π― What it does: A black-box adversarial method called TRAP is proposed, which injects semantic information into the CLIP latent space using diffusion models, thereby inducing multimodal proxy systems to prefer altered target images.
Treatment Effect Estimation for Optimal Decision-Making
Dennis Frauen (Ludwig Maximilian University of Munich), Stefan Feuerriegel (Ludwig Maximilian University of Munich)
CodeOptimizationMeta LearningTabular
π― What it does: This study investigates the optimality of a two-stage CATE meta-learner under constrained function classes and proposes a method for redirecting CATE to enhance the performance of threshold decisions.
Tree Ensemble Explainability through the Hoeffding Functional Decomposition and TreeHFD Algorithm
Clement Benard
CodeExplainability and InterpretabilityTabular
π― What it does: This paper proposes the TreeHFD algorithm, which estimates the Hoeffding Functional Decomposition (HFD) of tree ensembles using data samples, achieving an interpretable breakdown of tree models.
Tree of Preferences for Diversified Recommendation
Hanyang Yuan (Zhejiang University), Mingli Song (Zhejiang University)
CodeRecommendation SystemGraph Neural NetworkTransformerLarge Language ModelText
π― What it does: This paper proposes a diversified recommendation framework based on large language models (LLM) called ToP-Rec. It constructs a Tree of Preferences using user attributes and interaction records to infer unexplored user interests through LLM systems, generating composite interactions based on these preferences to enhance training data, ultimately training a recommendation model that balances diversity and relevance.
π― What it does: This paper proposes Tree-Sliced Entropy Partial Transport (PartialTSW), a distance that transforms partial transport into a balanced optimal transport in tree metric spaces and achieves efficient computation through tree slicing.
π― What it does: This paper proposes TreeSplat, which utilizes a mergeable hierarchical tree structure to collaboratively model the motion of dynamic Gaussian Splatting.
TreeSynth: Synthesizing Diverse Data from Scratch via Tree-Guided Subspace Partitioning
Sheng Wang (University of Hong Kong), Chuan Wu (University of Hong Kong)
CodeGenerationData SynthesisLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes a tree-guided subspace partitioning method (TREESYNTH) that recursively divides the task data space into mutually exclusive and complementary atomic subspaces from a global perspective. Samples are then synthesized using LLM within each subspace, and finally aggregated into a high-diversity, fully covered dataset.
π― What it does: Proposes the TRIDENT tri-modal molecular representation learning framework, integrating SMILES, textual descriptions, and hierarchical classification annotations for global and local alignment.
Tropical Attention: Neural Algorithmic Reasoning for Combinatorial Algorithms
Baran Hashemi (Data Science Lab Technical University of Munich), Ruriko Yoshida (Naval Postgraduate School)
CodeOptimizationTransformerSequentialBenchmark
π― What it does: This paper proposes the Tropical Attention mechanism, which replaces the traditional softmax pointwise attention, reasoning in the tropical projective space, and proves its approximation to maximum sum dynamic programming and closure.
CodeTransformerPrompt EngineeringVision Language ModelVideoMultimodality
π― What it does: An automated method called TROVE has been developed to discover static feature biases that lead to errors in temporal visual-language models, and to improve model performance on real tasks through prompt learning.
True Zero-Shot Inference of Dynamical Systems Preserving Long-Term Statistics
Christoph JΓΌrgen Hemmer (Heidelberg University), Daniel Durstewitz (Heidelberg University)
CodeRecurrent Neural NetworkMixture of ExpertsTime SeriesSequentialMagnetic Resonance Imaging
π― What it does: Developed the DynaMix model, achieving the reconstruction of long-term statistical properties of any dynamic system under zero-shot conditions.
π― What it does: This paper proposes a non-adversarial inverse reinforcement learning framework TRRO and implements a practical algorithm PIRO to stably learn the expert reward function and generate high-quality imitation policies.
π― What it does: A testing-time adaptive method called TRUST for Visual State Space Models (VMamba) is proposed, which generates multiple causal perspectives by traversing different directional scanning arrangements and updates model parameters using pseudo-labels, ultimately performing weight averaging to enhance cross-domain robustness.
