Wonjeong Choi (Korea Advanced Institute of Science and Technology), Jaekyun Moon (Korea Advanced Institute of Science and Technology)
CodeClassificationAdversarial AttackSupervised Fine-TuningVision Language ModelImageMultimodality
π― What it does: This paper proposes AdvMask, a few-shot adversarial robustness enhancement framework for vision-language models (VLMs), which screens robust subnetworks by applying binary masks on frozen pre-trained weights;
IGC-Net for conditional average potential outcome estimation over time
Konstantin Hess (LMU Munich), Stefan Feuerriegel (LMU Munich)
CodeRecurrent Neural NetworkTransformerTabularTime SeriesElectronic Health Records
π― What it does: Proposed an end-to-end neural network named IGC-Net for estimating time-varying conditional average potential outcomes (CAPO) in the presence of time-varying confounders;
IGGT: Instance-Grounded Geometry Transformer for Semantic 3D Reconstruction
Hao Li (Northwestern Polytechnical University), Ziwei Liu (Singapore Lab, Nanyang Technological University)
CodeSegmentationGenerationDepth EstimationTransformerVision Language ModelContrastive LearningGaussian SplattingImagePoint Cloud
π― What it does: Propose the IGGT framework, a unified end-to-end Transformer combining 3D geometric reconstruction and instance-level semantic understanding; build the InsScene-15K large-scale instance consistency dataset; propose an instance-driven scene understanding strategy;
IGU-LoRA: Adaptive Rank Allocation via Integrated Gradients and Uncertainty-Aware Scoring
Xuan Cui (Chongqing Technology and Business University), Zhanpeng Zhou (Shanghai Jiao Tong University)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: To address parameter-efficient fine-tuning for large language models, this paper proposes an adaptive rank allocation LoRA framework called IGU-LoRA, which calculates the importance of each parameter in the low-rank matrix using integrated gradients in the parameter space, and dynamically adjusts the LoRA rank per layer by combining uncertainty-awareness (EMA + variance) to select the most important singular values.
iLLaVA: An Image is Worth Fewer Than 1/3 Input Tokens in Large Multimodal Models
Lianyu Hu (Tianjin University), Wei Feng (Tianjin University)
CodeComputational EfficiencyLarge Language ModelVision Language ModelImageVideo
π― What it does: Proposed and implemented iLLaVA, which performs two-phase token merging on both the image encoder and LLM to significantly accelerate large vision-language models.
π― What it does: Proposed a multi-modal adversarial attack framework called RECALL, which uses adversarial image prompts combined with original text prompts to jointly attack machine-unlearned image generation models, thereby evaluating and weakening their unlearning effects.
CodeRobotic IntelligenceSupervised Fine-TuningVision Language ModelVision-Language-Action ModelImageBenchmark
π― What it does: Established an image quality assessment (IQA) task applicable to embodied AI, constructed a large-scale database named Embodied-IQA, and conducted a systematic evaluation of the performance of existing IQA methods on this database.
Imagine How To Change: Explicit Procedure Modeling for Change Captioning
Jiayang Sun (Soochow University), Jorma Laaksonen (Aalto University)
CodeGenerationTransformerVision Language ModelAuto EncoderImageMultimodality
π― What it does: Propose a two-stage ProCap framework that addresses the limitations of traditional methods which only compare static images, by explicitly modeling the image change process and implicitly describing it through queries.
π― What it does: Achieved text-to-image generation by optimizing a frequency-aware implicit neural representation (INR) on a frozen CLIP image encoder, without requiring a pre-trained decoder or CLIP fine-tuning.
Implicit Regularization of SGD Reduces Shortcut Learning
Nahal Mirzaie (Sharif University of Technology), Mohammad Hossein Rohban (Sharif University of Technology)
CodeOptimizationExplainability and InterpretabilityData-Centric LearningConvolutional Neural NetworkTransformerImageBenchmark
π― What it does: Investigated the impact of implicit regularization in stochastic gradient descent (SGD) under large learning rates and small batch sizes on models relying on spurious features, and validated the mechanism through theoretical proofs and experiments.
π― What it does: Proposed a fast βp regression algorithm based on iterative reweighted least squares (IRLS), with an implementation achieving iteration complexity that matches the theoretical optimum.
