🎯 What it does: A framework-aware learning curve extrapolation model LC-GODE has been developed, which utilizes graph neural networks to encode network structures and embeds them into neural ODEs to predict the learning curves of different neural networks during the early training phase using variational inference.
Are Expressive Models Truly Necessary for Offline RL?
Guan Wang (Tsinghua University), Xianyuan Zhan (Shanghai AI Laboratory)
CodeReinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes a Recursive Step Planning (RSP) method that uses a shallow two-layer MLP to predict sub-goals and executes actions through a goal-conditioned policy, addressing the long-term error problem in offline reinforcement learning.
Are Key-Phrases All That Reviewers Care About? A Comprehensive Benchmarking of Reviewer Matchmaking Systems
Sourish Dasgupta (Dhirubhai Ambani Institute of Information and Communication Technology), Anil K. Roy (Dhirubhai Ambani Institute of Information and Communication Technology)
🎯 What it does: A comprehensive benchmark test of existing Reviewer Matchmaking (RM) systems is conducted, systematically comparing two types of models: Document Representation (DR) and Keyword Extraction (KPE). The shortcomings of traditional evaluation metrics (Precision@K, Kendall Loss) are pointed out, and the use of correlation coefficients (Pearson, Spearman, Kendall) is proposed as a reliable evaluation method.
Argumentative Large Language Models for Explainable and Contestable Claim Verification
Gabriel Freedman (Imperial College London), Francesca Toni (Imperial College London)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes an 'Argument LLM (ArgLLM)' that combines large language models (LLM) with a formal argumentation framework (QBAF), achieving interpretable and debatable decision outputs by constructing argument trees that support and refute.
ARNet: Self-Supervised FG-SBIR with Unified Sample Feature Alignment and Multi-Scale Token Recycling
Jianan Jiang (Hunan University), Di Wu (University of Warwick)
CodeRetrievalTransformerContrastive LearningImage
🎯 What it does: A self-supervised FG-SBIR framework ARNet has been developed, which achieves intra- and inter-sample feature alignment using a dual-weight shared encoder, and enhances fine-grained image retrieval performance by reusing discarded patch tokens in ViT through a multi-scale token recovery module.
ARTICLE: Annotator Reliability Through In-Context Learning
Sujan Dutta (Rochester Institute of Technology), Ashiqur R. KhudaBukhsh (Rochester Institute of Technology)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the ARTICLE framework, which utilizes context learning from large language models to evaluate the self-consistency of annotators, thereby identifying reliable annotators and modeling the perception of aggression among different political groups.
🎯 What it does: This paper proposes the AdaptFG model, which enhances feature diversity and avoids pseudo-label noise in sparse annotation object detection through adaptive feature generation, thereby improving detection performance.
🎯 What it does: This paper proposes an adaptive sampling single-stage 3D detector AS-Det, which can achieve high-precision detection on various point clouds (LiDAR and 4D radar).
ASER: Activation Smoothing and Error Reconstruction for Large Language Model Quantization
Weibo Zhao (Alibaba Cloud Computing), Yong Li (Alibaba Cloud Computing)
CodeTransformerLarge Language ModelText
🎯 What it does: This study investigates the low-rank characteristics of low-bit quantization errors in LLMs and proposes the ASER algorithm, which achieves efficient post-training quantization through activation smoothing and error reconstruction.
ASP-Driven Emergency Planning for Norm Violations in Reinforcement Learning
Sebastian Adam (Vienna University of Technology), Thomas Eiter (Vienna University of Technology)
CodeReinforcement Learning
🎯 What it does: A framework based on Answer Set Programming (ASP) is proposed, which corrects the behavior of reinforcement learning (RL) agents by generating emergency action plans (policy fixes) when ethical violations are detected during execution, thus achieving compliance without retraining.
🎯 What it does: An association pattern-aware information enhancement plugin, APMP, is proposed, which improves the atomic representations of existing molecular base models by sampling, filtering, and utilizing high-confidence paths for message passing in molecular graphs.
