🎯 What it does: Propose an evolutionary generalized zero-shot learning (EGZSL) framework that enables the model to continuously adapt and improve performance on unlabeled test streams.
Exactly Solving Minimum Dominating Set and Its Generalization
Ziliang Xiong (University of Electronic Science and Technology of China), Mingyu Xiao (University of Electronic Science and Technology of China)
CodeOptimizationGraphBenchmark
🎯 What it does: Proposed two exact algorithms for solving the minimum dominating set problem based on branch-and-bound, using combinatorial lower bounds and LP relaxation lower bounds respectively.
Expected Work Search: Combining Win Rate and Proof Size Estimation
Owen Randall (University of Alberta), Ryan Hayward (University of Alberta)
CodeOptimizationReinforcement Learning
🎯 What it does: Proposed a novel search algorithm called Expected Work Search (EWS), which predicts and minimizes the computational effort required to solve positions by simultaneously estimating win rates and proof sizes.
🎯 What it does: Propose the ELIMIPL algorithm, which improves the disambiguation performance in multi-instance partial label learning by leveraging conjugate information between candidate and non-candidate labels, as well as the sparsity of true labels.
CodeOptimizationRepresentation LearningImageTabularBiomedical Data
🎯 What it does: An auxiliary multi-label learning (MLL) process is proposed in label distribution learning (LDL), leveraging low-rank label correlations on the auxiliary process to enhance LDL's predictive performance; label distributions are converted into multi-labels via threshold or top-k methods, followed by jointly optimizing the mapping between label distributions and multi-labels, forming two methods: TLRLDL and TKLRLDL.
🎯 What it does: Proposed a multi-scale graph neural network (MSGNN) that utilizes internal multi-scale self-similarity and external example similarity to aggregate neighboring image patches via graph neural networks for single-image deraining.
Exploring Cross-Domain Few-Shot Classification via Frequency-Aware Prompting
Tiange Zhang (Ocean University of China), Junyu Dong (Ocean University of China)
CodeClassificationDomain AdaptationMeta LearningConvolutional Neural NetworkPrompt EngineeringImageBiomedical Data
🎯 What it does: Studied cross-domain few-shot classification, proposing Frequency-Aware Prompting and mutual attention mechanisms to enhance the cross-domain generalization ability of meta-learning models.
🎯 What it does: Designed and implemented a phased weight sharing (SWS) framework to learn a compact learngene, and extended it to multi-scale Transformers to adapt to model deployment under different resource constraints.
Exploring Urban Semantics: A Multimodal Model for POI Semantic Annotation with Street View Images and Place Names
Dabin Zhang (Shandong University), Kai Zhao (Georgia State University)
CodeClassificationConvolutional Neural NetworkTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a multimodal model M3PA that fuses street view images with POI names and their neighbors' information to achieve semantic annotation of POIs.
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the FACTCHD benchmark to evaluate LLM's fact conflict hallucination detection, and proposed TRUTH-TRIANGULATOR for multi-source evidence triangulation.
Renqiang Luo (Dalian University of Technology), Feng Xia (RMIT University)
CodeClassificationSafty and PrivacyTransformerGraph
🎯 What it does: Proposed a fairness-adaptive graph transformer called FairGT to address the bias issue in traditional Graph Transformers on sensitive features;
🎯 What it does: Propose the FastKGE framework, which utilizes an incremental low-rank adapter (IncLoRA) to enable efficient learning and storage for continuously growing knowledge graphs.
🎯 What it does: This paper proposes a one-stage, end-to-end unsupervised domain adaptation person search framework named FOUS, which simultaneously performs detection and identity recognition in unlabeled target domains.
FBLG: A Local Graph Based Approach for Handling Dual Skewed Non-IID Data in Federated Learning
Yi Xu (Anhui University), Xiao Liu (Deakin University)
CodeFederated LearningGraph Neural NetworkImage
🎯 What it does: Propose a federated learning algorithm called FBLG based on a local graph, specifically designed to handle non-IID data with dual skewness in label distribution and sample size.
🎯 What it does: Proposed a federated learning algorithm called FNR-FL based on feature norm regularization, aiming to address model update bias caused by non-i.i.d. data.
