Decomposing and Composing: Towards Efficient Vision-Language Continual Learning via Rank-1 Expert Pool in a Single LoRA
Zhan Fa (Nanjing University), Yinghuan Shi (Nanjing University)
CodeComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningMixture of ExpertsContrastive LearningMultimodality
π― What it does: In the continual learning scenario, we propose decomposing a single LoRA module into a dynamically combinable Rank-1 expert pool, using CLS semantics to guide expert selection, and designing an Activation-Guided Orthogonal (AGO) loss to reduce task interference.
CodeOptimizationExplainability and InterpretabilityTabular
π― What it does: The study decomposes direct and indirect bias in linear models under demographic parity constraints and proposes a post-processing analytical fair linear regression method.
π― What it does: Propose the DECON framework to achieve disentanglement and reconstruction of multi-human full-body clothing geometry from a single RGB image, and restore realistic spatial relationships through perspective-aware position optimization.
Decoupled Spatiotemporal Forecasting from Extreme Sparse Observations via Quantized Latent Space
Zhongnan Weng (Xiamen University), Xiangrong Liu (Xiamen University)
CodeRepresentation LearningData-Centric LearningTransformerAuto EncoderPhysics Related
π― What it does: By decoupling spatial reconstruction from temporal extrapolation of extremely sparse observations, high-precision spatial reconstruction and long-term temporal prediction are achieved in the quantized latent space.
Yifu Guo (Sun Yat-sen University), Ruixuan Wang (Sun Yat-sen University)
CodeSegmentationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImageText
π― What it does: Proposed the DecoupleCSS two-stage framework, decoupling class-aware detection from class-agnostic segmentation to achieve continual semantic segmentation; using task-specific LoRA adapters for class detection driven by language, and leveraging SAM to generate positional prompts for segmentation.
Decoupling Knowledge and Reasoning in LLMs: An Exploration Using Cognitive Dual-System Theory
Mutian Yang (Tsinghua University), Ji Wu (Tsinghua University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This work proposes to split the reasoning process of large language models into two stages: knowledge retrieval (fast thinking) and reasoning adjustment (slow thinking), and quantifies the contribution of knowledge and reasoning to performance by comparing two prompting methods. Subsequently, systematic experiments were conducted on 15 models and three categories of datasets.
Decoupling Shared and Personalized Knowledge: A Dual-Branch Federated Learning Framework for Multi-Domain with Non-IID Data
Yiran Pang (Florida Atlantic University), Xiangnan Zhong (Florida Atlantic University)
CodeDomain AdaptationFederated LearningImageBiomedical Data
π― What it does: Proposes a dual-branch personalized federated learning framework called pFedDB, which uses a two-phase training process to first learn expert models locally and then collaborates across multi-domain non-IID data through a shared branch to address catastrophic forgetting and negative transfer problems.
Decoupling Understanding from Reasoning via Problem Space Mapping for Small-Scale Model Reasoning
Li Wang (Beihang University), Wenjun Wu (Beihang University)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
π― What it does: Propose a framework that decouples understanding and reasoning by mapping natural language questions to a low-dimensional normalized question space, and implement a three-step alternating training algorithm called DURIT.
Decoupling What to Count and Where to See for Referring Expression Counting
Yuda Zou (Wuhan University), Yongchao Xu (Wuhan University)
CodeObject DetectionTransformerVision Language ModelImageMultimodality
π― What it does: Designed the W2-Net framework, addressing the mismatch between annotation points and attribute-related visual regions in Referring Expression Counting through dual query mechanisms (what-to-count and where-to-see) and Subclass Separable Matching (SSM), achieving more fine-grained subclass counting and localization.
Deep (Predictive) Discounted Counterfactual Regret Minimization
Hang Xu (Institute of Automation, Chinese Academy of Sciences), Jian Cheng (Tencent AI Lab)
CodeReinforcement LearningBenchmark
π― What it does: Proposed two model-agnostic neural CFR variants, VR-DeepDCFR+ and VR-DeepPDCFR+, which approximate the updates of DCFR+ and PDCFR+ through bootstrapping, discounting, and truncating cumulative advantages, while employing a value baseline to reduce variance;
Deep Clustering Based on Sparse Kolmogorov-Arnold Network and Spectral Constraint
Zixuan Bi (Northwestern Polytechnical University), Ganchao Liu (Northwestern Polytechnical University)
CodeOptimizationRepresentation LearningImageText
π― What it does: Proposes a deep clustering framework based on sparse Kolmogorov-Arnold Network (KAN) and spectral constraints, utilizing an adaptive adjacency matrix for unsupervised clustering.