π― What it does: Propose the CODA framework, combining input-independent registration slots and contrastive alignment loss to improve diffusion-based object-centric learning
π― What it does: Proposes the Inter-Slice Consistent Stochasticity (ISCS) method, which utilizes spherical linear interpolation (Slerp) to generate correlated noise between adjacent slices during the resampling step of 2D diffusion models, thereby enhancing cross-slice consistency in 3D medical image reconstruction.
π― What it does: This paper proposes a two-stage framework called Structured Policy Initialization (SPIN), which first obtains an Action Structure Model (ASM) to capture the structure of discrete composite actions through self-supervised pre-training, and then freezes this representation and uses a lightweight policy head for control;
π― What it does: Studied black-box generative adversarial attacks, proposing to enhance the transfer performance across different models, domains, and tasks by applying Mean Teacher's semantic consistency constraints on early residual blocks of the generator.
Improving Block-Wise LLM Quantization by 4-bit Block-Wise Optimal Float (BOF4): Analysis and Variations
Patrick Blumenberg (Technische UniversitΓ€t Braunschweig), Tim Fingscheidt (Technische UniversitΓ€t Braunschweig)
CodeComputational EfficiencyLarge Language ModelSupervised Fine-TuningText
π― What it does: Studied and improved LLM weight quantization methods, proposing BOF4/BOF4-S and OPQ, and evaluated their effectiveness in inference and fine-tuning.
π― What it does: Propose and implement a capacity manipulation method that reserves and allocates model capacity specifically for the minority class, thereby improving the generation quality of diffusion models under imbalanced data scenarios.
Improving the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models
Weiqing He (University of Pennsylvania), Qi Long (University of Pennsylvania)
CodeGenerationOptimizationTransformerLarge Language ModelText
π― What it does: This paper investigates the trade-off between watermark strength and speculative sampling efficiency in large language model outputs, proposes a quantifiable watermark strength metric, converts the trade-off into a constrained optimization problem using this metric, derives the Pareto curve, and designs a pseudo-random acceptance mechanism to achieve maximum watermark strength while maintaining the highest sampling efficiency. The effectiveness of this approach is validated through experiments.
IMSE: Intrinsic Mixture of Spectral Experts Fine-tuning for Test-Time Adaptation
Sunghyun Baek (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
CodeDomain AdaptationComputational EfficiencyTransformerMixture of ExpertsImage
π― What it does: The IMSE framework is proposed for Test-Time Adaptation (TTA) and Continuous Test-Time Adaptation (CTTA). It leverages SVD decomposition of the linear layer in ViT, updating only the singular values to achieve parameter-efficient adaptation, and suppresses feature collapse through a diversity maximization loss. In CTTA, domain-aware spectral code retrieval is introduced to rapidly reuse previously adapted parameters from other domains.
In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations
Mohammad Aflah Khan (Max Planck Institute for Software Systems), Abhilasha Ravichander (Max Planck Institute for Software Systems)
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: Investigate and quantify the potential source preferences of large language models (LLMs) across different domains (news, academic papers, e-commerce) through direct and indirect comparative experiments validated on real-world data;
Jiashuo Fan (Duke University), Xiang Cheng (Duke University)
CodeClassificationTransformerImageGraph
π― What it does: Proposed and implemented a two-stage Transformer architecture for semi-supervised context learning (IC-SSL), which adaptively learns representations of unlabeled data and performs classification prediction within the same network.
π― What it does: Proposes a lightweight teacher selection score named GRACE, which leverages the gradient distribution characteristics of the student on teacher-generated data to evaluate the compatibility between teacher and student, thereby predicting the student's performance after distillation without using a validator, teacher logits, or test data.
π― What it does: This paper demonstrates that a frozen-weight Transformer (using only softmax attention) can simulate a class of widely ranging algorithms, including gradient descent, linear regression, ridge regression, etc., through prompts; in two modes: in the task-specific mode, single-head attention can approximate any continuous function f(wxα΅βy); in the prompt-programmable mode, a single two-layer attention module can achieve all these algorithms by altering prompts;
CodeMeta LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningImage
π― What it does: Studied the pure exploration problem, proposing the ICPE framework that uses Transformer to actively collect data and infer correct hypotheses through Meta-RL in contextual scenarios.
π― What it does: Proposes FIM-PP, a foundational inference model for marked time point processes (MTPP), capable of directly estimating the conditional intensity function of event sequences under zero-shot or few-shot settings.