Asymmetric Learning for Spectral Graph Neural Networks
Fangbing Liu (Australian National University), Qing Wang (Australian National University)
CodeOptimizationGraph Neural NetworkGraph
🎯 What it does: An in-depth analysis of the optimization process of spectral graph convolutional networks is conducted, and a method to improve the optimization landscape through gradient asymmetric preprocessing is proposed.
Asymmetric Visual Semantic Embedding Framework for Efficient Vision-Language Alignment
Yang Liu (Sichuan University), Jiancheng Lv (Sichuan University)
CodeRetrievalTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the Heterogeneous Visual-Semantic Embedding (AVSE) framework, which obtains multi-view image features through radial bias sampling and achieves efficient text-image matching using asymmetric embedding optimal matching.
Asynchronous Federated Clustering with Unknown Number of Clusters
Yunfan Zhang (Guangdong University of Technology), Yiu-ming Cheung (Hong Kong Baptist University)
CodeFederated LearningTabular
🎯 What it does: This paper proposes an Asynchronous Federated Clustering method (AFCL) that can automatically learn the global cluster distribution and complete cluster number estimation under conditions of unsynchronized communication, highly heterogeneous client data, and unknown number of clusters.
Attack-in-the-Chain: Bootstrapping Large Language Models for Attacks Against Black-Box Neural Ranking Models
Yu-An Liu (Chinese Academy of Sciences), Xueqi Cheng (Chinese Academy of Sciences)
CodeRetrievalAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A chain reasoning framework based on large language models, Attack-in-the-Chain, is proposed for attacking neural ranking models under black-box conditions.
🎯 What it does: AENet is proposed, which enhances visual representation in zero-shot learning through concept alignment attention and semantic enhancement prompts.
Attention-Driven GUI Grounding: Leveraging Pretrained Multimodal Large Language Models Without Fine-Tuning
Hai-Ming Xu (Australian Institute for Machine Learning), Lingqiao Liu (University of Wollongong)
CodeRecognitionObject DetectionTransformerLarge Language ModelPrompt EngineeringImageTextMultimodality
🎯 What it does: A no-fine-tuning Attention-driven GUI Grounding (TAG) method is proposed, which achieves precise localization of GUI elements by leveraging the attention mechanism of a pre-trained multimodal large language model.
Attribute Inference Attacks for Federated Regression Tasks
Francesco Diana (Universite Cote d'Azur), Eoin Thomas (Amadeus)
CodeFederated LearningSafty and PrivacyTabular
🎯 What it does: This paper proposes an Attribute Inference Attack (AIA) method for regression tasks within the framework of federated learning, exploring the feasibility of model-based attacks in this scenario.
Attribution Analysis Meets Model Editing: Advancing Knowledge Correction in Vision Language Models with VisEdit
Qizhou Chen (East China Normal University), Tingting Liu (Exacity Inc.)
CodeExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This study investigates the impact of visual representations on text prediction in Vision-Language Large Models (VLLM) and proposes a new model editor, VisEdit, to correct knowledge errors in VLLM without retraining the entire model.
🎯 What it does: This paper constructs a hallucination attribution framework based on internal states and proposes Differential Penalty Decoding (DPD) to reduce hallucinated outputs from LLMs.
CodeClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringContrastive LearningMultimodalityAudio
🎯 What it does: This paper proposes the Audio Entailment task to evaluate the reasoning capabilities of audio-language models (ALM); it constructs two datasets, ACE and CLE, and benchmarks existing contrastive learning and next-word prediction ALMs under zero-shot and linear probe settings; it introduces the intermediate step 'caption-before-reason', which significantly enhances reasoning performance.
AudioGenX: Explainability on Text-to-Audio Generative Models
Hyunju Kang (Sungkyunkwan University), Hogun Park (NCSOFT)
CodeGenerationData SynthesisExplainability and InterpretabilityTransformerTextAudio
🎯 What it does: AudioGenX is proposed - an explainable method based on factual and counterfactual reasoning to explain the impact of text inputs on audio generated by text-to-audio generation models.