Federated Adaptation for Foundation Model-based Recommendations
Chunxu Zhang (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)
CodeRecommendation SystemFederated LearningSafty and PrivacyKnowledge DistillationTransformerTabular
🎯 What it does: Proposed a federated recommendation framework called FedPA based on foundational models, which uses pre-trained large-scale models as a shared knowledge base. It achieves user personalization by learning lightweight low-rank adapters only at the client side, and fuses general knowledge with personalized features through an adaptive gating mechanism.
FedGCS: A Generative Framework for Efficient Client Selection in Federated Learning via Gradient-based Optimization
Zhiyuan Ning (Chinese Academy of Sciences), Yuanchun Zhou (Chinese Academy of Sciences)
CodeOptimizationFederated LearningRecurrent Neural NetworkLarge Language ModelImageText
🎯 What it does: Propose the FedGCS framework, reformulating the client selection task in federated learning as a generation task, generating the optimal subset of clients through gradient optimization in a continuous representation space.
🎯 What it does: Propose the FedPFT method, which performs proxy fine-tuning on a base model through federated learning without sharing server models or client data.
FedSSA: Semantic Similarity-based Aggregation for Efficient Model-Heterogeneous Personalized Federated Learning
Liping Yi (Nankai University), Xiaoxiao Li (University Of British Columbia)
CodeClassificationFederated LearningImage
🎯 What it does: Propose a FedSSA framework for model heterogeneity personalized federated learning, which can achieve local and global knowledge transfer through semantic similarity aggregation and adaptive parameter stabilization without sharing a public dataset.
🎯 What it does: Propose FlagVNE, a flexible and generalizable Virtual Network Embedding (VNE) framework that leverages reinforcement learning, graph neural networks, and meta-learning.
🎯 What it does: Designed and implemented a scene text recognition model called Character Features Enriched (CFE), specifically addressing Large Intra-Class Variance (LICV) and Small Inter-Class Variance (SICV) issues to enhance the discriminability of character features.
Fostering Collective Action in Complex Societies Using Community-Based Agents
Jonathan Skaggs (Brigham Young University), Jacob W. Crandall (Brigham Young University)
CodeGraph
🎯 What it does: This paper proposes Junior High Game (JHG) as a testbed for simulating complex social collective actions, and designs a community-based agent CAB that learns to form and maintain communities to achieve collective goals.
FreqFormer: Frequency-aware Transformer for Lightweight Image Super-resolution
Tao Dai (Shenzhen University), Zexuan Zhu (Shenzhen University)
CodeSuper ResolutionTransformerImage
🎯 What it does: Introduces FreqFormer, a lightweight Transformer that integrates spatial, frequency, and channel information to enhance image super-resolution performance.
🎯 What it does: This paper addresses image quality imbalance caused by equipment differences in medical images, proposing the FedISM method under a federated learning framework to achieve unified sharpness across clients, thereby improving the model's generalization performance on low-quality images while maintaining performance on high-quality images.
🎯 What it does: Propose a generative unified adversarial attack framework called GenSeg, designed to attack semantic, instance, and panoptic segmentation models.
Gradformer: Graph Transformer with Exponential Decay
Chuang Liu (Wuhan University), Wenbin Hu (Wuhan University)
CodeRepresentation LearningDrug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerGraphBiomedical Data
🎯 What it does: Designed a novel Graph Transformer called Gradformer, which incorporates graph structural priors naturally during information aggregation by multiplying an exponential decay mask on the self-attention matrix, using graph structure distances (short paths) to decay attention weights.
CodeTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoMultimodalityBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: Propose and release the GRASP benchmark, utilizing Unity-generated videos to evaluate the ability of multimodal large language models (LLMs) in language localization and intuitive physics understanding.
🎯 What it does: Proposed Group-Aware Coordination Graph (GACG), a dynamic coordination graph that simultaneously considers individual pairwise relationships and group-level dependencies, for information exchange and decision-making in multi-agent reinforcement learning;
🎯 What it does: Propose a learning framework based on ranking, replacing traditional heuristics with pairwise ranking to guide GBFS, and integrating it with DirectRanker;
Hard-Thresholding Meets Evolution Strategies in Reinforcement Learning
Chengqian Gao (Mohamed bin Zayed University of Artificial Intelligence), Zhiqiang Xu (Mohamed bin Zayed University of Artificial Intelligence)
CodeReinforcement LearningImage
🎯 What it does: Proposed an algorithm called NESHT that combines hard thresholding (Hard-Thresholding) with natural evolution strategies (NES) to remove task-irrelevant features and achieve sparse policies in reinforcement learning.