Truth over Tricks: Measuring and Mitigating Shortcut Learning in Misinformation Detection
Herun Wan (Xi'an Jiaotong University), Zhixiong Su (Xi'an Jiaotong University)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Proposes the TRUTHOVERTRICKS evaluation framework, which systematically assesses the dependence of rumor detection models on internal and external shortcuts; constructs two new factual rumor datasets, NQ-Misinfo and Streaming-Misinfo; and introduces the LLM-assisted data augmentation framework SMF to mitigate the model's reliance on shortcuts.
π― What it does: A two-stage multi-objective fine-tuning framework, TS-MOF, is proposed to address the class imbalance problem in long-tail recognition. First, high-quality general features are obtained through pre-training; then, the feature network is frozen, and only the multi-task classification head undergoes multi-objective fine-tuning.
π― What it does: This paper proposes TS-RAG, a retrieval-augmented generation (RAG) framework that utilizes a pre-trained time series foundation model (TSFM) and retrieved similar time series segments to achieve high-accuracy predictions in a zero-shot setting.
TSENOR: Highly-Efficient Algorithm for Finding Transposable N:M Sparse Masks
Xiang Meng (Massachusetts Institute of Technology), Rahul Mazumder (Massachusetts Institute of Technology)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: An efficient algorithm TSENOR is proposed for generating binary masks that satisfy the transposed N:M sparse constraints, enabling forward and backward acceleration on large-scale language models.
π― What it does: A testing time scaling framework for visual autoregressive (VAR) models, TTS-VAR, is proposed to enhance the quality of generated images through adaptive batching, coarse-scale clustering diversity search, and fine-scale potential resampling.
π― What it does: An approximate, distributed, accelerated sketch-and-project algorithm ADASAP is proposed for large-scale Gaussian process inference.
π― What it does: A sequence recommendation model based on time-varying convolutional filters, TV-Rec, is proposed, which captures the temporal changes in user behavior sequences using node transformation filters from graph signal processing, completely omitting the self-attention module.
Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning
Chaofan Lin (Tsinghua University), Mingyu Gao (Tsinghua University)
CodeTransformerLarge Language ModelText
π― What it does: Designed and implemented the Twilight framework, which incorporates an adaptive budget mechanism into the existing top-k sparse attention models, using topp sampling to dynamically prune key tokens in the KV cache.
π― What it does: This paper proposes the U-REPA framework, which aligns the hidden states of U-Net with the Vision Transformer encoder to accelerate the training of diffusion models.
UFO-RL: Uncertainty-Focused Optimization for Efficient Reinforcement Learning Data Selection
Yang Zhao (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
CodeOptimizationReinforcement LearningText
π― What it does: This paper proposes the UFO-RL framework, which utilizes a confidence assessment method based on single forward inference to efficiently filter RL training data for LLMs, thereby focusing learning on the model's Zone of Proximal Development (ZPD).
UFO: A Unified Approach to Fine-grained Visual Perception via Open-ended Language Interface
Hao Tang (Peking University), Liwei Wang (Peking University)
CodeObject DetectionSegmentationTransformerVision Language ModelImageTextMultimodality
π― What it does: Proposes the UFO framework, which unifies the mapping of fine-grained visual tasks such as detection and segmentation to an open language interface, achieving a unified model without a task-specific decoder;
π― What it does: A unified post-training framework UFT is proposed, which integrates Supervised Fine-Tuning (SFT) and Reinforcement Learning Fine-Tuning (RFT) into a continuous process, guiding the model to smoothly transition between exploration and memory using hints.
UI-Genie: A Self-Improving Approach for Iteratively Boosting MLLM-based Mobile GUI Agents
Han Xiao (CUHK MMLab), Hongsheng Li (CUHK MMLab)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
π― What it does: A self-improvement framework called UI-Genie is proposed for training and enhancing mobile GUI agents based on multimodal large language models (MLLMs), focusing on addressing the challenges of trajectory result validation and the lack of high-quality training data.
Ultrametric Cluster Hierarchies: I Want βem All!
Andrew Draganov (Aarhus University), Ira Assent (Aarhus University)
CodeOptimizationTabular
π― What it does: It is proposed that center-based clustering (k-means, k-median, k-center) can obtain optimal solutions for all k in O(n) time on any ultrametric (represented by LCA-tree), and these solutions form a hierarchy; based on this, the SHiP framework is constructed, which can quickly generate various clustering hierarchies and partitions after a single ultrametric fitting.
