Sirui Wang (Beijing Jiaotong University), Jie Liu (Beijing Jiaotong University)
CodeAutonomous DrivingTransformerVideoTime Series
π― What it does: Proposes CaTFormer, a causal temporal Transformer for driving intent prediction, explicitly modeling the mutual causal relationships between driver behavior and the external environment.
CATS: Category-Aware Token-level Steering for Training-Free Redundancy Reduction in Large Reasoning Models
Mengfei Zhang (Zhejiang University), Zhenglin Wang (Southeast University)
CodeCompressionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose CATS, a training-agnostic, lightweight category-based token-level activation-driven method for compressing redundant reasoning chains in large inference models.
π― What it does: Proposed a causal disentanglement-based remote sensing change detection network, CDDGNet, for unsupervised change detection in cross-domain environments;
Causal Reward Adjustment: Mitigating Reward Hacking in External Reasoning via Backdoor Correction
Ruike Song (Institute of Software Chinese Academy of Sciences), Wenwen Qiang (Institute of Software Chinese Academy of Sciences)
CodeExplainability and InterpretabilityAuto EncoderText
π― What it does: Proposes the Causal Reward Adjustment (CRA) method using a causal inference framework and backdoor correction, leveraging a sparse autoencoder to extract interpretable semantic features and eliminate reward hacking issues, thereby improving the accuracy of external reasoning systems.
Causal-HalBench: Uncovering LVLMs Object Hallucinations Through Causal Intervention
Zhe Xu (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)
CodeRestorationVision Language ModelImageTextMultimodalityBenchmark
π― What it does: This paper constructs a causal analysis framework, proposes Visual Content Intervention (VCI) to automatically generate counterfactual samples, and creates the Causal-HalBench benchmark for systematically evaluating object hallucination issues in LVLMs.
π― What it does: In domain generalization semantic segmentation tasks, by performing frequency domain (DCT) decomposition on Vision Foundation Models, causal and non-causal features are extracted, and causal features are optimized using learnable causal markers.
Causality Matters: How Temporal Information Emerges in Video Language Models
Yumeng Shi (Nanyang Technological University), Wenya Wang (Nanyang Technological University)
CodeComputational EfficiencyVision Language ModelVideoMultimodalityBenchmark
π― What it does: This paper investigates the temporal understanding mechanisms in video language models (VideoLM), revealing that they do not rely on positional encoding but instead generate sequential sensitivity through causal attention; meanwhile, it proposes two efficient inference strategies based on this mechanism;
Causally-Aware Attribute Completion for Incomplete Federated Graph Clustering
Jingxin Liu (Hainan University), Xiangyan Tang (Hainan University)
CodeFederated LearningGraph Neural NetworkGraph
π― What it does: This study proposes IFedGC, a framework capable of performing node-level clustering in federated graph clustering scenarios with missing attributes;
Causally-Grounded Dual-Path Attention Intervention for Object Hallucination Mitigation in LVLMs
Liu Yu (University of Electronic Science and Technology of China), Gillian Dobbie (University of Electronic Science and Technology of China)
CodeExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningMultimodality
π― What it does: Propose the Owl framework by causal modeling of visual and textual attention, using fine-grained attention intervention and dual-path contrastive decoding to reduce object misreporting in large vision-language models.
CauVQ: Causal Vector Quantization for Graph OOD Generalization
Weihong Zhang (Shanxi University), Xian Yang (The University of Manchester)
CodeDomain AdaptationExplainability and InterpretabilityGraph Neural NetworkGraph
π― What it does: Proposes a graph learning framework based on causal vector quantization (CauVQ), which enhances the generalization and interpretability of graph neural networks in out-of-distribution (OOD) environments by mapping local substructures to discrete codebooks and leveraging causal discovery with counterfactual regularization.
π― What it does: Proposes a connected component-aware hypergraph contrastive learning framework named CCAHCL, which constructs multi-layer representations for nodes, hyperedges, and connected components, and balances the learning of different-scale connected components through hierarchical contrastive loss.
CCFQA: A Benchmark for Cross-Lingual and Cross-Modal Speech and Text Factuality Evaluation
Yexing Du (Harbin Institute of Technology), Yang Xiang (Pengcheng Laboratory)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmarkAudio
π― What it does: Proposed CCFQA, a cross-lingual and cross-modal (speech and text) factuality evaluation benchmark, and used it to systematically assess the fact consistency of existing multimodal large language models.
CD-DPE: Dual-Prompt Expert Network Based on Convolutional Dictionary Feature Decoupling for Multi-Contrast MRI Super-Resolution
Xianming Gu (Guizhou University), Yi Chen (Guizhou University)
CodeSuper ResolutionConvolutional Neural NetworkMixture of ExpertsBiomedical DataMagnetic Resonance Imaging
π― What it does: Achieve multi-contrast MRI super-resolution reconstruction using a dual-prompt expert network and convolutional dictionary feature decoupling method, precisely leveraging structural information from high-resolution reference images to guide the restoration of low-resolution target images.