Augmenting Math Word Problems via Iterative Question Composing
Haoxiong Liu (Institute for Interdisciplinary Information Sciences), Andrew C Yao (Institute for Interdisciplinary Information Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper enhances the mathematical reasoning ability of basic LLMs by constructing a new mathematical problem dataset called MMIQC and proposing the Iterative Question Composing (IQC) method.
🎯 What it does: A pluggable enhancement plugin BASRec is proposed in sequential recommendation to generate new training samples while maintaining semantic relevance and introducing diversity.
🎯 What it does: A novel method for enhancing adversarial robustness, AUTE, is designed, which combines Peer-Alignment and Self-Unlearning techniques to train ensemble models.
🎯 What it does: This paper proposes an autoregressive diffusion model called ARDHOI, which achieves text-driven 3D human-object interaction sequence generation using continuous HOI labels.
Auto-Regressive Moving Diffusion Models for Time Series Forecasting
Jiaxin Gao (Shanghai Jiao Tong University), Yuntian Chen (Ningbo Institute of Digital Twin, Eastern Institute of Technology)
CodeDiffusion modelTime SeriesFinance Related
🎯 What it does: The Auto-Regressive Moving Diffusion (ARMD) model is proposed, treating the future sequence of a time series as the initial state of a diffusion process and the historical sequence as the final state. It utilizes a sliding mechanism to generate continuous intermediate states, achieving unconditional continuous diffusion time series forecasting.
AutoFEA: Enhancing AI Copilot by Integrating Finite Element Analysis Using Large Language Models with Graph Neural Networks
Shifu Hou (University of Notre Dame), Yanfang Ye (University of Notre Dame)
CodeAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelGraphRetrieval-Augmented Generation
🎯 What it does: Integrate large language models with finite element analysis (FEA) to build the AutoFEA system, achieving automatic generation of FEA input files and executing simulations, thereby reducing AI hallucinations;
Automated Creation of Reusable and Diverse Toolsets for Enhancing LLM Reasoning
Zhiyuan Ma (University of Science and Technology of China), Xin Li (Tianjin University)
CodeLarge Language ModelAgentic AITextTabular
🎯 What it does: This paper proposes a two-stage knowledge-driven tool creation and evolution framework (KTCE) that can automatically generate reusable and diverse toolsets to enhance the reasoning capabilities of large language models.
Automatic Selection of Macro-Events for Heuristic-Search Temporal Planning
Alessandro La Farciola (Fondazione Bruno Kessler), Andrea Micheli (Fondazione Bruno Kessler)
CodeOptimizationReinforcement LearningSequential
🎯 What it does: This study investigates the concept of 'macro-event' used in heuristic time planning and proposes a statistical ranking method for automatically selecting useful macro-events from a verified set of plans.
AutoMMLab: Automatically Generating Deployable Models from Language Instructions for Computer Vision Tasks
Zekang Yang (SenseTime Research and Tetras.AI), Wentao Liu (SenseTime Research and Tetras.AI)
CodeObject DetectionSegmentationHyperparameter SearchTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageTextBenchmark
🎯 What it does: The AutoMMLab platform has been built to realize a complete end-to-end automated process from natural language requests to deployable computer vision models.
🎯 What it does: A hierarchical reinforcement learning and planning framework called CHiRP has been developed, which can automatically invent, represent, and utilize symbolic options in a continual learning environment.
AutoSGNN: Automatic Propagation Mechanism Discovery for Spectral Graph Neural Networks
Shibing Mo (Xidian University), Jing Liu (Xidian University)
CodeGraph Neural NetworkLarge Language ModelPrompt EngineeringGraph
🎯 What it does: This paper proposes AutoSGNN, a method for automatically generating spectral graph neural network propagation mechanisms using LLM and evolutionary strategies.
🎯 What it does: A backdoor attack model for no-reference image quality assessment (NR-IQA) is proposed, which injects noise in the DCT domain using a scalable trigger to achieve arbitrary control over the model output score.