🎯 What it does: Propose HeterGCL, an unsupervised graph contrastive learning framework for heterogeneous graphs, integrating structural and semantic information through adaptive neighbor aggregation and multi-layer contrastive loss for representation learning.
🎯 What it does: Construct a gene-microbe-disease heterogeneous graph, extract subgraphs using six causal meta-paths, and design a causal semantic shared message passing network combined with subgraph attention fusion to predict triplet associations.
🎯 What it does: This paper proposes a visual reinforcement learning framework called MIIR based on information theory, which improves the generalization performance of visual RL by learning domain-invariant representations.
HVOFusion: Incremental Mesh Reconstruction Using Hybrid Voxel Octree
Shaofan Liu (Zhejiang University), Jianke Zhu (Zhejiang University)
CodeOptimizationSimultaneous Localization and MappingPoint CloudMesh
🎯 What it does: Proposes HVOFusion, an incremental mesh reconstruction method utilizing a hybrid voxel-octree structure, capable of online generating and directly optimizing explicit triangular meshes, combining sparse storage with high surface quality.
Hybrid Frequency Modulation Network for Image Restoration
Yuning Cui (Technical University of Munich), Alois Knoll (Technical University of Munich)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: Proposes the CSNet network based on channel and spatial dual-frequency modulation for image restoration tasks such as dehazing, defocusing blur removal, and snow removal.
HypBO: Accelerating Black-Box Scientific Experiments Using Experts’ Hypotheses
Abdoulatif Cissé (University of Liverpool), Andrew I. Cooper (University of Liverpool)
CodeOptimizationHyperparameter Search
🎯 What it does: Proposed a dual-layer Bayesian optimization method called HypBO based on expert hypotheses, which utilizes parameter interval constraints provided by experts to guide the search and accelerate convergence in black-box scientific experiments.
🎯 What it does: Proposed the SE-HSSL framework, leveraging unsampled CCA objectives and hierarchical member contrastive learning to achieve self-supervised representation learning on hypergraphs.
🎯 What it does: This paper investigates the optimistic bias and unsafe behaviors that occur when using typical unbiased estimators (e.g., IPS) as proxy targets for automatic hyperparameter optimization (HPO) in offline (off-policy) learning, and proposes the CIR-HPO algorithm to address these two issues.
🎯 What it does: Propose a post-training quantization (PTQ) method for hybrid models combining CNN and Vision Transformer, achieving 8-bit integer inference while maintaining high accuracy.
Improved Approximation Algorithms for Capacitated Location Routing
Jingyang Zhao (University of Electronic Science and Technology of China), Shunwang Wang (University of Electronic Science and Technology of China)
CodeOptimizationBenchmark
🎯 What it does: Proposed two improved approximation algorithms (Tree-Alg and Path-Alg), achieving approximation ratios of 4.169 and 4.092, respectively, for solving the capacity-constrained location routing problem (CLR).
Improved Encodings of Acyclicity for Translating Answer Set Programming into Integer Programming
Masood Feyzbakhsh Rankooh (Tampere University), Tomi Janhunen (Tampere University)
CodeOptimizationBenchmark
🎯 What it does: This paper proposes two novel integer programming (IP) translation methods to improve acyclicity constraints in answer set programming (ASP), thereby more effectively solving ASP problems.
Individual Causal Structure Learning from Population Data
Wei Chen (Guangdong University of Technology), Zhifeng Hao (Shantou University)
CodeTabularBiomedical DataFinance Related
🎯 What it does: This paper proposes a new individual linear acyclic model (ILAM) and develops an individual causal structure learning (ICSL) method based on shared independent component analysis (ShICA), which can utilize aggregate data from multiple individuals to recover their specific causal structures even when each individual's sample size is limited.
🎯 What it does: Design and implement the InfoMatch framework, combining information entropy neural estimation with pseudo labels and strong contrastive learning to enhance semi-supervised image classification performance.
🎯 What it does: This paper proposes using spherical harmonic transform (SHT) coefficients as auxiliary input, designing parallel and serial dual-encoder networks for multi-channel speech enhancement, achieving unified processing for different microphone array configurations.