π― What it does: Proposes a fully amortized multi-objective black-box optimization framework TAMO, which utilizes transformers to directly generate the next query based on historical data, without requiring modeling or hyperparameter tuning.
Yepeng Liu (University Of California Santa Barbara), Yuheng Bu (University Of California Santa Barbara)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes an In-Context Watermark (ICW) technique achieved solely through prompt engineering, which can embed detectable watermarks into generated text without accessing the internal decoding process of LLMs;
π― What it does: This paper proposes a Reinforcement Learning-based Functional Token Tuning (RFTT) framework, which internalizes the reasoning behavior of LLMs through learnable functional tokens (e.g., <analyze>, <verify>), and achieves self-improvement via self-generated tree search samples.
Incomplete Data, Complete Dynamics: A Diffusion Approach
Zihan Zhou (Chinese University of Hong Kong Shenzhen), Tianshu Yu (Chinese University of Hong Kong Shenzhen)
CodeRestorationDiffusion modelImagePhysics Related
π― What it does: Propose a diffusion model-based framework that can directly learn and reconstruct complete physical dynamics processes from data with only incomplete observations.
π― What it does: Studies the multi-view multi-label classification problem with both views and labels incomplete, proposing the SCSD framework to achieve joint learning.
π― What it does: Propose the RePB framework to address the issues of score context drift and gradient time bias in dynamic data pruning, achieving efficient training through local window pruning, uniform resampling, and cumulative temporal rescaling.
Incorporating Expert Priors into Bayesian Optimization via Dynamic Mean Decay
Chongqi Qu (Xi'an Jiaotong University), Zhunga Liu (Northwestern Polytechnical University)
CodeOptimizationHyperparameter SearchBenchmark
π― What it does: Introduce expert priors in Bayesian optimization and achieve gradual abandonment of priors through a dynamically decaying mean function, enabling the algorithm to rapidly leverage priors for accelerated search in the early stages, while gradually restoring the robustness of standard BO in later stages.
Inductive Reasoning for Temporal Knowledge Graphs with Emerging Entities
Ze Zhao (Shanghai Jiao Tong University), Chenghu Zhou (Chinese Academy Of Sciences)
CodeRepresentation LearningTransformerLarge Language ModelTextGraph
π― What it does: Proposed the TRANSFIR framework for inductive reasoning in temporal knowledge graphs when dealing with new entities lacking historical interactions;
Inference-Cost-Aware Dynamic Tree Construction for Efficient Inference in Large Language Models
Yinrong Hong (Beihang University), Kai Hu (Beihang University)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Proposes CAST (Cost-Aware Speculative Tree) β a method for accelerating LLM inference by dynamically constructing tree drafts and adaptively pruning based on system costs such as GPU devices, batch sizes, etc.;
Inferring brain plasticity rule under long-term stimulation with structured recurrent dynamics
Zhichao Liang (Southern University of Science and Technology), Quanying Liu (Southern University of Science and Technology)
CodeRecurrent Neural NetworkTime SeriesBiomedical Data
π― What it does: This study proposes the STEER framework to learn brain plasticity rules under long-term stimulation, formalizing long-term plasticity as a potential dynamic system under stimulation conditions.
Inferring the Invisible: Neuro-Symbolic Rule Discovery for Missing Value Imputation
Wendi Ren (Chinese University of Hong Kong), Shuang Li (Chinese University of Hong Kong)
CodeData SynthesisExplainability and InterpretabilityTabular
π― What it does: Propose a neural-symbolic framework that achieves reasoning and imputation of missing entries in incomplete tables through bidirectional feedback loops between rule learning and missing value imputation.
InfLLM-V2: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation
Weilin Zhao (Tsinghua University), Zhiyuan Liu (Tsinghua University)
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: Propose an InfLLM-V2 framework that can switch to sparse attention without additional parameters or significant distribution changes after short-sequence pretraining, achieving long-sequence processing.
Influence-Preserving Proxies for Gradient-Based Data Selection in LLM FineTuning
Sirui Chen (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
CodeOptimizationComputational EfficiencyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Designed the IPROX two-stage framework to generate an influence-preserving proxy model for gradient data selection in large language models.
π― What it does: By constructing an unbiased mutual information estimator, Info Bridge, the efficient estimation of mutual information for any random variables is achieved through the diffusion bridge matching method.