π― What it does: Propose a deep incomplete multi-view clustering framework DIMVC-HIA, achieving high-quality clustering through hierarchical missing value imputation and alignment.
π― What it does: Propose a scalable deep reinforcement learning method to solve the offline 3D packing problem, capable of efficiently planning layouts for a large number of items (20β1000 pieces).
DeepPhy: Benchmarking Agentic VLMs on Physical Reasoning
Xinrun Xu (Taobao & Tmall Group of Alibaba), Bo Zheng (Taobao & Tmall Group of Alibaba)
CodeAgentic AIVision Language ModelVision-Language-Action ModelWorld ModelMultimodalityBenchmarkPhysics Related
π― What it does: Propose the DeepPHY benchmark to systematically evaluate the capabilities of vision-language models in interactive physical reasoning tasks.
DeepProofLog: Efficient Proving in Deep Stochastic Logic Programs
Ying Jiao (KU Leuven), Giuseppe Marra (University of Siena)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement LearningImageTextGraph
π― What it does: Propose DeepProofLog (DPrL), a neural network-based deep stochastic logic program that can guide proof steps in real-time during reasoning and model reasoning as a Markov Decision Process (MDP).
π― What it does: Proposed an end-to-end differentiable DeepRAHT framework, achieving complete training and compression workflow of RAHT in deep learning.
DeepSenseMoE: Harnessing Power of Time Series Foundation Models for Few-Shot Human Activity Recognition
Zenan Fu (Nanjing Normal University), Hao Wu (Yunnan University)
CodeRecognitionTransformerSupervised Fine-TuningMixture of ExpertsContrastive LearningMultimodalityTime Series
π― What it does: This paper proposes the DeepSenseMoE module, which performs parameter-efficient fine-tuning of pre-trained time series foundation models through multi-scale convolution Mixture-of-Experts to address the scarcity and heterogeneity of wearable sensor data.
DeepTracer: Tracing Stolen Model via Deep Coupled Watermarks
Yunfei Yang (Chinese Academy of Sciences), He Li (Chinese Academy of Sciences)
CodeSafty and PrivacyAdversarial AttackImage
π― What it does: DeepTracer proposes a deeply coupled watermarking framework that enhances watermark robustness by strengthening the coupling between the watermark task and the main task to defend against model stealing attacks.
π― What it does: Propose the DeFB framework to achieve an end-to-end real-time system for multi-person face detection, tracking, and eye movement detection;
π― What it does: Proposes a stealthy data poisoning method called Deferred Poisoning Attack (DPA), which utilizes Hessian matrix singularization to enhance model local curvature, making the model perform normally during training and validation stages but become highly vulnerable to adversarial attacks and natural noise after deployment.
Delayed Feedback Modeling with Influence Functions
Chenlu Ding (University of Science and Technology of China), Andrew Rabinovich (Upwork)
CodeRecommendation SystemTabularSequential
π― What it does: This paper proposes an influence function-based delayed feedback modeling framework, IF-DFM, which can directly estimate the impact of label flipping and new incoming data on model parameters, enabling efficient updates to the CVR prediction model without requiring full retraining.
DeLightMono: Enhancing Self-Supervised Monocular Depth Estimation in Endoscopy by Decoupling Uneven Illumination
Mingyang Ou (Southern University of Science and Technology), Jiang Liu (Southern University of Science and Technology)
CodeDepth EstimationConvolutional Neural NetworkBiomedical Data
π― What it does: Proposed a self-supervised monocular depth estimation framework called DeLightMono, which separates endoscopic images using an illumination-reflection-depth (IRD) model to mitigate the impact of uneven illumination on depth estimation.
Xiwei Liu (Mohamed bin Zayed University of Artificial Intelligence), Imran Razzak (Mohamed bin Zayed University of Artificial Intelligence)
CodeComputational EfficiencyMixture of ExpertsMultimodalityBenchmark
π― What it does: Proposes the DeLo framework, which utilizes Dual Decomposed Low-Rank Experts to address the Continual Missing Modality Learning (CMML) problem, balancing parameter efficiency and robustness to modality missing scenarios.