Long Vu (IBM Research), Horst Samulowitz (IBM Research)
CodeAnomaly DetectionMeta LearningTabular
🎯 What it does: This work proposes a novel anomaly detection method called T-AutoOD based on instance-level meta-learning. It learns the structural relationships between abnormal and normal instances from an imbalanced labeled classification dataset, and builds a meta-model to combine scores from multiple unsupervised anomaly detection pipelines, enabling anomaly detection on new unlabeled datasets.
InstructME: An Instruction Guided Music Edit Framework with Latent Diffusion Models
Bing Han (Shanghai Jiao Tong University), Xuchen Song (ByteDance)
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelAuto EncoderTextAudio
🎯 What it does: An instruction-driven music editing framework called InstructME based on latent diffusion models, supporting operations such as adding, deleting, replacing, and mixing audio tracks.
🎯 What it does: Proposes IntensPure, an adversarial attack purification method for pedestrian re-identification, based on attack intensity awareness and quadratic domain adaptive diffusion, purifying only low/mid-frequency coefficients.
Jinmin Li (Tsinghua University), Jingyun Zhang (Tencent)
CodeSuper ResolutionFlow-based ModelImage
🎯 What it does: Propose a reversible residual rescaling model (IRRM) that can efficiently learn bijective relationships between high-resolution and low-resolution images, and enhance high-frequency information reconstruction through residual modules.
🎯 What it does: This paper constructs a cross-platform cascade dataset for similar topics spreading on Twitter and Weibo, and proposes a dual-channel source localization framework (DSLF), achieving precise localization of cross-platform source nodes through technologies such as self-loop attention GCN, dual structure KL regularization, and exponential VAE.
🎯 What it does: Proposed the KDDC framework, which combines knowledge graph semantic representation, causal separation, and metric learning to model users' interests and herd behavior, achieving intent inference and recommendation for pre-travel POI (Points of Interest).
🎯 What it does: Proposed a kernel-based graph pooling method called KerRead, which maps node embeddings to an infinite-dimensional enhanced space via kernel computations after using learnable adaptive centers.
KG-CoT: Chain-of-Thought Prompting of Large Language Models over Knowledge Graphs for Knowledge-Aware Question Answering
Ruilin Zhao (Huazhong University of Science and Technology), Guandong Xu (University of Technology Sydney)
CodeExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the KG-CoT framework, combining a small-scale graph reasoning model with a large LLM. The framework leverages knowledge graphs to generate high-confidence chain-of-thought reasoning paths, providing interpretable and responsible knowledge support for LLMs.
Knowledge Compilation for Incremental and Checkable Stochastic Boolean Satisfiability
Che Cheng (National Taiwan University), Jie-Hong R. Jiang (National Taiwan University)
CodeExplainability and InterpretabilityComputational EfficiencyBenchmark
🎯 What it does: Proposed a knowledge compilation method based on levelized dec-DNNF to implement incremental queries and verifiable SSAT solving in the SSAT solver SharpSSAT, and extended CPOG to generate verifiable proof logs.
🎯 What it does: In the open-world object detection task, the KTCN framework is proposed, which utilizes the Segment Anything Model (SAM) to generate pseudo labels and improves the recall rate of unknown categories by combining the Dual Matching Label Assignment method with the Class-Awareness Neutralizer method.
🎯 What it does: Established the ScribbleSC benchmark for semantic scene completion requiring only sparse scribble annotations, and proposed the Scribble2Scene weakly supervised method, which utilizes geometric self-supervised labelers and range-guided offline-to-online distillation to achieve 3D semantic occupancy prediction under sparse annotations.
Large Language Model as a Policy Teacher for Training Reinforcement Learning Agents
Zihao Zhou (Zhejiang Lab), Bin Liu (Zhejiang Lab)
CodeKnowledge DistillationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes the LLM4Teach framework, which utilizes a pre-trained large language model (LLM) as a teacher to provide uncertainty-aware soft instructions to a lightweight reinforcement learning (RL) student agent. Through policy distillation, the student rapidly acquires high-level planning capabilities in the early training phase and gradually transitions to relying solely on environmental feedback to improve performance in later stages.
Layered and Staged Monte Carlo Tree Search for SMT Strategy Synthesis
Zhengyang Lu (University of Waterloo), Vijay Ganesh (Georgia Institute of Technology)
CodeOptimizationComputational EfficiencyBenchmark
🎯 What it does: This paper proposes Z3alpha, an automated SMT strategy synthesis method based on Monte Carlo Tree Search (MCTS), designed to generate efficient solving strategies on given instance sets.