InfoDet: A Dataset for Infographic Element Detection
Jiangning Zhu (BNRist, Tsinghua University), Shixia Liu (BNRist, Tsinghua University)
CodeObject DetectionData SynthesisConvolutional Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityChain-of-Thought
π― What it does: Constructed a large-scale InfoDet dataset and trained an information graph-oriented object detection model using it, further applying it to chart understanding and graphic layout detection in VLM.
π― What it does: Propose the Information Gain Reward Framework (IGPO) in multi-round search agents, achieving dense reinforcement learning supervision through improvements in the probability of the true answer at each step.
Xiaoyuan Cheng (University College London), Zhuo Sun (Shanghai University of Finance and Economics)
CodeRepresentation LearningAuto EncoderContrastive LearningGraphTime Series
π― What it does: Propose a Koopman representation learning framework based on the information bottleneck theory, and design Lagrangian regularization to balance conciseness and expressiveness, achieving stable and interpretable Koopman representations.
π― What it does: Propose a differentiable planning module named VIΒ²N, which integrates value iteration networks with information value assessment for decision-making in partially observable environments
Initialization Schemes for KolmogorovβArnold Networks: An Empirical Study
Spyros Rigas (National and Kapodistrian University of Athens), Yixuan Wang (California Institute of Technology)
CodeHyperparameter SearchPhysics Related
π― What it does: Investigated initialization schemes for piecewise spline KAN and systematically evaluated LeCun, Glorot, and empirical power law methods.
CodeTransformerLarge Language ModelAgentic AIBenchmark
π― What it does: Proposed the InnoGym framework and iBench benchmark for systematic evaluation of AI agents' innovation potential, combining two dimensions: performance improvement and method novelty;
π― What it does: Designed and implemented InnovatorBench, an end-to-end evaluation benchmark for LLM research, along with theι ε₯ ResearchGym platform, to assess AI research agents in real-world, long-latency, distributed experimental environments.
INO-SGD: Addressing Utility Imbalance under Individualized Differential Privacy
Xiao Tian (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
CodeOptimizationImageBiomedical Data
π― What it does: Investigate the issue of efficacy imbalance caused by individualized differential privacy (IDP) and propose a novel individualized noise-ordered stochastic gradient descent (INO-SGD) algorithm to improve the data performance of high-privacy groups while satisfying IDP constraints.
CodeLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelText
π― What it does: Propose a reinforcement learning framework IGPO based on masked diffusion large models, leveraging the model's inpainting capability to inject partial real reasoning snippets during sampling to guide exploration and alleviate the zero-advantage problem; meanwhile, introduce length-aligned supervised fine-tuning and entropy filtering strategies to enhance training stability.
Chongyi Zheng (Princeton University), Benjamin Eysenbach (Princeton University)
CodeFlow-based ModelImageTabular
π― What it does: Pre-train a flow occupancy model (InFOM) incorporating user intent on an offline RL dataset, and achieve efficient fine-tuning in downstream tasks through autoregressive occupancy prediction and implicit general policy improvement.
Interaction-aware Representation Modeling With Co-Occurrence Consistency for Egocentric Hand-Object Parsing
YUEJIAO SU, Lap-Pui Chau (Hong Kong Polytechnic University)
CodeSegmentationTransformerImage
π― What it does: Propose InterFormer for pixel-level segmentation of hand-object interactions in first-person perspective, addressing the challenge of modeling interaction relationships.
Interleaving Reasoning for Better Text-to-Image Generation
Wenxuan Huang (East China Normal University), Shaohui Lin (East China Normal University)
CodeGenerationTransformerLarge Language ModelVision Language ModelAuto EncoderImageTextMultimodalityChain-of-Thought
π― What it does: Constructed an alternating framework for text reasoning and image generation (Interleaving Reasoning Generation, IRG), achieving a multi-round generation process of 'text-image-text-image' by first generating text reasoning, then generating images, reflecting on the generated images, and finally producing improved images.
Internal Evaluation of Density-Based Clusterings with Noise
Anna Beer (Webster Vienna Private University), Claudia Plant (University of Vienna)
CodeBenchmark
π― What it does: This paper proposes a new internal clustering validation metric called DISCO, designed to evaluate density-based clustering results with a particular focus on the correct labeling of noise points.
InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models
Nianchen Deng (Shanghai AI Laboratory), Hongjie Zhang (Shanghai AI Laboratory)
CodeData SynthesisTransformerSupervised Fine-TuningVision Language ModelImageMultimodalityPoint CloudBenchmark
π― What it does: Proposed the InternSpatial dataset and InternSpatial-Bench evaluation benchmark, constructed 12M QA pairs, supporting 19 instruction formats, covering single-view and multi-view spatial reasoning, and fine-tuned Vision-Language Models (VLMs) using these data, significantly enhancing their spatial reasoning capabilities.
π― What it does: Proposes Interp3D, an untrained, correspondence-aware 3D texture deformation framework that achieves smooth transitions from source to target 3D models.
π― What it does: Proposed two conditioning strategies without training during inference (Interpolate-Integrate and Replacement Guidance), achieving 3D molecule generation using the pre-trained SemlaFlow flow matching model for bioequivalent ligand design.
Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders
Yizhou Wang (Shanghai Artificial Intelligence Laboratory), Botian Shi (Shanghai Artificial Intelligence Laboratory)
CodeComputational EfficiencyLarge Language ModelMultimodality
π― What it does: Investigate the redundancy in multi-encoder multimodal large language models, systematically evaluate and quantify the independent contributions of each visual encoder;
Invisible Safety Threat: Malicious Finetuning for LLM via Steganography
Guangnian Wan (National University of Singapore), Xinchao Wang (National University of Singapore)
CodeAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Investigated and implemented a malicious fine-tuning method enabling LLMs to learn and use invisible character steganography techniques, hiding and generating malicious content in dialogues with an innocuous appearance; simultaneously verified the feasibility of this attack on GPT-4.1 and multiple open-source models.
IR-Agent: Expert-Inspired LLM Agents for Structure Elucidation from Infrared Spectra
Heewoong Noh (Korea Advanced Institute Of Science And Technology), Chanyoung Park (Korea Advanced Institute Of Science And Technology)
CodeDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextTabularRetrieval-Augmented Generation
π― What it does: Propose IR-Agent, a multi-agent large language model framework that mimics expert workflows in infrared spectroscopy analysis. It automatically infers molecular SMILES structures from infrared spectra through three specialized agent roles: table parsing, retrieval, and structural reasoning.
CodeTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: This paper systematically evaluates whether in-context learning (ICL) in autoregressive large language models truly possesses learning capabilities through a theoretical framework and large-scale experiments.
Is the Reversal Curse a Binding Problem? Uncovering Limitations of Transformers from a Basic Generalization Failure
Boshi Wang (Ohio State University), Huan Sun (Ohio State University)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningText
π― What it does: Investigate and address the 'Reversal Curse' that arises when large language models learn reversible factual associations, propose its root causes from the perspective of cognitive binding issues, and validate through experiments.
It's All Just Vectorization: einx, a Universal Notation for Tensor Operations
Florian Fervers (Fraunhofer IOSB), Michael Arens (Fraunhofer IOSB)
Code
π― What it does: Proposed a generic tensor operation symbol representation called einx, unifying all tensor operations into a vectorized form of low-level operations and implementing it in Python.
Iterative Training of Physics-Informed Neural Networks with Fourier-enhanced Features
Yulun Wu (KTH Royal Institute of Technology), Matthieu Barreau (KTH Royal Institute of Technology)
CodeOptimizationComputational EfficiencyPhysics Related
π― What it does: Propose an iteratively trained physics-informed neural network (IFeF-PINN), which expands the latent space by incorporating random Fourier features (RFF) on the hidden layer features, and significantly reduces the spectral bias of PINNs through a two-stage hierarchical optimization (upper layer generates bases, lower layer performs linear regression), achieving high-accuracy approximation for high-frequency and multi-scale partial differential equations.
IV-Bench: A Benchmark for Image-Grounded Video Perception and Reasoning in Multimodal LLMs
David Ma, Xiaojie Jin
CodeRepresentation LearningLarge Language ModelVision Language ModelImageVideoTextBenchmark
π― What it does: Propose the IV-Bench benchmark, containing 966 videos, 2560 external images, and text queries, covering 13 categories of perception and reasoning tasks, to evaluate the ability of multi-modal LLMs in joint image-video understanding.