Backdoor Token Unlearning: Exposing and Defending Backdoors in Pretrained Language Models
Peihai Jiang (Xidian University), Jing Ma (Xidian University)
CodeAnomaly DetectionTransformerLarge Language ModelText
🎯 What it does: A method for actively detecting and eliminating backdoor triggers in pre-trained language models during the training phase is proposed—Backdoor Token Unlearning (BTU);
Balanced Adaptive Subspace Collaboration for Mixed Pareto-Lexicographic Multi-Objective Problems with Priority Levels
Wenjing Hong (Shenzhen University)
CodeOptimization
🎯 What it does: A new Balanced Adaptive Subspace Collaboration (BASC) algorithm is proposed to address mixed Pareto-Lexicographic multi-objective optimization problems (PL-MPL-MOPs) with multiple priority levels. It generates new solutions by sampling within subspaces and dynamically balances exploration and exploitation across different priority layers during the search process.
Batch Selection for Multi-Label Classification Guided by Uncertainty and Dynamic Label Correlations
Ao Zhou (Chongqing University of Posts and Telecommunications), Grigorios Tsoumakas (Aristotle University of Thessaloniki)
CodeClassificationImageVideoText
🎯 What it does: A multi-label batch selection method based on uncertainty is proposed, which evaluates label uncertainty using a combination of sliding window differences and entropy, and further measures sample uncertainty through dynamic label associations, thereby selecting more informative samples during training.
Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation
HyunGi Kim (Seoul National University), Sungroh Yoon (Seoul National University)
CodeTransformerTime SeriesBenchmark
🎯 What it does: This paper proposes a test-time adaptive framework for non-stationary time series forecasting, TAFAS, which can actively utilize partially observed true values to adjust the pre-trained model during the inference phase.
BearLLM: A Prior Knowledge-Enhanced Bearing Health Management Framework with Unified Vibration Signal Representation
Haotian Peng (Shenyang Institute of Automation, Chinese Academy of Sciences), Wei Wang (Shenyang Institute of Automation, Chinese Academy of Sciences)
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningMultimodalityTime Series
🎯 What it does: This paper presents BearLLM, an integrated multi-task bearing health management framework that achieves anomaly detection, fault diagnosis, maintenance recommendations, and risk analysis through a unified vibration signal representation and large language models.
CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerReinforcement LearningImage
🎯 What it does: A method named BEE is proposed, which is based on baseline exploration-exploitation for path integral interpretation. It automatically learns and samples baseline distributions suitable for specific evaluation metrics, thereby generating explanation maps that better align with the metrics.
🎯 What it does: This paper proposes a visual attribution analysis method to diagnose the defects of content features in multimodal recommendation, and designs a general behavior-driven feature adapter (BeFA) based on behavioral information to reconstruct content features.
🎯 What it does: The BiGNAS framework is proposed, which jointly optimizes the GNN architecture and the importance of source domain behavior in cross-domain recommendation.
Benchmarking and Understanding Compositional Relational Reasoning of LLMs
Ruikang Ni (Beijing University of Posts and Telecommunications), Hongliang Liang (Beijing University of Posts and Telecommunications)
CodeTransformerLarge Language ModelTextBenchmark
🎯 What it does: A new synthetic benchmark called Generalized Associative Recall (GAR) was designed and evaluated, systematically analyzing the performance of LLMs in Combinatorial Relation Reasoning (CRR), and discovering core circuits and true/false heads through mechanism explanations.
Better Understandings and Configurations in MaxSAT Stochastic Local Search Solvers via Anytime Performance Analysis
Furong Ye (Chinese Academy of Sciences), Shaowei Cai (Chinese Academy of Sciences)
CodeOptimizationHyperparameter SearchTabular
🎯 What it does: This study is the first to use the empirical cumulative distribution function (ECDF) to evaluate the 'anytime' performance of MaxSAT's stochastic local search (SLS) solvers and to improve hyperparameter configurations based on the evaluation results.