π― What it does: By constructing a spectral-to-spatial mapping theorem (S2SMT) and proposing a DeloopSGNN, the spectral GNN is transformed into spatial aggregation, eliminating cycles and over-smoothing caused by multiple aggregations, thereby enhancing expressiveness and adversarial robustness.
π― What it does: Developed DeNC++, a lightweight neural-enhanced video streaming (NeVS) solution capable of compressing and restoring videos on edge devices without relying on media servers;
DenoDet V2: Phase-Amplitude Cross Denoising for SAR Object Detection
Kang Ni (Nanjing University of Posts and Telecommunications), Yimian Dai (Nankai University)
CodeObject DetectionImagePhysics Related
π― What it does: This paper proposes DenoDet V2, which leverages the complementary characteristics of SAR image amplitude and phase in the frequency domain, achieving more refined denoising and detection through phase-guided soft threshold denoising and phase-amplitude token exchange mechanisms.
DensiCrafter: Physically-Constrained Generation and Fabrication of Self-Supporting Hollow Structures
Shengqi Dang (Shanghai Research Institute for Intelligent Autonomous Systems), Nan Cao (Shanghai Research Institute for Intelligent Autonomous Systems)
CodeGenerationOptimizationImageTextPhysics Related
π― What it does: Propose DensiCrafter, achieving lightweight and self-supporting hollow structures by optimizing the continuous density field of the voxel grid output from 3D generative models.
Dep-MAP: A Multi-level Alignment Framework with Semantic Prototypes for Video-based Automatic Depression Assessment
Hao Wang (Qilu University of Technology), Qingxiang Wang (Qilu University of Technology)
CodeClassificationConvolutional Neural NetworkVision Language ModelContrastive LearningVideo
π― What it does: Proposed the Dep-MAP video-based automatic depression assessment framework, which utilizes a dual-branch structure to extract visual and emotional semantic features, and achieves key frame selection and depression severity judgment through semantic prototype clustering, cross-layer contrastive learning, and multi-scale fusion.
π― What it does: This work proposes the Departures framework, which utilizes the SchrΓΆdinger Bridge (SB) approximation to achieve single-cell perturbation prediction. By combining discrete (gene activation states) and continuous (gene expression levels) bridge models, and adopting MiniBatch OT to directly obtain control-perturbed sample pairings, it avoids the challenges of bidirectional models and reverse processes in traditional SB. Ultimately, it achieves high-precision distribution-level prediction of single-cell gene expression and activation states under different gene or compound perturbation conditions.
Depth-Synergized Mamba Meets Memory Experts for All-Day Image Reflection Separation
Siyan Fang (Huazhong University of Science and Technology), Yuehuan Wang (Huazhong University of Science and Technology)
CodeRestorationDepth EstimationTransformerMixture of ExpertsImage
π― What it does: This paper proposes the Depth-Memory Decoupling Network (DMDNet), achieving all-weather (especially nighttime) image reflection separation through depth-guided scanning and memory experts.
Description Logics with Two Types of Definite Descriptions: Complexity, Expressiveness, and Automated Deduction
MichaΕ SochaΕski (University of ΕΓ³dΕΊ), MichaΕ Zawidzki (University of ΕΓ³dΕΊ)
CodeComputational Efficiency
π― What it does: Studied the extension of description logic ALC with two types of deterministic descriptions (local {ΞΉC} and global ΞΉC.D), analyzing their complexity, expressiveness, and automatic reasoning capabilities.
Detecting Emotional Dynamic Trajectories: An Evaluation Framework for Emotional Support in Language Models
Zhouxing Tan (Peking University), Junfei Liu (Peking University)
CodeLarge Language ModelTextBenchmark
π― What it does: This paper constructs an evaluation framework based on emotional trajectories, using large-scale simulated dialogues to assess the long-term dynamic performance of LLMs in emotional support.
π― What it does: Propose DEF (Detection-based Event Forecasting), which utilizes matching loss to simultaneously and in parallel predict multiple future event sequences over a long period.