IVC-Prune: Revealing the Implicit Visual Coordinates in LVLMs for Vision Token Pruning
Zhichao Sun (Wuhan University), Yongchao Xu (Wuhan University)
CodeComputational EfficiencyVision Language ModelMultimodality
π― What it does: Propose a training-agnostic visual token pruning method called IVC-Prune, which retains implicit visual coordinate (IVC) tokens and semantic foreground tokens, reducing the number of visual tokens in LVLM by approximately 50% without significantly affecting performance;
Jailbreaking the Matrix: Nullspace Steering for Controlled Model Subversion
Vishal Pramanik (University of Florida), Sumit Kumar Jha (University of Florida)
CodeAdversarial AttackTransformerLarge Language ModelTextBenchmark
π― What it does: Propose Head-Masked Nullspace Steering (HMNS) on large language models, achieving efficient jailbreaking by identifying and suppressing key attention heads while injecting orthogonal perturbations.
JailbreakLoRA: Your Downloaded LoRA from Sharing Platforms might be Unsafe
Fanjunduo Wei (Northeastern University), Bo Han (Hong Kong Baptist University)
CodeAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
π― What it does: Designed a multi-task jailbreak LoRA method that balances downstream task performance with attack success rate, leveraging uncertainty weighting, gradient conflict projection, and trigger prefix injection to enhance hallucination attack effectiveness during inference.
JailNewsBench: Multi-Lingual and Regional Benchmark for Fake News Generation under Jailbreak Attacks
Masahiro Kaneko (Mohamed bin Zayed University of Artificial Intelligence), Timothy Baldwin (Mohamed bin Zayed University of Artificial Intelligence)
CodeAnomaly DetectionSafty and PrivacyAdversarial AttackLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Proposed JailNewsBench, a benchmark for evaluating fake news generation and jailbreaking attacks, covering 34 regions, 22 languages, and approximately 300,000 instances.
JALMBench: Benchmarking Jailbreak Vulnerabilities in Audio Language Models
Zifan Peng (Hong Kong University of Science and Technology), Xinyi Huang (Nanjing University of Aeronautics and Astronautics)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringMultimodalityBenchmarkAudio
π― What it does: Proposed the JALMBench benchmark for systematic evaluation of the security of large audio language models (LALMs) against jailbreak attacks.
JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence
Qiushi Sun (University of Hong Kong), Fei Yuan (Shanghai AI Laboratory)
CodeGenerationData SynthesisAI Code AssistantTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBenchmark
π― What it does: Propose the JANUSCODER series of models, constructing a unified visual-program interface to achieve the capability of generating code from text instructions, visual inputs, or a combination of both; simultaneously releasing a complete data synthesis tool and the JANUSCODE-800K multimodal code corpus; competing with large commercial models (e.g., GPT-4o) on multiple text, visual, and cross-modal coding benchmarks and outperforming them in multiple scenarios.
π― What it does: Propose VeloxSeg, a lightweight 3D medical image segmentation framework that achieves efficient segmentation through a dual-stream CNN-Transformer architecture combined with Johnson-Lindenstrauss-guided convolution and Paired Window Attention.
Joint DistributionβInformed Shapley Values for Sparse Counterfactual Explanations
Lei You (Technical University of Denmark), Lele Cao (Microsoft)
CodeOptimizationExplainability and InterpretabilityTabular
π― What it does: This paper proposes the COLA framework, which generates p-SHAP feature importance by performing optimal transport matching between the joint distributions of real instances and counterfactual instances, thereby significantly reducing the required feature modifications while maintaining counterfactual effectiveness.
π― What it does: This paper proposes the DATAMASK framework, which employs a mask learning method based on policy gradients to jointly select quality and diversity from large-scale pre-trained data, ultimately extracting a 1.5 trillion-token subset called FineWeb-Mask from the 15 trillion-token FineWeb.
π― What it does: The study proposes Light-Geometry Interaction (LGI) maps to jointly generate shadows and relighting in a single view, achieving physically consistent shadows and lighting effects when inserting objects.
Journey to the Centre of Cluster: Harnessing Interior Nodes for A/B Testing under Network Interference
Qianyi Chen (Tsinghua University), Yong Wang (Tencent Inc)
CodeGraph Neural NetworkGraph
π― What it does: In social network A/B testing, a mean of internal nodes (MII) estimator is proposed, which is corrected by a trained counterfactual predictor (AMII), significantly reducing bias while maintaining low variance.
JUDO: A Juxtaposed Domain-Oriented Multimodal Reasoner for Industrial Anomaly QA
Hyunju Kang (Sungkyunkwan University), Hogun Park (Sungkyunkwan University)
CodeAnomaly DetectionExplainability and InterpretabilityTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: Propose the JUDO framework, which combines visual contrast, domain knowledge injection, and reinforcement learning through a three-stage training process to improve the accuracy and interpretability of industrial anomaly question-answering.