Beyond Accuracy: On the Effects of Fine-Tuning Towards Vision-Language Model’s Prediction Rationality
Qitong Wang (University of Delaware), Xi Peng (University of Delaware)
CodeClassificationExplainability and InterpretabilityTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: This study investigates the impact of mainstream VLM fine-tuning methods on prediction rationality (based on the credibility of evidence) and proposes two new metrics: prediction credibility and inference reliability.
Beyond Human Data: Aligning Multimodal Large Language Models by Iterative Self-Evolution
Wentao Tan (South China University of Technology), Changxing Ding (South China University of Technology)
CodeRecognitionGenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: A multimodal self-evolution framework (SENA) is constructed using unlabeled images through three main mechanisms: self-questioning, self-enhanced answering, and image content alignment, achieving continuous improvement of LLM in visual tasks without the need for manual or external model annotations.
Beyond IID: Optimizing Instruction Finetuning from the Perspective of Instruction Interaction and Dependency
Hanyu Zhao (Beijing Academy of Artificial Intelligence), Tengfei Pan (Beijing Academy of Artificial Intelligence)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a systematic analysis and optimization of the instruction set of large language models (SFT) from the perspective of instruction interaction and dependency.
🎯 What it does: Proposed a Federated Participation Framework (FPF) and Local Multi-Federation (LMF) framework to address the issues of low information sharing and participation in traditional mandatory federations; simultaneously balance the three main objectives of self-interest, utilitarianism, and equality in mixed-motivation games; experiments demonstrate that FPF/LMF outperforms various baselines and reduces the self-interest of non-participants.
Beyond Prompt Engineering: A Reinforced Token-Level Input Refinement for Large Language Models
Guang Huang (University of Macau), Pengyang Wang (University of Macau)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A token-level input optimization framework based on reinforcement learning, RTLIR, is proposed to automatically remove irrelevant information from LLM inputs.
Beyond Skip Connection: Pooling and Unpooling Design for Elimination Singularities
Chengkun Sun (University of Florida), Jie Xu (University of Florida)
CodeClassificationSegmentationConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: Proposed and implemented the Pool Skip module, which combines max pooling, max unpooling, 3×3 convolution, and skip connections to alleviate the learning degradation caused by the elimination of singularities in deep CNNs.
🎯 What it does: A module based on Bernoulli-Gaussian Decision Blocks (BGDB) is proposed, which simulates the probability distribution of multiple Bernoulli trials using an improved Denoising Diffusion Probabilistic Model (IDDPM) in a single training session, thereby enhancing the stability and performance of the discriminative classifier.
BGHR: Bridging the Gap Between HBox-Supervised and RBox-Supervised Oriented Object Detection via Adaptive Fine-Grained Sample Mining
Chenlin Fu (Shenzhen University), Yingying Zhu (Shenzhen University)
CodeObject DetectionImage
🎯 What it does: A rotation target detection framework BGHR based on horizontal box annotations has been developed, utilizing adaptive fine-grained sample mining and self-supervised branch loss to enhance detection performance.
🎯 What it does: This paper proposes the BiMSGC framework, which first selects the optimal medium-scale subgraph through information bottleneck, and then performs bidirectional (from large to small and from small to large) graph dataset distillation based on this, ultimately generating multi-scale high-quality synthetic graphs.
BiDeV: Bilateral Defusing Verification for Complex Claim Fact-Checking
Yuxuan Liu (Renmin University of China), Rui Yan (Renmin University of China)
CodeTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented the BiDeV framework, achieving de-ambiguation and de-redundancy of complex claims through multi-role collaboration of LLMs, thereby enabling more accurate fact-checking.
🎯 What it does: This paper proposes two methods: Bidirectional Logits Tree (BiLT) and Adaptive Intra-Granularity Difference Learning (AIGDL) to address the feature learning bias caused by the competition between coarse and fine levels in fine-grained classification.
CodeRecommendation SystemTransformerTabularAgriculture Related
🎯 What it does: Proposes a green food recommendation task and designs the GRAPE model to achieve dual optimization of personalization and sustainability.