π― What it does: This paper proposes DFDT (Dynamic Fast Decision Tree), a data stream learning algorithm for IoT edge devices, which achieves efficient online decision tree construction through activity-aware pre-pruning and adaptive parameter control.
π― What it does: Propose an unsupervised multimodal change detection method that uses a dual-branch GKAN (Graph KolmogorovβArnold Network) to construct an autoencoder, extracting common features of spatial-spectral structures and directly comparing them to obtain a change map.
DGP: A Dual-Granularity Prompting Framework for Fraud Detection with Graph-Enhanced LLMs
Yuan Li (National University Of Singapore), Cheng Chen (National University Of Singapore)
CodeAnomaly DetectionLarge Language ModelPrompt EngineeringTextGraphTabularFinance Related
π― What it does: Proposed and implemented the Dual-Granularity Prompting (DGP) framework, which classifies heterogeneous fraud detection graphs using text prompts and graph-enhanced LLMs, retaining fine-grained text of target nodes and performing coarse-grained compression of neighbor information.
π― What it does: Proposed the DHMRec framework in multi-modal recommendation, first separating public and exclusive features through collaborative modal decoupling, then enhancing recommendation performance via graph neural network multi-view learning and hierarchical fusion.
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Proposed the DIAA method, achieving model-agnostic and draft-model-free inference acceleration for large language models (LLMs) on edge devices;
π― What it does: For semi-supervised multi-label learning, a method is proposed to enhance model performance by calibrating the distribution of pseudo-label weights.
π― What it does: This study proposes the DICE method, which trains a lightweight sharpener to weight text embeddings without guided sampling, thereby achieving image quality comparable to traditional Classifier-Free Guidance (CFG) without increasing additional model evaluations;
π― What it does: This paper proposes Diff-NAT, a naturalistic physical adversarial patch generation method based on class-optimized diffusion, which can significantly interfere with object detection models while maintaining a highly natural appearance.
DIFFA: Large Language Diffusion Models Can Listen and Understand
Jiaming Zhou (Nankai University), Xuelong Li (China Telecom)
CodeRecognitionLarge Language ModelDiffusion modelAudio
π― What it does: Developed DIFFA, the first large-scale audio-language model based on diffusion, which utilizes a frozen diffusion LLM and a lightweight dual adapter to achieve speech understanding and instruction execution.
Differentially Private Subspace Fine-Tuning for Large Language Models
Lele Zheng (Xidian University), Yulong Shen (Institute of Science Tokyo)
CodeSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Introduce differential privacy into the fine-tuning process of large language models, proposing the DP-SFT method that injects noise only within a task-specific low-dimensional subspace;
π― What it does: Proposed a diffusion model based on difficulty control, using learning difficulty as a conditional signal to synthesize training data.
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Proposes a dual-evaluation curriculum learning framework called DUCL for LLM fine-tuning, which prioritizes low-difficulty, high-benefit samples during early training and gradually introduces high-difficulty samples later.
π― What it does: MonoDLGD provides explicit geometric supervision to enhance monocular 3D object detection by applying difficulty-aware perturbation and reconstruction to 3D object labels during training;
π― What it does: Proposed a task difficulty-aware learning curve extrapolation framework DA-LCE, which achieves precise extrapolation of learning curves in automated machine learning by dynamically quantifying difficulty through early curves, generating synthetic data via conditional diffusion, and employing Transformer predictors.
DiffOP: Reinforcement Learning of Optimization-Based Control Policies via Implicit Policy Gradients
Yuexin Bian (University of California, San Diego), Yuanyuan Shi (University of California, San Diego)
CodeOptimizationReinforcement Learning
π― What it does: Propose the DiffOP framework, which employs implicit gradients to achieve reinforcement learning optimization control strategies without value function approximation, jointly learning cost and dynamics models;
DiffRefiner: Coarse to Fine Trajectory Planning via Diffusion Refinement with Semantic Interaction for End to End Autonomous Driving
Liuhan Yin (Polytechnic Institute, Zhejiang University), Erkang Cheng (Nullmax)
CodeAutonomous DrivingTransformerDiffusion model
π― What it does: Propose DiffRefiner, a two-stage end-to-end autonomous driving trajectory prediction framework that first generates coarse-grained trajectory candidates using a Transformer, then refines them with a diffusion model to satisfy scene constraints.