Laying the Foundations for Solving FOND HTN Problems: Grounding, Search, Heuristics (and Benchmark Problems)
Mohammad Yousefi (Australian National University), Pascal Bercher (Australian National University)
CodeOptimizationBenchmark
🎯 What it does: This paper proposes the first method capable of solving hierarchical task network (HTN) problems in fully observable non-deterministic (FOND) environments. The approach encompasses 'all-outcome determinization' compilation from FOND HTN domains to deterministic domains, a strong solution search algorithm based on AO*, heuristic evaluation leveraging deterministic HTN and classical planning heuristics, and corresponding solver implementation.
🎯 What it does: Proposed an efficient single-image de-raining model ESDNet based on spiking neural networks, addressing the challenges of binary activation information loss and training difficulties caused by rain pixels.
Learning Conditional Preference Networks: An Approach Based on the Minimum Description Length Principle
Pierre-François Gimenez, Jérôme Mengin
CodeRecommendation SystemTabular
🎯 What it does: This study proposes an unsupervised learning method based on the Minimum Description Length (MDL) principle to automatically infer Conditional Preference Networks (CP-nets) from sales history data, and evaluates it on car configuration recommendation tasks.
🎯 What it does: Propose a fair representation learning method (FairIB) based on the Information Bottleneck principle to generate user representations in recommendation systems that retain collaborative filtering information while suppressing sensitive attribute information.
🎯 What it does: Proposed and implemented a semantic separation POI category embedding model (SD-CEM) based on decoupled mobile sequences, generating hierarchically enhanced POI category representations.
🎯 What it does: Propose a single Hypernet that learns the Pareto set for multi-objective continuous robot control problems, representing the entire Pareto front as a low-dimensional curve in a high-dimensional parameter space.
Learning Robust Classifiers with Self-Guided Spurious Correlation Mitigation
Guangtao Zheng (University of Virginia), Aidong Zhang (University of Virginia)
CodeClassificationVision Language ModelImage
🎯 What it does: Propose a self-guided, unlabeled pseudo-correlation elimination framework called LBC, which automatically leverages pre-trained vision-language models (VLMs) to extract image attributes, computes pseudo-correlation scores (spuriousness scores) between each attribute and category, clusters in the spuriousness embedding space to generate fine-grained labels, and trains a classifier less sensitive to pseudo-correlations and with higher robustness by reshaping the classification head into (K·C) classes, balanced sampling, and iterative updates.
🎯 What it does: Propose a few-shot object counting network SSD based on 4D spatial similarity distribution, combining feature cross enhancement (FCE), dynamic image scaling, and generalization loss;
Learning Translations: Emergent Communication Pretraining for Cooperative Language Acquisition
Dylan Cope (King's College London), Peter McBurney (King's College London)
CodeReinforcement LearningText
🎯 What it does: In multi-agent collaborative environments, the Cooperative Language Acquisition Problem (CLAP) challenge is proposed, constructing a joiner agent to learn and adapt to the communication protocols of the target community.
🎯 What it does: Propose a new visual Transformer architecture called LeMeViT, which achieves sparse representation in high-resolution remote sensing images by utilizing learnable meta tokens, and significantly reduces the computational complexity of self-attention through Dual Cross-Attention (DCA).
LeRet: Language-Empowered Retentive Network for Time Series Forecasting
Qihe Huang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
CodeTransformerLarge Language ModelTime Series
🎯 What it does: Proposes the LeRet framework, which splits time series into patches, employs RetNet to capture causal dependencies, pretrains on patch-level autoregressive tasks, leverages pre-trained large language models (LLMs) to extract time series-related language knowledge, and achieves language-enabled sequence prediction through two-stage cross-modal integration.
LG-FGAD: An Effective Federated Graph Anomaly Detection Framework
Jinyu Cai (National University of Singapore), See-Kiong Ng (National University of Singapore)
CodeAnomaly DetectionFederated LearningSafty and PrivacyKnowledge DistillationGraph Neural NetworkGenerative Adversarial NetworkGraph
🎯 What it does: Proposed a federated graph anomaly detection framework named LG-FGAD, which generates abnormal graphs at each client via a self-adversarial generator, detects anomalies using a discriminator, and achieves model personalization and collaborative learning through a local-global mutual information module and dual knowledge distillation.