BLADE: Enhancing Black-Box Large Language Models with Small Domain-Specific Models
Haitao Li (Tsinghua University), Yiqun Liu (Tsinghua University)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By constructing a small domain-specific language model and working in conjunction with a general large language model, the question-and-answer performance in vertical fields (such as law and medicine) is enhanced.
Blend the Separated: Mixture of Synergistic Experts for Data-Scarcity Drug-Target Interaction Prediction
Xinlong Zhai (Beijing University of Posts and Telecommunications), Chuan Shi (Beijing University of Posts and Telecommunications)
CodeDrug DiscoveryGraph Neural NetworkMixture of ExpertsBiomedical Data
🎯 What it does: A hybrid expert model named MoseDTI is proposed to simultaneously handle structural information and knowledge graph information in drug-target interaction prediction, alleviating issues of missing input data and sparse annotations through a self-supervised pseudo-labeling approach.
🎯 What it does: We propose BloomScene, a lightweight and structured 3D Gaussian splatting framework that can automatically generate high-quality 3D scenes from cross-modal inputs such as text or images.
🎯 What it does: This paper proposes a BLS-GAN framework that can separate the upper and lower bone layers from a single hand X-ray image, eliminating bone overlap.
🎯 What it does: A new high-dimensional Bayesian optimization method called BOIDS is proposed, which enhances the search efficiency on high-dimensional expensive black-box functions by utilizing direction lines guided by dominant solutions and subspace embedding.
🎯 What it does: By utilizing a small number of coarsely labeled anomalous samples, an energy discriminator drives an autoencoder to align the reconstruction results with normal samples in the feature space, thereby improving the accuracy of fine-grained visual anomaly detection and localization.
Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inference
Zhihang Lin (Xiamen University), Rongrong Ji (Xiamen University)
CodeRecognitionComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodality
🎯 What it does: The Visual Tokens Withdrawal (VTW) module is proposed, which can remove visual tokens in deep networks during inference of multimodal large language models (MLLM), significantly reducing computational overhead.
🎯 What it does: The MI-DELIGHT model is proposed, which significantly improves short text classification performance through multi-source information exploration (words, POS, entity graphs) and dual-layer contrastive learning (instance-level ICL and cluster-level CCL), employing a hierarchical task architecture.
Boosting Test Performance with Importance Sampling--a Subpopulation Perspective
Hongyu Shen (University of Illinois), Zhizhen Zhao (University of Illinois)
CodeClassificationDomain AdaptationImageText
🎯 What it does: A subpopulation bias analysis framework (DBA) based on importance sampling is proposed, and within this framework, a subpopulation bias correction method (DBCM) for a single estimator is designed to reweight the training set to improve testing performance under subpopulation imbalance.
🎯 What it does: This paper presents FreeGS, an unsupervised 3D Gaussian Splatting framework that utilizes the IDentity-coupled Semantic Field (IDSF) to simultaneously encode semantic features and cross-view instance indexing on Gaussians, achieving view-consistent 3D scene understanding.
Jacob Adamczyk (University of Massachusetts Boston), Rahul V. Kulkarni (Texas Tech University)
CodeReinforcement LearningTabular
🎯 What it does: This paper proposes a bootstrapping reward shaping method (BSRS) that dynamically shapes rewards by using the agent's current estimated state value function as a potential function.
Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach
Hang Gao (National Key Laboratory of Space Integrated Information System Institute of Software Chinese Academy of Sciences), Huaping Liu (Tsinghua University)
CodeRepresentation LearningGraph Neural NetworkLarge Language ModelGraph
🎯 What it does: This paper proposes a general heterogeneous graph representation learning framework GHGRL, which utilizes large language models to automatically identify the types and formats of nodes and edges, and combines it with a GNN with adaptive parameters for graph information aggregation, without the need for pre-provided type labels or unified feature formats.
BotSim: LLM-Powered Malicious Social Botnet Simulation
Boyu Qiao (Institute of Information Engineering, Chinese Academy of Sciences), Songlin Hu (Institute of Information Engineering, Chinese Academy of Sciences)
CodeAnomaly DetectionGraph Neural NetworkLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: This paper designs a scalable LLM-driven malicious social bot simulation framework called BotSim, and based on this framework, generates the BotSim-24 dataset, which includes real human users and LLM-driven bot accounts.