π― What it does: Propose a framework combining score distillation with Direct Preference Optimization (Distillation-DPO) for efficient 3D LiDAR scene completion;
Diffusion Once and Done: Degradation-Aware LoRA for All-in-One Image Restoration
Ni Tang (Xiamen University), Yanyun Qu (Xiamen University)
CodeRestorationDiffusion modelImage
π― What it does: Proposes a one-stage, one-sampling stable diffusion model called Diffusion Once and Done (DOD) for universal image restoration across multiple degradation types.
Diffusion-Assisted Progressive Learning for Weakly Supervised Phrase Localization
Pengyue Lin (Beijing University of Posts and Telecommunications), Ruifan Li (Beijing University of Posts and Telecommunications)
CodeObject DetectionGenerationVision Language ModelDiffusion modelImageTextMultimodality
π― What it does: Proposes the DAPO (Diffusion-Assisted Progressive Learning) framework, which first generates images of scenes with few objects using Stable Diffusion, enhances semantic alignment through attention guidance and diffusion-guided loss, and subsequently trains a weakly supervised phrase localization network in a progressive manner based on sample difficulty.
π― What it does: Proposed the DiCoSeg framework, which utilizes diffusion models to perform contextual semantic reconstruction on sparsely annotated point clouds, recovering complete segmentation information from noise.
π― What it does: Propose a diffusion-based personalized pathology disentanglement model, DPPD, which can simultaneously perform visual gait scoring, dementia subtyping classification, and visualize abnormal joint localization.
DiM-TS: Bridge the Gap Between Selective State Space Models and Time Series for Generative Modeling
Zihao Yao (Tongji University), Yaying Zhang (Tongji University)
CodeGenerationData SynthesisDiffusion modelTime SeriesFinance Related
π― What it does: Propose the DiM-TS framework, applying the selective state space model (Mamba) to unsupervised time series generation, and achieving joint modeling of temporal and channel dimension features through a dual-channel encoder-decoder structure.
π― What it does: Proposes the DIMM framework, combining a 3D decoupled multi-level Kalman filter with an adaptive fusion network based on reinforcement learning, achieving high-precision tracking of unknown dynamic targets.
π― What it does: This paper proposes a Dual Impulse Network (DIN), which extracts information from both the attribute axis and channel axis, and generates an attention matrix through an integration network to achieve stronger multi-view representations;
π― What it does: Propose the SIEVE framework, enhancing representation robustness through controlled uncertainty injection and contrastive learning, and improve information diffusion prediction by introducing uncertainty-aware directed aggregation.
π― What it does: This paper proposes a direction-sensitivity-based knowledge distillation method called DSKD, which achieves more efficient low-rank knowledge transfer by dynamically selecting singular directions most sensitive to loss.
DISC: Dynamic Feature Selection for Cost-Sensitive Medical Diagnosis
Yu-sheng Li (Nanjing University), Han-jia Ye (Nanjing University)
CodeComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsTextMultimodalityTabularElectronic Health Records
π― What it does: This paper proposes a dynamic feature selection framework called DISC, which gradually selects experiments during the medical diagnosis process according to the cost sensitivity of individual patients, thereby reducing resource consumption while maintaining diagnostic accuracy.
π― What it does: Propose the DisCo DETR self-supervised pre-training framework, enhancing DETR's localization and semantic feature learning through distance-aware multi-view object query fusion and contrastive learning.
Discovering Decoupled Functional Modules in Large Language Models
Yanke Yu (Hong Kong University of Science and Technology), Yi Zheng (Huawei Technologies Ltd)
CodeExplainability and InterpretabilityLarge Language ModelText
π― What it does: Proposed an unsupervised LLM cross-layer module discovery framework (ULCMOD), aiming to simultaneously decouple neurons in large language models and discover input sample topics associated with these modules.
Discretization Is Not Always Better: Rethinking Deep Quantization for Asymmetric Image Retrieval
Xinze Liu (Institute of Information Engineering, Chinese Academy of Sciences), Pengwen Dai (Baidu Research)
CodeRetrievalConvolutional Neural NetworkImage
π― What it does: In resource-constrained scenarios, a novel deep cross-modal alignment hashing framework (DCAH) is proposed, which enhances cross-modal image retrieval performance through implicit binary quantization of the query model via a correlation-aligned quantization strategy (CAQ).