UMoE: Unifying Attention and FFN with Shared Experts
Yuanhang Yang (Institute of Science Tokyo), Jing Li (Hong Kong Polytechnic University)
CodeTransformerMixture of ExpertsText
π― What it does: This paper proposes a unified sparse expert mixture model UMoE, which integrates the attention layer and feedforward network layer of the Transformer into a shareable FFN-style expert through a pre-mixing attention reformulation, achieving improved model performance while maintaining the same parameter count.
un$^2$CLIP: Improving CLIP's Visual Detail Capturing Ability via Inverting unCLIP
Yinqi Li (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)
CodeSegmentationGenerationTransformerVision Language ModelDiffusion modelContrastive LearningImageMultimodality
π― What it does: Based on the pre-trained CLIP image encoder, reverse fine-tuning is achieved using the unCLIP generative model to enhance its ability to capture visual details.
π― What it does: A unified multi-modal, multi-task object re-identification (MΒ³T-ReID) framework is proposed to address the problem of simultaneous retrieval across modalities and multi-modalities for different categories (people, vehicles).
π― What it does: The ACUS framework and unbiased sliced Wasserstein RBF kernel are proposed to address exposure bias and cross-modal alignment issues in audio caption generation.
π― What it does: This paper proposes a semi-supervised confidence distribution learning-based UKG completion framework called ssCDL, which can simultaneously perform confidence prediction and link prediction; by transforming single confidence into a confidence distribution, it enhances the supervisory information of minority or unseen confidence levels, and utilizes a Pseudo Confidence Distribution Generator (PCDG) and meta-learning strategies for self-supervised training of unlabeled triples to improve embedding learning quality.
Uncertainty Estimation by Flexible Evidential Deep Learning
Taeseong Yoon (Korea Advanced Institute of Science and Technology), Heeyoung Kim (Korea Advanced Institute of Science and Technology)
CodeClassificationAnomaly DetectionImage
π― What it does: A flexible evidence deep learning (F-EDL) framework is proposed, using a flexible Dirichlet distribution to infer the uncertainty of classification tasks in a single forward pass.
π― What it does: This paper proposes a method for uncertainty quantification in physics-informed neural networks (PINNs) based on extended fiducial inference (EFI), utilizing a narrow-neck supernetwork to learn PINN parameters and achieve reliable confidence set construction through random error interpolation.
π― What it does: A 3D molecular diffusion model framework guided by multi-objective reinforcement learning based on uncertainty perception is proposed to generate molecules that meet multiple drug-related properties.
Uncertainty-aware Preference Alignment for Diffusion Policies
Runqing Miao (Chinese University of Hong Kong), Guiliang Liu (Chinese University of Hong Kong)
CodeOptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelSequential
π― What it does: This paper proposes an uncertainty-aware preference alignment algorithm for diffusion strategies, called Diff-UAPA, which optimizes directly on the policy without the need to learn a reward function.
Uncertainty-Based Smooth Policy Regularisation for Reinforcement Learning with Few Demonstrations
Yujie Zhu (University of Warwick), Giovanni Montana (University of Warwick)
CodeReinforcement Learning
π― What it does: The SPReD framework is proposed, which achieves smooth policy regularization of demonstration behaviors by using ensemble Q-value distributions and dynamically calculates uncertainty weights at each step to determine imitation strength.
Uncertainty-Informed Meta Pseudo Labeling for Surrogate Modeling with Limited Labeled Data
Xingyu Ren (Zhejiang University), Dong Ni (Zhejiang University)
CodeDomain AdaptationOptimizationComputational EfficiencyKnowledge DistillationMeta LearningSupervised Fine-TuningTabularTime SeriesPhysics Related
π― What it does: A meta pseudo-label learning framework based on uncertainty, UMPL, is proposed to enhance the surrogate modeling performance of physical systems under limited labeled data.
π― What it does: A privilege learning framework based on uncertainty perception, USPL, is proposed, which utilizes an observation encoder to estimate the uncertainty of missing information and inputs it into the privileged policy, achieving alignment between the deployment policy and the privileged policy behavior.
UnCLe: Towards Scalable Dynamic Causal Discovery in Non-linear Temporal Systems
Tingzhu Bi (Peking University), Ping Wang (Peking University)
CodeAnomaly DetectionOptimizationComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkAuto EncoderTime SeriesFinance Related
π― What it does: The UnCLe method is proposed, which uses parameter-shared TCN Uncoupler/Recoupler to semantically decompose time series and achieves scalable dynamic causal graph learning through autoregressive dependency matrices and temporal perturbations.