Boundary Decomposition for Finding Nadir Objective Vector in Multi-Objective Discrete Optimization
Ruihao Zheng (Southern University of Science and Technology), Zhenkun Wang (Southern University of Science and Technology)
CodeOptimization
🎯 What it does: A boundary decomposition-based bi-level optimization algorithm BDNC is proposed to accurately calculate the nadir vector of multi-objective discrete optimization problems (MODOP).
🎯 What it does: This paper studies the introduction of bounded rationality and receding horizon methods in Mean Field Games (MFG), proposing Quantal Response Equilibria (QRE) and its variant under limited foresight, and designing general fixed point iteration and fictitious game algorithms to learn these equilibria.
BrainMAP: Learning Multiple Activation Pathways in Brain Networks
Song Wang (University of Virginia), Jundong Li (University of Virginia)
CodeOptimizationExplainability and InterpretabilityGraph Neural NetworkMixture of ExpertsContrastive LearningGraphBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes the BrainMAP framework for learning and interpreting multiple activation pathways in fMRI functional connectivity maps, enhancing brain network prediction and interpretability.
Breaking Barriers in Physical-World Adversarial Examples: Improving Robustness and Transferability via Robust Feature
Yichen Wang (Huazhong University of Science and Technology), Minghui Li (Huazhong University of Science and Technology)
CodeAdversarial AttackVision Language ModelImage
🎯 What it does: A robust feature coverage-based physical world adversarial attack method (RFCoA) is proposed, which generates adversarial samples through robust feature injection and semantic pattern minimization.
🎯 What it does: Proposes the Bridge Diffusion Model (BDM), which generates images from Chinese text through a backend-branch architecture, while also being compatible with English community plugins.
Bridging Molecular Graphs and Large Language Models
Runze Wang (Dalian University of Technology), Yanming Shen (Dalian University of Technology)
CodeDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextMultimodalityGraph
🎯 What it does: The molecular graph structure is mapped to a special graph token through a graph encoder, and this token is aligned to the LLM's vocabulary space using a cross-attention mechanism, allowing the LLM to understand and reason about the molecular graph while maintaining the original weights; subsequently, the IUPAC name and task description are added to the prompt, and the task head is used to output the prediction results.
Bridging Sequence-Structure Alignment in RNA Foundation Models
Heng Yang (University of Exeter), Ke Li (University of Exeter)
CodeTransformerLarge Language ModelSequentialBiomedical Data
🎯 What it does: The OmniGenome RNA base model is proposed, achieving bidirectional alignment of sequences and secondary structures, and performing excellently in RNA design and structure prediction tasks.
🎯 What it does: A Hyperdimensional Computing (HDC) encoding algorithm called NysHD based on the Nyström method is proposed, which can map any positive definite kernel function to low-precision high-dimensional vectors, achieving efficient HDC learning.
🎯 What it does: This paper proposes a test-time adaptive method for multimodal emotion analysis (CASP), which addresses the issue of model performance degradation under target domain distribution shift by randomly dropping multimodal inputs and ensuring consistency through contrastive learning, while generating stable pseudo-labels for self-supervised training.
Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models
Zheng Hu (University of Electronic Science and Technology of China), Fuji Ren (University of Electronic Science and Technology of China)
CodeRecommendation SystemGraph Neural NetworkLarge Language ModelAuto EncoderContrastive LearningGraph
🎯 What it does: This paper utilizes large language models to infer user interests from historical behavior, constructing a Collaborative Interest Knowledge Graph (CIKG) for recommendations.
Bridging Traffic State and Trajectory for Dynamic Road Network and Trajectory Representation Learning
Chengkai Han (Beihang University), Junjie Wu (Beihang University)
CodeAutonomous DrivingRepresentation LearningGraph Neural NetworkTransformerContrastive LearningGraphTime Series
🎯 What it does: By jointly modeling traffic state data and trajectory data, the TRACK framework is proposed to learn low-dimensional representations of dynamic road networks and trajectories.