JULI: Jailbreak Large Language Models by Self-Introspection
Jesson Wang (University of Southern California), David Wagner (University of California, Berkeley)
CodeAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: By leveraging a lightweight BiasNet plugin, the model output is biased using only the top5 token log probabilities returned by the target LLM, enabling unlocking attacks on LLMs with open weights and API calls.
Just Do It!? Computer-Use Agents Exhibit Blind Goal-Directedness
Erfan Shayegani (Microsoft Research AI Frontiers), Vibhav Vineet (Microsoft Research AI Frontiers)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringWorld ModelMultimodalityBenchmark
π― What it does: Investigated the Blind Goal-Directedness (BGD) phenomenon in Computer-Use Agents (CUA) and constructed the BLIND-ACT benchmark, comprising 90 tasks, to evaluate the blind goal-pursuing behavior of CUA under different contexts, hypothesis uncertainties, and contradictory/implementable goals.
K-Prism: A Knowledge-Guided and Prompt Integrated Universal Medical Image Segmentation Model
Bangwei Guo (Rutgers University), Dimitris N. Metaxas
CodeSegmentationTransformerPrompt EngineeringMixture of ExpertsMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
π― What it does: Proposes a unified medical image segmentation framework K-Prism, capable of seamlessly switching between three knowledge paradigms (semantic prior, contextual examples, interactive feedback) and jointly training within the same model.
K-Sort Eval: Efficient Preference Evaluation for Visual Generation via Corrected VLM-as-a-Judge
Zhikai Li (Institute of Automation, Chinese Academy of Sciences), Kurt Keutzer (University of California, Berkeley)
CodeGenerationPrompt EngineeringVision Language ModelImageVideoTextMultimodalityBenchmark
π― What it does: Propose a scalable evaluation framework K-Sort Eval based on vision-language models (VLM), which rapidly estimates human preference rankings for generative models using Bayesian calibration and dynamic matching.
Jin Lee (University of California, Santa Barbara), Zheng Zhang (University of California, Santa Barbara)
CodeExplainability and InterpretabilityComputational EfficiencyPhysics Related
π― What it does: Proposed and implemented the Kolmogorov-Arnold Neural Operator (KANO), a neural operator that simultaneously sparsely represents and is interpretable in both spatial and frequency domains;
π― What it does: Propose the KAVA framework, which uses compressed KV-cache for self-distillation to train the implicit reasoning student to generate only continuous latent thoughts instead of lengthy CoT.
Keep the Best, Forget the Rest: Reliable Alignment with Order-Aware Preference Optimization
Jiahui Zhu (Washington State University), Honghao Wei (Washington State University)
CodeReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
π― What it does: Propose an improved DPO algorithm called RAPPO, which reduces reference strategy bias by filtering low-alignment samples, thereby enhancing the alignment of language models.
KeepLoRA: Continual Learning with Residual Gradient Adaptation
Mao-Lin Luo (Southeast University), Tong Wei (Southeast University)
CodeRepresentation LearningMeta LearningTransformerVision Language ModelMultimodalityBenchmark
π― What it does: Propose a LoRA-based continual learning method called KeepLoRA, which balances pre-trained knowledge, previously learned tasks, and new tasks by projecting gradients into the residual subspace for updates;
Kimi-Dev: Agentless Training as Skill Prior for SWE-agents
Zonghan Yang (Moonshot AI), Tianyu Liu (Shanghai Qi Zhi Institute)
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
π― What it does: Propose a code LLM called Kimi-Dev based on Agentless training. First, they develop bug repair and test writing capabilities through mid-training, cold start, reinforcement learning, and self-play. Then, they fine-tune with a small amount of SWE-Agent trajectories to achieve high-performance SWE-Agent.