π― What it does: The study investigates the bottlenecks of autoregressive image generation models in visual understanding and proposes a self-guided training framework ST-AR to enhance the model's learning of high-level visual semantics.
π― What it does: This study investigates the effectiveness of message passing mechanisms in heterogeneous graphs and proposes the CMGNN model, which utilizes a Compatibility Matrix (CM) to enhance node distinguishability.
Understanding and Improving Adversarial Robustness of Neural Probabilistic Circuits
Weixin Chen (University of Illinois Urbana-Champaign), Han Zhao (University of Illinois Urbana-Champaign)
CodeExplainability and InterpretabilityAdversarial AttackImage
π― What it does: A robust neural probabilistic circuit (RNPC) is proposed, enhancing adversarial robustness through category-level ensemble in the concept bottleneck model.
Understanding and Improving Fast Adversarial Training against $l_0$ Bounded Perturbations
Xuyang Zhong (City University of Hong Kong), Chen Liu (City University of Hong Kong)
CodeOptimizationAdversarial AttackImage
π― What it does: This paper studies fast adversarial training against l0 sparse perturbations, analyzing and addressing the catastrophic overfitting (CO) problem caused by one-step attacks.
Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference
Jiayi Yuan (Rice University), Zirui Liu (University of Minnesota Twin Cities)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A systematic study on the impact of floating-point precision (BF16, FP16, FP32) and runtime configurations (GPU type, number, batch size) on the reproducibility of large language model inference, and the proposal of the LayerCast mixed precision inference scheme.
Understanding and Rectifying Safety Perception Distortion in VLMs
Xiaohan Zou (Pennsylvania State University), Lu Lin (Pennsylvania State University)
CodeSafty and PrivacyVision Language ModelImageTextMultimodality
π― What it does: This study investigates the failure mechanisms of visual-language models in safety alignment and proposes a method for activation offset correction during inference called ShiftDC.
π― What it does: Explores the Bayesian perspective of prompt tuning and meta-learning, and empirically validates the effectiveness of soft prefixes on a simplified coin flipping sequence.
π― What it does: We propose PoET-2, a multimodal retrieval-augmented protein language model that can accurately predict the effects of mutations and generate new protein sequences under zero-shot and few-shot supervision.
Understanding the Evolution of the Neural Tangent Kernel at the Edge of Stability
Kaiqi Jiang (Princeton University), Yuanzhi Li (Carnegie Mellon University)
CodeOptimizationTransformerImage
π― What it does: This study investigates the evolution of the neural tangent kernel (NTK) feature vectors of neural networks under the edge of stability (EoS) during gradient descent, as well as their alignment behavior with the target vectors over time.
Understanding while Exploring: Semantics-driven Active Mapping
Liyan Chen (Stevens Institute of Technology), Philippos Mordohai (Stevens Institute of Technology)
CodeSegmentationRobotic IntelligenceGaussian SplattingSimultaneous Localization and MappingPoint Cloud
π― What it does: This paper proposes an active semantic mapping framework called ActiveSGM based on 3D Gaussian Splatting, which can actively select the most informative viewpoints and generate high-quality geometric and semantic maps in unknown environments.
Kaiyang Li (University of Connecticut), Shihao Ji (University of Connecticut)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A unified framework called Uni-LoRA is proposed, which maps all LoRA variants to projections in a low-dimensional subspace, allowing LLM fine-tuning with the training of just one vector.
Uni-MuMER: Unified Multi-Task Fine-Tuning of Vision-Language Model for Handwritten Mathematical Expression Recognition
Yu Li (Peking University), Liangcai Gao (Peking University)
CodeRecognitionTransformerSupervised Fine-TuningVision Language ModelImageMultimodalityChain-of-Thought
π― What it does: A unified multi-task fine-tuning framework called Uni-MuMER has been constructed to enhance the performance of VLM in the handwritten mathematical expression recognition (HMER) task.
Uni-RL: Unifying Online and Offline RL via Implicit Value Regularization
Haoran Xu (University of Texas at Austin), Amy Zhang (University of Texas at Austin)
CodeReinforcement Learning
π― What it does: This paper proposes Uni-RL, a unified offline/online/offline-to-online reinforcement learning framework that normalizes different learning scenarios using implicit value regularization.