π― What it does: Propose the CellStream framework, which jointly learns low-dimensional embeddings and cell dynamics to recover continuous cell trajectories from single-cell snapshot data.
Center-Outward q-Dominance: A Sample-Computable Proxy for Strong Stochastic Dominance in Stochastic Multi-Objective Optimisation
Robin van der Laag (Leiden University), Yingjie Fan (Leiden University)
CodeOptimizationTabularBenchmark
π― What it does: Propose the center-outward q-advantage relationship as a computable proxy for strong first-order stochastic dominance, validated in two scenarios: hyperparameter optimization and noisy multi-objective evolutionary algorithms.
Dieter Vandesande (Vrije Universiteit Brussel), Bart Bogaerts (Universitat de Girona)
CodeOptimizationExplainability and InterpretabilityBenchmark
π― What it does: Added verifiable proof logging functionality to the branch-and-bound MaxSAT solver MAXCDCL, enabling evidence generation for look-ahead bound estimation and MDD encoding.
CharBench: Evaluating the Role of Tokenization in Character-Level Tasks
Omri Uzan (Stanford University), Yuval Pinter (Ben-Gurion University of the Negev)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: Constructed a benchmark framework named CHARBENCH containing 175,000 examples with two character-level tasks (counting and localization), systematically evaluating character reasoning capabilities of multiple public and proprietary large language models.
CHASE: Contextual History for Adaptive and Simple Exploitation in Large Language Model Jailbreaking
Zhiqiang Hao (Nanjing University), Vincent Ng (University of Texas at Dallas)
CodeAdversarial AttackLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Proposes CHASE, a jailbreak method based on multi-turn dialogue, which first generates 'jailbroken history' on vulnerable LLMs and then transfers this history to target LLMs, inducing them to output harmful content while maintaining conversational coherence.
CHIMERA: Controllable High-quality Image-Mask Extraction for Reliable Diffusion-based Anomaly Synthesis
JoungBin Lee (KAIST AI), Seungryong Kim (KAIST AI)
CodeGenerationData SynthesisAnomaly DetectionTransformerVision Language ModelDiffusion modelImageTextMultimodality
π― What it does: This paper proposes the CHIMERA framework, which can synthesize high-quality, controllable, and generalizable anomaly samples in industrial images based on natural language instructions.
Chinese Two-part Allegorical Sayings Reading Comprehension: Exploration from Reasoning to Metaphor
Dongyu Su (Key Laboratory of Smart Farming for Agricultural Animals), Ying Sha (Huazhong Agricultural University)
CodeTransformerLarge Language ModelContrastive LearningTextBenchmark
π― What it does: Constructed the Chinese Two-Fable Idiom (TPAS) Reading Comprehension Dataset (CTRC) and proposed a Multi-Perspective TPAS Contrastive Learning Network (MTCLN) to address three major challenges: rhetoric identification, logical reasoning, and metaphor understanding.
CiNuSeg: Class Incremental Nuclei Segmentation via Anchor-driven Consistency Learning with Dual Region Regularization
Xuexin Wu (Shenzhen University), Jing Qin (Hong Kong Polytechnic University)
CodeSegmentationKnowledge DistillationConvolutional Neural NetworkBiomedical Data
π― What it does: Proposes the CiNuSeg method to achieve incremental nuclear segmentation, addressing the balance between old class forgetting and new class learning.
Federico Di Valerio (Sapienza University), Roberto Capobianco (Sony AI)
CodeClassificationExplainability and InterpretabilityComputational EfficiencyContrastive LearningImage
π― What it does: Propose a sample-free self-explaining prototype network called CIP-Net, achieving continual learning through a shared prototype layer.
Paul GΓΆlz, Philipp C. Verpoort (Sortition Foundation)
CodeOptimizationTabular
π― What it does: This paper studies the two-stage sampling problem when holding citizens' assemblies in countries such as Germany, and designs an algorithm that ensures equal sampling probabilities for all citizens while limiting the number of participating cities, the number of letters, and upper limits under municipal registry constraints.
π― What it does: Propose a cross-modal knowledge separation and alignment method (CKDA), which separates the cross-modal shared knowledge and the modality-specific knowledge exclusive to visible/infrared modalities through visual prompting, and achieves dual alignment of new and old knowledge without data replay, thereby mitigating catastrophic forgetting in visible-infrared lifelong person re-identification;
π― What it does: Propose a contrastive learning-based dynamic multi-modal data fusion model (CL-DMDF), aiming to achieve efficient and adaptive multi-modal information fusion in scenarios with missing or noisy modalities;
CL-Guard: Defending DNNs Against Backdoors via Fine-Grained Neuron Analysis and Collaborative Dual-Network Learning
Jie Xiao (Zhejiang University of Technology), Fan Terry Zhang (Zhejiang University of Technology)
CodeSafty and PrivacyKnowledge DistillationAdversarial AttackConvolutional Neural NetworkImage
π― What it does: Proposes the CL-Guard pre-deployment defense framework, which selects critical neurons through recursive hierarchical partitioning, sparsely trains non-critical neurons, and introduces dual-network collaborative learning to eliminate backdoors in deep neural networks.