🎯 What it does: This paper proposes a Transformer-based Layered Graph Aggregation Network (TGCNet) that utilizes dynamic directed graphs to learn communication strategies among multiple agents. During centralized training, it approximates the global state through a graph aggregation network, and during decentralized execution, it uses a Transformer decoder to complete feature fusion.
🎯 What it does: This paper presents BSAFusion, a single-stage bidirectional step feature alignment network for the registration and fusion of unaligned multimodal medical images.
CodeRecognitionKnowledge DistillationTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: A multi-modal spatio-temporal expert framework STDD based on CLIP is proposed, specifically addressing the problem of collaborative understanding of visual and textual spatio-temporal dynamics in zero-shot action recognition (ZSAR).
🎯 What it does: C2F-TP is proposed, using a coarse-to-fine denoising framework for vehicle trajectory prediction. It first learns the multimodal trajectory distribution through a spatial-temporal interaction module, and then gradually denoises to generate more accurate trajectories through a conditional denoising module.
C3oT: Generating Shorter Chain-of-Thought Without Compromising Effectiveness
Yu Kang (Beike Inc), Wei Zou (Beike Inc)
CodeCompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: The C3oT framework is proposed, which significantly shortens intermediate reasoning steps through Conditional Compressed Chain-of-Thought (CoT) while maintaining the accuracy of the final answer.
🎯 What it does: This paper proposes an unsupervised video generation model called CAGE, which can achieve scene composition and object animation by placing visual tokens in both spatial and temporal dimensions.
CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuning
Peiyuan Liu (Tsinghua University), Shu-Tao Xia (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningTime Series
🎯 What it does: By constructing the CALF framework and using a cross-modal fine-tuning method, the text representation of the LLM is aligned with the time series input in terms of distribution, features, and output space, achieving long-term and short-term multivariate forecasting.
🎯 What it does: In the multi-label learning task with a candidate label set that includes noisy labels, a calibration-based curriculum learning method PML-CD is proposed. It utilizes a transferable calibrator learned from the confidence histogram during the training process to adaptively assign weights to samples and enhances robustness through prototype alignment regularization.
Calibrating Large Language Models with Sample Consistency
Qing Lyu (University of Pennsylvania), Chris Callison-Burch (University of Pennsylvania)
CodeGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper estimates the confidence of large language models (LLMs) by sampling multiple generations and calculating their consistency (using three metrics: agreement, entropy, and FSD), thereby achieving post-processing calibration of the model's prediction results.
🎯 What it does: Utilizing instance conditional generative models (such as Stable Diffusion and ICGAN) to generate semantically consistent image enhancements, enriching the views of self-supervised learning and improving representation quality.
Can Large Language Models Derive High-Level Cognition from Low-Level and Fragmented Foundational Information?
Yang Liu (Hunan University), Kai Lu (National University of Defense Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A dataset specifically designed to assess the high-level cognitive (HLC) abilities of large language models (MatchIntel) has been constructed, and three evaluation tasks (MCQ, SCQ, TFQ) have been proposed. The performance of existing LLMs on HLC tasks has been verified through fine-tuning and comparative experiments.
Can LLMs Obfuscate Code? A Systematic Analysis of Large Language Models into Assembly Code Obfuscation
Seyedreza Mohseni (University of Maryland), Manas Gaur (University of Maryland)
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper investigates whether large language models can generate obfuscated assembly code and proposes the METAMORPHASM benchmark and the MAD dataset for systematic evaluation.
Joseph Rance (University of Cambridge), Filip Svoboda (University of Cambridge)
CodeFederated LearningImageTabular
🎯 What it does: This paper proposes a new attack method that deliberately introduces unfairness through malicious clients during the federated learning process, and systematically evaluates the impact of existing robust aggregation methods on fairness.
🎯 What it does: A knowledge distillation strategy based on bias elimination and correction is proposed, enabling the student model to surpass the teacher model.