LLMem: Estimating GPU Memory Usage for Fine-Tuning Pre-Trained LLMs
Taeho Kim (University of Colorado Boulder), Sangtae Ha (University of Colorado Boulder)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes LLMem, a framework that accurately predicts GPU memory consumption during fine-tuning of large language models (LLMs) in a multi-GPU environment and automatically selects the optimal distributed fine-tuning method based on the prediction results.
🎯 What it does: Propose M2Beats, a framework for action beat detection and enhancement in short videos, which includes a high-quality action beat dataset AIST-M2B, a spatiotemporal graph convolutional network-based model M2BNet, and an algorithm for aligning and enhancing action beats with music beats.
Huiqiang Chen (University of Technology Sydney), Wanlei Zhou (City University of Macau)
CodeSafty and PrivacyImage
🎯 What it does: Proposes a machine unlearning method called UNSC under zero-space calibration, which efficiently and accurately enables the model to forget specific samples when data deletion requests are made.
🎯 What it does: Proposed the MARCO framework, which integrates a memory module with construction and improvement methods from neural combinatorial optimization (NCO) to enhance search efficiency and solution quality.
🎯 What it does: The MARS framework achieves precise perception of joint position, orientation, and status by fusing RGB images with point clouds and combining an active sensing strategy;
🎯 What it does: Proposed the MAS-SAM framework, customizing the Segment Anything Model (SAM) for marine animal segmentation, primarily by incorporating LoRA and Adapter into the SAM encoder for low-parameter fine-tuning, constructing a Hypermap Extraction Module (HEM) to extract multi-scale features, and designing a Progressive Prediction Decoder (PPD) and Fusion Attention Module (FAM) to achieve multi-source feature fusion, ultimately achieving fine-grained segmentation results.
🎯 What it does: In distributed learning, this paper investigates the robustness of the mean aggregator and existing robust aggregators under label poisoning attacks, and proves that the mean aggregator is theoretically and practically superior when data heterogeneity is sufficiently large;
MEDVOC: Vocabulary Adaptation for Fine-tuning Pre-trained Language Models on Medical Text Summarization
Gunjan Balde (Indian Institute of Technology Kharagpur), Niloy Ganguly (Indian Institute of Technology Kharagpur)
CodeGenerationHyperparameter SearchData-Centric LearningTransformerSupervised Fine-TuningTextBiomedical Data
🎯 What it does: This paper proposes a dynamic vocabulary adaptation strategy called MEDVOC for fine-tuning pre-trained language models, thereby improving the quality of medical text summarization.
Memorizing Documents with Guidance in Large Language Models
Bumjin Park (KAIST AI), Jaesik Choi (KAIST AI)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose a document-based memory architecture and document-guided loss, enabling large language models to map document content to trackable memory entries during training.
Meta In-Context Learning Makes Large Language Models Better Zero and Few-Shot Relation Extractors
Guozheng Li (Southeast University), Zijie Xu (Southeast University)
CodeClassificationMeta LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTabular
🎯 What it does: Propose the MICRE framework, which utilizes meta-learning to train LLMs for context learning, aiming to enhance performance in zero-shot and few-shot relation extraction tasks.
MLP-DINO: Category Modeling and Query Graphing with Deep MLP for Object Detection
Guiping Cao (Southern University of Science and Technology), Yaowei Wang (Peng Cheng Laboratory)
CodeObject DetectionTransformerImage
🎯 What it does: Proposed MLP-DINO, which integrates a deep MLP backbone network with the DINO framework, and introduces two strategies, QICS and GQS, to decouple class prediction from box regression and balance the query distribution;
MMVQA: A Comprehensive Dataset for Investigating Multipage Multimodal Information Retrieval in PDF-based Visual Question Answering
Yihao Ding (University of Melbourne), Soyeon Caren Han (University of Melbourne)
CodeRetrievalTransformerVision Language ModelMultimodalityBiomedical DataBenchmark
🎯 What it does: Constructed the MMVQA dataset (multi-page, multi-modal PDF Visual Question Answering (VQA)), and proposed a joint coarse-to-fine entity retrieval framework capable of locating paragraphs, tables, figures, and other entities across pages in documents to answer questions.