π― What it does: Two methods, prototype-based calibrated distillation (PGCD) and bidirectional aligned prototype distillation (DAPD), are proposed to address the category incremental segmentation problem in medical imaging.
π― What it does: Propose a pure point cloud scene generation framework that directly generates object bounding boxes, categories, and latent features using class-partitioned VQ-VAE and latent space flow matching models, avoiding external database retrieval.
Clean-Label Physical Backdoor Attacks with Data Distillation
Thinh Dao (VinUniversity), Kok-Seng Wong (VinUniversity)
CodeKnowledge DistillationAdversarial AttackImage
π― What it does: This paper proposes a Clean Label Physical Backdoor Attack (CLPBA), which implants a backdoor by applying tiny, invisible perturbations to a small number of target class samples, without modifying labels or injecting triggers.
CLER: Improving Multimodal Financial Reasoning by Cross-MLLM Error Reflection
Shuangyan Deng (University of Auckland), Jiamou Liu (University of Auckland)
CodeLarge Language ModelContrastive LearningMultimodalityFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the CLER framework, which enhances the accuracy of multimodal financial reasoning through the retrieval of cross-model errors and iterative reflection.
CliCARE: Grounding Large Language Models in Clinical Guidelines for Decision Support over Longitudinal Cancer Electronic Health Records
Dongchen Li (Northeastern University), Kun Yu (Northeastern University)
CodeGraph Neural NetworkTransformerLarge Language ModelBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
π― What it does: Proposes the CliCARE framework, which converts long-term cancer electronic health records (EHR) into a temporal knowledge graph (TKG), aligns it with clinical guideline knowledge graphs, and generates clinical summaries and treatment recommendations.
π― What it does: Propose CLIPDet3D, a multi-view 3D object detection framework based on audio-visual collaboration, specifically addressing the long-tailed distribution problem of rare categories.
CodeSuper ResolutionPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Propose the CLIPPan framework, using CLIP as a supervisor for unsupervised full-resolution pansharpening, guided by text descriptions to direct the fusion process;
CLM-Access: A Specialized Foundation Model for High-Dimensional Single-Cell ATAC-Seq Analysis
Ziqiang Liu (Hangzhou Institute of Medicine, Chinese Academy of Sciences), Xiaolin Li (Hangzhou Institute of Medicine, Chinese Academy of Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBiomedical Data
π― What it does: Designed and trained a Transformer-based foundation model called CLM-Access specifically for scATAC-seq, incorporating steps such as unified preprocessing, patch embedding, and masked reconstruction.
π― What it does: Propose CloudMamba, a Mamba-based point cloud analysis network that integrates sequence expansion and merging, chained bidirectional Mamba, and grouped selective state space model (GS6), achieving efficient long-range modeling.
CLUENet: Cluster Attention Makes Neural Networks Have Eyes
Xiangshuai Song (National University of Defense Technology), Chang Tang (Huazhong University of Science and Technology)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkImageBenchmark
π― What it does: Proposed CLUENet clustering attention network to address the shortcomings of traditional convolutional/attention models in modeling irregular spatial patterns and interpretability.
π― What it does: A self-learning graph regression (SGR) method is proposed, which uses the affinity graph itself to construct regularization for graph clustering, significantly improving clustering quality.
π― What it does: Proposed a UNet with Continuous Neural Memory ODE (CNM-UNet), achieving efficient and lightweight medical image segmentation by replacing the entire original UNet decoder with a single CNM-Block;
CO-Bench: Benchmarking Language Model Agents in Algorithm Search for Combinatorial Optimization
Weiwei Sun (Carnegie Mellon University), Yiming Yang (Carnegie Mellon University)
CodeOptimizationLarge Language ModelAgentic AIBenchmarkChain-of-Thought
π― What it does: This study proposes the CO-Bench benchmark, which includes 36 real-world combinatorial optimization (CO) problems, and systematically evaluates them on 15 LLMs and 9 agent frameworks, comparing their performance with human-designed classical solvers;
CodeRestorationSuper ResolutionVision Language ModelContrastive LearningImageMultimodality
π― What it does: Designed and implemented a language-guided high-resolution spectral image fusion framework, CO IF, capable of reconstructing high-quality HR-HSI by integrating low-resolution hyperspectral images with high-resolution multispectral images.
π― What it does: A CoCo-MILP framework is proposed for mixed integer linear programming (MILP) problems, which predicts high-quality initial solutions by contrastive learning on variables and constructing competitive mechanisms within constraints, thereby accelerating traditional solvers.
π― What it does: Achieving synthetic translation from structural MRI to Alzheimer's AΞ² PET scans via the ControlNet-conditional latent diffusion model CoCoLIT.