CodeGenerationDrug DiscoveryConvolutional Neural NetworkScore-based ModelMultimodalityBiomedical Data
π― What it does: A unified atomic-level molecular generation framework called FuncBind is proposed, which can generate small molecules, cyclic peptides, and antibody CDRs under the condition of a given target protein structure.
Unifying and Enhancing Graph Transformers via a Hierarchical Mask Framework
Yujie Xing (Beijing University of Posts and Telecommunications), Chuan Shi (Beijing University of Posts and Telecommunications)
CodeClassificationGraph Neural NetworkTransformerMixture of ExpertsGraph
π― What it does: A unified hierarchical masking framework is proposed, and the M3Dphormer Graph Transformer is designed to efficiently model multi-level interactions in graphs through multi-level masks (local, clustering, global) and a two-layer expert routing.
Unifying Attention Heads and Task Vectors via Hidden State Geometry in In-Context Learning
Haolin Yang, Naoya Inoue
CodeClassificationGenerationTransformerLarge Language ModelText
π― What it does: A unified framework based on the geometry of hidden states is constructed to explain the roles of attention heads and task vectors in ICL.
π― What it does: A unified image and video lighting reshaping framework called UniLumos is proposed, which achieves physically feasible lighting control while maintaining scene geometry and temporal consistency.
π― What it does: A unified motion framework called UniMotion is proposed, capable of simultaneously handling motion simulation, trajectory prediction, and self-driving planning tasks.
π― What it does: A unified multimodal segmentation framework, UniMRSeg, is proposed, achieving robust segmentation of missing modalities through hierarchical self-supervised compensation.
π― What it does: A multi-site dataset at the UniProt level, UniSite-DS, has been constructed, and an end-to-end Ligand binding site detection framework, UniSite-1D/3D, has been proposed, along with the introduction of an IoU-based AP evaluation metric.
π― What it does: A unified neural solver called UniteFormer is proposed, capable of handling vehicle routing problems (VRP) with only node input, only edge input, and a mixture of node and edge inputs within the same model.
UniTok: a Unified Tokenizer for Visual Generation and Understanding
Chuofan Ma (University of Hong Kong), XIAOJUAN QI
CodeRecognitionGenerationTransformerVision Language ModelAuto EncoderContrastive LearningImageMultimodality
π― What it does: A unified visual tokenizer, UniTok, is proposed, which supports both visual generation and understanding tasks with the same tokenizer.
UniTraj: Learning a Universal Trajectory Foundation Model from Billion-Scale Worldwide Traces
Yuanshao Zhu (Southern University of Science and Technology), Yuxuan Liang (Hong Kong University of Science and Technology)
CodeClassificationRepresentation LearningTransformerMultimodalityTime Series
π― What it does: A universal trajectory foundation model called UniTraj is proposed, along with the construction of a global trajectory dataset named WorldTrace, covering 70 countries.
π― What it does: Proposes Universal Few-shot Control (UFC), achieving fine control over arbitrary spatial conditions with only a small amount of labeled data on pre-trained text-to-image diffusion models.
Universally Invariant Learning in Equivariant GNNs
Jiacheng Cen (Renmin University of China), Wenbing Huang (Renmin University of China)
CodeGraph Neural NetworkGraphPhysics Related
π― What it does: A universal dynamic method Uni-EGNN is proposed, achieving the integrity of graph neural networks invariant to alignment transformations (capable of approximating any continuous equivariant function).
UniZyme: A Unified Protein Cleavage Site Predictor Enhanced with Enzyme Active-Site Knowledge
Chenao Li (Hong Kong University of Science and Technology), Enyan Dai (Hong Kong University of Science and Technology)
CodeProtein Structure PredictionGraph Neural NetworkTransformerSupervised Fine-TuningBiomedical Data
π― What it does: A unified protease cleavage site prediction model, UniZyme, is proposed, which can generalize across different enzymes and improve prediction accuracy by utilizing knowledge of enzyme active sites.
Unlabeled Data Improves Fine-Grained Image Zero-shot Classification with Multimodal LLMs
Yunqi Hong (University of California), Cho-Jui Hsieh (University of California)
CodeClassificationRecognitionLarge Language ModelPrompt EngineeringImageMultimodality
π― What it does: This study investigates the enhancement of fine-grained image zero-shot classification performance of multimodal large language models using self-supervised prompt learning with unlabeled data.
π― What it does: This paper proposes the Cluster Attention Adapter (CLAdapter), which utilizes the knowledge of large pre-trained models to provide adaptive feature transfer and fine-tuning solutions for data-scarce scientific tasks.