Know When to Abstain: Optimal Selective Classification with Likelihood Ratios
Alvin Heng (National University of Singapore), Harold Soh (National University of Singapore)
CodeClassificationDomain AdaptationImageText
π― What it does: This paper proposes to design a selector function using the Neyman-Pearson lemma, unifying and improving selective classification methods, particularly for covariate shift scenarios;
KnowGuard: Knowledge-Driven Abstention for Multi-Round Clinical Reasoning
Xilin Dang (Chinese University of Hong Kong), Pheng-Ann Heng
CodeTransformerLarge Language ModelMultimodalityGraphBiomedical DataBenchmarkRetrieval-Augmented Generation
π― What it does: Proposes a knowledge graph-based proactive exploration framework (KnowGuard) for deciding when to stop answering (abstention) and collect evidence in multi-turn clinical dialogues;
Knowledge Exchange with Confidence: Cost-Effective LLM Integration for Reliable and Efficient Visual Question Answering
Mahsa Mozaffari (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
CodeKnowledge DistillationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: Propose a confidence-based collaborative framework named Uni-VQA that integrates large language models (LLMs) with task-specific VQA models to achieve knowledge exchange and improve visual question answering performance.
Knowledge Fusion of Large Language Models via Modular SkillPacks
Guodong DU (Harbin Institute of Technology), Jing Li (Harbin Institute of Technology)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical DataBenchmarkFinance Related
π― What it does: Propose the GraftLLM framework to achieve cross-model capability transfer, encoding source model knowledge into a lightweight SkillPack, supporting the fusion and continuous learning of heterogeneous LLMs.
Knowledge Reasoning Language Model: Unifying Knowledge and Language for Inductive Knowledge Graph Reasoning
Xingrui Zhuo (Hefei University of Technology), Xindong Wu (Hefei University of Technology)
CodeGraph Neural NetworkTransformerLarge Language ModelGraphBenchmark
π― What it does: Integrate LLM knowledge with KG context to realize KRLM, addressing the knowledge distortion problem in open-domain induced knowledge graph reasoning
KnowledgeSmith: Uncovering Knowledge Updating in LLMs with Model Editing and Unlearning
Yinyi Luo (Carnegie Mellon University), Jindong Wang (Carnegie Mellon University)
CodeLarge Language ModelGraphBenchmark
π― What it does: This paper proposes the KnowledgeSmith framework, which unifies the analysis of knowledge editing and machine forgetting (unlearning) in LLMs, and constructs a large-scale structured benchmark through automated knowledge graph data generation;
KnowProxy: Adapting Large Language Models by Knowledge-guided Proxy
Gukhyeon Lee (Korea University), SangKeun Lee (Korea University)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Propose KNOWPROXY, an agent method for fine-tuning large language models using textual knowledge rather than probability distributions;
KRAMABENCH: A Benchmark for AI Systems on Data-to-Insight Pipelines over Data Lakes
Eugenie Lai (MIT CSAIL), Tim Kraska (MIT CSAIL)
CodeData-Centric LearningTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Evaluate the capability of large language models (LLMs) to design and execute complete data-to-insight pipelines on real-world data lakes, constructing a benchmark called KRAMABENCH with 104 multi-domain tasks.
KVComm: Enabling Efficient LLM Communication through Selective KV Sharing
Xiangyu Shi (KTH Royal Institute of Technology), Dejan Kostic (KTH Royal Institute of Technology)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose the KVComm framework, which achieves efficient communication between LLMs by sharing key-value (KV) pairs in critical layers only.
Zonglin Di (University of California, Santa Cruz), Yang Liu (University of California, Santa Cruz)
CodeSafty and PrivacyComputational EfficiencyImageTextMultimodality
π― What it does: Proposed a machine unlearning method called UGradSL, which combines gradient ascent with label smoothing (negative label smoothing), enabling rapid elimination of the memory of specific data while maintaining model accuracy and enhancing local differential privacy;
Landscape of Thoughts: Visualizing the Reasoning Process of Large Language Models
Zhanke Zhou (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: Propose LoT (Landscape of Thoughts), a visualization tool for 2D visualization of step-by-step reasoning trajectories of large language models (LLMs), and develop a lightweight validator based on this to improve reasoning accuracy.
π― What it does: Proposes the Label-Aware Noise Elimination (LANE) method, which assigns dynamic weights to each sample based on semantic similarity between labels and training dynamics to mitigate the impact of noisy labels.
Language Agents for Hypothesis-driven Clinical Decision Making with Reinforcement Learning
David Bani-Harouni (Technical University of Munich), Matthias Keicher (Technical University of Munich)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextMultimodalityTabularBiomedical DataElectronic Health Records
π― What it does: Built and trained a dual-agent language model (LA-CDM) composed of a hypothesis generator and a decision maker, achieving iterative narrowing of diagnostic hypotheses through repeated requests and explanations of diagnostic tests, ultimately providing a diagnosis.