Distillation Dynamics: Towards Understanding Feature-Based Distillation in Vision Transformers
Huiyuan Tian (Zhejiang University), Shijian Li (Zhejiang University)
CodeKnowledge DistillationTransformerImage
π― What it does: Investigated why feature-level knowledge distillation fails on Vision Transformers and proposed a multi-dimensional analysis framework called 'distillation dynamics'.
π― What it does: This paper proposes a frequency-domain decomposition-based cross-modal knowledge distillation method (FD-CMKD), which first decomposes the teacher and student features into low-frequency (cross-modal common information) and high-frequency (modality-specific information) components via Fourier transform, and then distills them using strong consistency (MSE) and weak consistency (logMSE) losses respectively, while incorporating feature normalization and a shared classifier to achieve scale and spatial alignment.
Distributionally Robust Online Markov Game with Linear Function Approximation
Zewu Zheng (Chinese University of Hongkong), Yuanyuan Lin (Chinese University of Hongkong)
CodeReinforcement LearningSequentialBenchmark
π― What it does: For multi-agent online distributed robust Markov games, a sample-efficient algorithm named DR-CCE-LSI based on linear function approximation is proposed to find an Ξ΅-approximate robust coarse correlated equilibrium.
Diversifying Counterattacks: Orthogonal Exploration for Robust CLlP Inference
Chengze Jiang (Southeast University), Jie Gui (Southeast University)
CodeAdversarial AttackTransformerVision Language ModelContrastive LearningImage
π― What it does: For test-time adversarial defense of the CLIP model, the Orthogonal Exploration Directional Adversarial method (DOC) is proposed, which enhances the diversity of adversarial examples by introducing orthogonal gradients and momentum, thereby improving the defense against adversarial attacks.
CodeSegmentationDomain AdaptationAdversarial AttackContrastive LearningImageBiomedical Data
π― What it does: This paper proposes a method for cross-domain few-shot semantic segmentation by splitting encoder features into class-specific private features and domain-related shared features, and achieving cross-domain adaptation through matrix-guided dynamic fusion.
π― What it does: This paper proposes a hierarchical style calibration prototype alignment framework called FedBCS, aimed at addressing feature bias caused by domain unevenness in federated medical segmentation.
π― What it does: Propose the DLDA framework, combining image-level photometric-aware contrastive translation and multi-scale conditional adversarial alignment to achieve unsupervised domain adaptation for low-light object detection.
π― What it does: Investigate whether existing audio-visual segmentation models truly integrate audio information, propose the AVSBench-Robust benchmark, and design a classifier-guided similarity learning method to enhance model robustness against negative audio scenarios.
Do It for HER: First-Order Temporal Logic Reward Specification in Reinforcement Learning
Pierriccardo Olivieri (Politecnico di Milano), Matteo Papini (Universita degli Studi di Milano)
CodeReinforcement Learning
π― What it does: This paper proposes a non-Markovian reward specification framework based on LTLfMT (Linear Temporal Logic modulo Theory), which can directly evaluate first-order logic predicates in continuous control environments using SMT solvers, thereby avoiding the need for manually writing label functions.
Do Language Models Associate Sound with Meaning? A Multimodal Study of Sound Symbolism
Jinhong Jeong (Yonsei University), Youngjae Yu (Seoul National University)
CodeExplainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelMultimodalityAudio
π― What it does: Investigate the understanding of phonosemantic symbols (onomatopoeia and constructed pseudo-words) by multimodal large language models (MLLMs), construct the LEX-ICON dataset, and conduct semantic dimension prediction and internal attention analysis.
Do Large Language Models Reason About Uncertainty Like Humans? A Benchmark on Hurricane Forecast Visualization Comprehension
Le Liu (Northwestern Polytechnical University), Peng Wang (Northwestern Polytechnical University)
CodeTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: Built and utilized the UnReason benchmark to systematically evaluate differences between humans and LLMs in visualizing uncertainties in hurricane prediction.