Unleashing the Potential of Multimodal LLMs for Zero-Shot Spatio-Temporal Video Grounding
Zaiquan Yang (City University of Hong Kong), Rynson W. H. Lau (City University of Hong Kong)
CodeRecognitionObject DetectionOptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoMultimodality
π― What it does: Utilizing multi-modal large language models (MLLM) to achieve spatiotemporal video localization (STVG) under unsupervised conditions, by identifying and optimizing specific 'localization words', and enhancing localization accuracy through query decomposition and temporal augmentation methods.
π― What it does: This paper proposes the Unlocker framework, which addresses the deadlock problem of noise label learning (NLL) and long-tail learning (LTL) on long-tail noisy label data through a dual-layer optimization.
π― What it does: This paper proposes a latent data distillation method using diffusion models, called LD3M, which achieves efficient distillation of a small number of synthetic samples by learning latent codes and conditional embeddings.
Unlocking SLM Potential for Data Analysis Code Generation via Non-Parametric Knowledge Distillation
Jinyang Li (University of Hong Kong), Reynold Cheng (Microsoft Research)
CodeOptimizationKnowledge DistillationAI Code AssistantTransformerLarge Language ModelTabularRetrieval-Augmented Generation
π― What it does: Achieving efficient and privacy-friendly deployment of data analysis code generation by transferring the knowledge of large language models (LLMs) to small language models (SLMs) in an unsupervised context learning manner.
π― What it does: A real-time, registration-free defense mechanism is proposed, utilizing the biometric features leaked in the pose-expression embeddings transmitted during AI voice and video conferences to detect and prevent 'puppeteering' attacks.
π― What it does: An unsupervised deep learning framework ULOT is proposed for predicting the optimal transport plan of FUGW between graphs, and directly minimizing the FUGW loss through amortized optimization;
Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards
Honghao Chen (Institute of Automation, Chinese Academy of Sciences), Xinlong Wang (Beijing Academy of Artificial Intelligence)
CodeTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
π― What it does: Proposes a Chain of Steps Reasoning (CoS) framework for visual language models, utilizing structured steps (Name-Thought-Reflection) and achieving more precise reinforcement learning and reasoning time scaling through fine-grained rewards.
Unveiling Extraneous Sampling Bias with Data Missing-Not-At-Random
Chunyuan Zheng (Peking University), Mengyue Yang (University of Bristol)
CodeRecommendation SystemTabular
π― What it does: A dual robust estimation method (SuperDR) is proposed under the condition of incomplete consistency between the training and testing set units, which eliminates additional covariance bias by correcting error estimates.
Unveiling the Compositional Ability Gap in Vision-Language Reasoning Model
Tianle Li (Chinese University of Hong Kong), Yu Cheng (Chinese University of Hong Kong)
CodeTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: This study investigates the combinatorial reasoning capabilities of visual-language models (VLM) in reinforcement learning (RL) fine-tuning and proposes the ComPABench benchmark to evaluate combinatorial generalization across modalities, tasks, and out-of-distribution (OOD) conditions.
Unveiling the Uncertainty in Embodied and Operational Carbon of Large AI Models through a Probabilistic Carbon Accounting Model
Xiaoyang Zhang (Hong Kong Polytechnic University), Dan Wang (Hong Kong University of Science and Technology)
CodeTime Series
π― What it does: A probabilistic carbon accounting model (PCAM) has been developed to systematically quantify the embedded carbon and operational carbon uncertainties of large AI models, providing distributed estimation results.
Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs
Mantas Mazeika (Center for AI Safety), Dan Hendrycks (Center for AI Safety)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper systematically analyzes and controls the internal value system of large language models (LLMs) by constructing a Utility Engineering framework.
π― What it does: The Video-to-Voxel (V2V) method is proposed, which directly converts traditional video frames into event voxel representations, avoiding explicit event stream generation and significantly reducing storage and I/O burdens.
VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment
Qing Li (Southwest Jiaotong University), Yu-Shen Liu (Tsinghua University)
CodeGaussian SplattingPoint Cloud
π― What it does: A 3D Gaussian point cloud geometry enhancement method based on view alignment is proposed to achieve more accurate surface reconstruction and high-quality novel view synthesis.
Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought
Chao Huang (Shenzhen Campus of Sun Yat-sen University), Xiaochun Cao (Shenzhen Campus of Sun Yat-sen University)
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoMultimodalityChain-of-Thought
π― What it does: This paper proposes the Video Anomaly Reasoning (VAR) task and constructs an end-to-end multimodal large language model framework called Vad-R1, which guides the model to reason and describe video anomalies step by step through the Perception-to-Cognition Chain-of-Thought (P2C-CoT);
Validating LLM-as-a-Judge Systems under Rating Indeterminacy
Luke Guerdan (Carnegie Mellon University), Alexandra Chouldechova (Microsoft Research)
CodeLarge Language ModelText
π― What it does: In response to the validation issues of the LLM-as-a-judge system under rating indeterminacy, a comprehensive theoretical framework is proposed, and large-scale experiments are conducted with 11 real evaluation tasks and 9 commercial LLMs.
Value-Guided KV Compression for LLMs via Approximated CUR Decomposition
Ayan Sengupta (Indian Institute of Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Delhi)
CodeRetrievalCompressionTransformerLarge Language ModelText
π― What it does: A KV cache compression method based on CUR decomposition, CurDKV (and its adaptive variant AdaCurDKV), is proposed. It selects the key-value pairs to retain by calculating the leverage scores of the key-value matrix, thereby reducing memory usage and inference latency of LLMs under long contexts.
Value-Guided Search for Efficient Chain-of-Thought Reasoning
Kaiwen Wang (Cornell University), Wen Sun (Cornell University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: This paper studies a block-level search method based on a token-level value model (Value-Guided Search, VGS) to enhance the computational efficiency and performance of Chain-of-Thought (CoT) large language models during testing.
π― What it does: Under the Edge of Stability framework, it is analyzed and verified that Variational Learning (VL) can find flatter solutions than gradient descent.
π― What it does: This paper proposes the Variational Polynomial Tree (VPT), which combines the PΓ³lya tree prior with deep generative models to achieve scalable non-parametric density estimation.
π― What it does: This paper proposes the Variational RUOT method, which utilizes variational methods to solve the Regularized Unbalanced Optimal Transport (RUOT) problem by learning a scalar field to simultaneously obtain particle velocity and growth rate, thereby reconstructing the continuous dynamics of high-dimensional systems.
π― What it does: This paper proposes VarCon, a supervised contrastive learning framework based on variational inference, which controls the embedding distribution through a posterior-weighted ELBO objective and introduces confidence-adaptive temperature to finely adjust intra-class dispersion.
Variational Uncertainty Decomposition for In-Context Learning
I. Shavindra Jayasekera (Imperial College London), Yingzhen Li (Imperial College London)
CodeLarge Language ModelPrompt EngineeringText
π― What it does: A Variational Uncertainty Decomposition (VUD) framework is proposed, which decomposes the uncertainty in context learning of large language models into distinguishable aleatoric and epistemic uncertainties using auxiliary queries.
VCM: Vision Concept Modeling with Adaptive Vision Token Compression via Instruction Fine-Tuning
Run Luo (Shenzhen Institute of Advanced Technology), Xiaobo Xia (National University of Singapore)
CodeClassificationObject DetectionSegmentationCompressionTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
π― What it does: Proposes a self-supervised visual concept model (VCM) that allows large-scale visual language models (LVLM) to extract only the visual concept tokens related to instructions, rather than full image tokens;
CodeComputational EfficiencyDrug DiscoveryTransformerSupervised Fine-TuningBiomedical Data
π― What it does: The Venus-MAXWELL framework is proposed, transforming the prediction of protein variant stability from sequence-to-label to sequence-to-landscape, utilizing a matrix scoring method to predict all single-point mutations' ΞΞG in one go.
VeriLoC: Line-of-Code Level Prediction of Hardware Design Quality from Verilog Code
Raghu Vamshi Hemadri (New York University), Siddharth Garg (New York University)
CodeLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a framework called VeriLoC, based on Verilog LLM embedding, to predict hardware design quality in real-time at the RTL level, including timing and routing congestion, and provides impact assessments for specific lines of code.
VeriThinker: Learning to Verify Makes Reasoning Model Efficient
Zigeng Chen (National University of Singapore), Xinchao Wang (National University of Singapore)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
π― What it does: By performing supervised fine-tuning on auxiliary verification tasks in large reasoning models, chain-of-thought (CoT) compression is achieved, significantly reducing the length of reasoning chains;