🎯 What it does: This paper proposes a model-free preference mining framework that directly learns user response probabilities to questions and the expected utility of recommendations through function approximation, achieving an expectation of information gain (EVOI)-driven multi-round querying strategy without relying on explicit Bayesian inference.
Modeling Personalized Retweeting Behaviors for Multi-Stage Cascade Popularity Prediction
Mingyang Zhou (Shenzhen University), Rui Mao (Shenzhen University)
CodeRecommendation SystemConvolutional Neural NetworkGraph Neural NetworkTransformerLarge Language ModelTextGraph
🎯 What it does: This study proposes the CasMS framework for predicting the popularity of information diffusion from the message generation stage to the long-term stage.
Modeling Selective Feature Attention for Lightweight Text Matching
Jianxiang Zang (Shanghai University of International Business and Economics), Hui Liu (Shanghai University of International Business and Economics)
CodeRetrievalRecurrent Neural NetworkText
🎯 What it does: Proposed Feature Attention (FA) and Selective Feature Attention (SFA) to enhance matching performance in lightweight text matching networks by modeling dependencies in the embedded feature layer.
MuEP: A Multimodal Benchmark for Embodied Planning with Foundation Models
Kanxue Li (Yunnan University), Xiaodong He (JD Explore Academy)
CodeRobotic IntelligenceLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed MuEP, a multi-modal embedded planning benchmark for evaluating embedded agents in complex scenarios with multi-round interactions;
MultifacetEval: Multifaceted Evaluation to Probe LLMs in Mastering Medical Knowledge
Yuxuan Zhou (Tsinghua University), Ji Wu (Tsinghua University)
CodeLarge Language ModelPrompt EngineeringBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Propose a multi-faceted evaluation framework, MultifactEval, which designs four evaluation dimensions (comparison, correction, differential diagnosis, verification) based on the same medical knowledge point, and constructs two multi-faceted medical evaluation datasets, MultiDiseK and MultiMedQA, to systematically assess the depth and breadth of LLMs' mastery of medical knowledge.
Multimodal Representation Distribution Learning for Medical Image Segmentation
Chao Huang (Sun Yat-sen University), Zhihua Wang (Shenzhen MSU-BIT University)
CodeSegmentationConvolutional Neural NetworkTransformerVision Language ModelMultimodalityBiomedical Data
🎯 What it does: This paper proposes a multimodal medical image segmentation method that enhances segmentation performance by fusing medical text annotations with image features.
NegativePrompt: Leveraging Psychology for Large Language Models Enhancement via Negative Emotional Stimuli
Xu Wang (School of Artificial Intelligence, Jilin University), Yuan Wu (School of Artificial Intelligence, Jilin University)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the NegativePrompt strategy, design and test 10 negative emotion stimuli, combine psychological theories (cognitive dissonance, social comparison, stress and coping) to enhance LLM prompts, and conduct experiments on five mainstream LLMs (Flan-T5-Large, Vicuna, Llama 2, ChatGPT, GPT-4).
NELLIE: A Neuro-Symbolic Inference Engine for Grounded, Compositional, and Explainable Reasoning
Nathaniel Weir, Benjamin Van Durme (Johns Hopkins University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a neural symbolic reasoning engine NELLIE based on large language models, which constructs interpretable proof trees through backward chaining search in natural language fact corpora to answer multiple-choice questions.
No Regularization Is Needed: Efficient and Effective Incomplete Label Distribution Learning
Xiang Li (Nanjing University of Aeronautics and Astronautics), Songcan Chen (Nanjing University of Aeronautics and Astronautics)
CodeClassification
🎯 What it does: This paper proposes a weighted incomplete label distribution learning method (WInLDL) without explicit regularization, modeling real-world scenarios where labels with small degrees are more prone to missingness.
🎯 What it does: This paper proposes a multi-view subspace clustering framework based on non-convex ℓq regularization, and provides an efficient solution algorithm along with convergence analysis;
Online Submodular Maximization via Adaptive Thresholds
Zhengchen Yang (Nanjing University of Aeronautics & Astronautics), Jiping Zheng (Nanjing University)
CodeOptimizationVideoTextTabular
🎯 What it does: Propose an online adaptive threshold algorithm ONLINEADAPTIVE for incrementally maximizing submodular functions in large streaming data under cardinality constraints.