Do Large Language Models Think like the Brain? Sentence-Level Evidences from Layer-Wise Embeddings and fMRI
Yu Lei (Beijing University of Posts and Telecommunications), Bolei Ma (LMU Munich)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelTextBiomedical DataMagnetic Resonance Imaging
π― What it does: Compare the correspondence between 14 types of LLM hierarchical representations and human brain fMRI at the sentence level, and investigate the relationship between instruction fine-tuning, semantic understanding, brain alignment, and hemispheric asymmetry.
Do LLMs Feel? Teaching Emotion Recognition with Prompts, Retrieval, and Curriculum Learning
Xinran Li (Dalian University of Technology), Xiujuan Xu (Dalian University of Technology)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: Propose the PRC-Emo framework, combining Prompt engineering, example retrieval, and curriculum learning, leveraging large language models to achieve explicit and implicit recognition of conversational emotions.
Do LLMs Really Struggle at NL-FOL Translation? Revealing Their Strengths via a Novel Benchmarking Strategy
Andrea Brunello (University of Udine), Nicola Saccomanno (University of Udine)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: A systematic evaluation of methods for assessing natural language to first-order logic (NL-FOL) translation was conducted, and a three-step benchmark based on ontology and logical translation splitting was proposed.
Do Retrieval Augmented Language Models Know When They Donβt Know?
Youchao Zhou (Beijing Institute of Technology), Yang Deng (Singapore Management University)
CodeExplainability and InterpretabilityTransformerSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: Investigate whether retrieval-augmented language models (RALM) can correctly refuse to answer when no answer is available, and systematically evaluate the relationship between their calibration, refusal behavior, and external retrieval context; propose a post-rejection method leveraging uncertainty thresholds and contextual information to balance refusal rates and answer quality.
Do We Truly Need So Many Samples? Multi-LLM Repeated Sampling Efficiently Scales Test-Time Compute
Jianhao Chen (Nanjing University), Shuyue Hu (Shanghai Artificial Intelligence Laboratory)
CodeComputational EfficiencyLarge Language ModelText
π― What it does: This paper proposes a test-time computational enhancement strategy called ModelSwitch, which leverages multi-model resampling and dynamically switches models based on consistency to improve answer accuracy while reducing sampling costs.
π― What it does: This paper proposes the Loddis framework, which improves infrared small target detection by learning to ignore domain differences. It uses auxiliary large-sample domain data in the limited-sample task domain to avoid the negative shift degradation problem.
DomainCQA: Crafting Knowledge-Intensive QA from Domain-Specific Charts
Yujing Lu (Zhejiang Lab), Qing Zhang (Zhejiang Lab)
CodeConvolutional Neural NetworkLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodalityBenchmarkPhysics RelatedChain-of-Thought
π― What it does: Design the DomainCQA framework and construct the AstroChart benchmark in the astronomy field, providing 482 charts and 1,690 QA pairs covering two-tier tasks of FQA and AQA.
Donβt Start Over: A Cost-Effective Framework for Migrating Personalized Prompts Between LLMs
Ziyi Zhao (University of Science and Technology of China), Fuli Feng (Huawei Technologies Co Ltd)
CodeRecommendation SystemLarge Language ModelPrompt EngineeringText
π― What it does: To address the issue of personalized soft prompts becoming ineffective after large language models (LLM) upgrades, the PUMA framework is proposed to achieve soft prompt transfer.
Donβt Stop the Multi-Party! On Generating Synthetic Written Multi-Party Conversations with Constraints
NicolΓ² Penzo (Fondazione Bruno Kessler), Sara Tonelli (Fondazione Bruno Kessler)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes using instruction-tuned large language models to generate synthetic written multi-party dialogues while satisfying constraints such as dialogue structure and speaker stance.
DOS: Directional Object Separation in Text Embeddings for Multi-Object Image Generation
Dongnam Byun (Seoul National University), Wonjong Rhee (Seoul National University)
CodeGenerationPrompt EngineeringVision Language ModelDiffusion modelImageTextBenchmark
π― What it does: Propose the DOS method, which improves the success rate of multi-object image generation by modifying CLIP text embeddings in text generation models.
π― What it does: Proposed the Double Rounding quantization method, combined with Adaptive Learning Rate Scaling (AdaScale) and Hessian-based Mixed Precision Search (HessBit), to achieve joint training of multi-precision and mixed-precision models in one go, enabling nearly lossless bit-width switching.