ICLR 2025 Papers — Page 5
International Conference on Learning Representations · 3704 papers
Bonsai: Gradient-free Graph Condensation for Node Classification
Mridul Gupta (Indian Institute of Technology Delhi), Sayan Ranu (Indian Institute of Technology Delhi)
ClassificationCompressionKnowledge DistillationGraph Neural NetworkGraph
🎯 What it does: This work proposes a gradient-independent graph data compression method based on computational trees, called Bonsai, for node classification tasks;
Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation
Sheng-Feng Yu (National Yang Ming Chiao Tung University), Wei-Chen Chiu (National Yang Ming Chiao Tung University)
Knowledge DistillationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A self-supervised data distillation method is proposed, which compresses large-scale unlabeled datasets into a small number of representative images and their features, achieving smaller sizes and better cross-model generalization through parameterization, predefined augmentation, and an approximate network.
Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation
Tiansheng Huang (Georgia Institute of Technology), Ling Liu (Georgia Institute of Technology)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A defense method named Booster is proposed for the alignment phase, which mitigates harmful loss reduction by incorporating regularization against harmful perturbations during the alignment process, thereby enhancing the model's robustness against subsequent harmful fine-tuning attacks.
Boosting Latent Diffusion with Perceptual Objectives
Tariq Berrada, Jakob Verbeek (FAIR at Meta)
GenerationData SynthesisComputational EfficiencyDiffusion modelAuto EncoderImage
🎯 What it does: A 'latent perceptual loss (LPL)' is proposed to strengthen the connection between the latent diffusion model and the autoencoder decoder, making the generated images more realistic in terms of details and textures.
Boosting Methods for Interval-censored Data with Regression and Classification
Yuan Bian (University of Western Ontario), Wenqing He (University of Western Ontario)
ClassificationOptimizationSupervised Fine-TuningTabularBiomedical Data
🎯 What it does: Two non-parametric boosting methods based on L2 Boost (L2Boost-CUT and L2Boost-IMP) are proposed, specifically designed to address regression and classification problems with interval censored data.
Boosting Multiple Views for pretrained-based Continual Learning
Quyen Tran (Qualcomm AI Research), Trung Le
ClassificationRepresentation LearningTransformerPrompt EngineeringImage
🎯 What it does: A continuous learning framework (BoostCL) that combines Multi-View Random Projection with adaptive AdaBoost is proposed, introducing a self-improvement process to more accurately select task-specific prompts.
Boosting Neural Combinatorial Optimization for Large-Scale Vehicle Routing Problems
Fu Luo (Southern University of Science and Technology), Qingfu Zhang (City University of Hong Kong)
OptimizationTransformerGraph
🎯 What it does: A lightweight cross-attention Transformer and self-improving training (SIT) scheme is proposed for large-scale vehicle routing problems.
Boosting Perturbed Gradient Ascent for Last-Iterate Convergence in Games
Kenshi Abe (CyberAgent), Atsushi Iwasaki (University of Electro-Communications)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: An improved gradient ascent GABP algorithm is proposed for monotonic games through an enhanced reward perturbation technique, and its convergence rate is provided.
Boosting Ray Search Procedure of Hard-label Attacks with Transfer-based Priors
Chen Ma (Zhejiang University of Technology), Qi Xuan (Zhejiang University of Technology)
OptimizationAdversarial AttackImage
🎯 What it does: Proposes two hard label attacks, Prior-OPT and Prior-Sign-OPT, which utilize transfer benchmarks to improve gradient estimation accuracy and significantly reduce query costs.
Boosting the visual interpretability of CLIP via adversarial fine-tuning
Shizhan Gong (Chinese University of Hong Kong), Farzan Farnia (Chinese University of Hong Kong)
SegmentationExplainability and InterpretabilityTransformerGenerative Adversarial NetworkContrastive LearningImageMultimodality
🎯 What it does: For the image encoder of CLIP, an unsupervised adversarial fine-tuning (AFT) method is proposed, which enhances its visual interpretability by incorporating norm regularization, and makes the fine-tuned encoder compatible with the original text encoder for application in multimodal large models.
Bootstrapped Model Predictive Control
Yuhang Wang (Xi'an Jiaotong University), Xuguang Lan (Xi'an Jiaotong University)
OptimizationReinforcement LearningWorld ModelSequential
🎯 What it does: The Bootstrapped Model Predictive Control (BMPC) algorithm is proposed, which enhances the learning efficiency and performance of continuous control tasks by using MPC as an expert to guide policy and value learning.
Bootstrapping Language Models with DPO Implicit Rewards
Changyu Chen (Sea AI Lab), Min Lin (Sea AI Lab)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Utilize DPO implicit rewards for self-alignment of LLMs, iteratively generate preference data, and perform DPO fine-tuning;
Bootstrapping Language-Guided Navigation Learning with Self-Refining Data Flywheel
Zun Wang (Shanghai AI Laboratory), Limin Wang (Shanghai AI Laboratory)
Data-Centric LearningRobotic IntelligenceTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: A fully automated self-improving data flywheel (SRDF) has been constructed, which generates high-quality large-scale language instruction-trajectory pair datasets through mutual screening and regeneration between the navigator and the instruction generator, enhancing the performance of visual-language navigation.
Both Ears Wide Open: Towards Language-Driven Spatial Audio Generation
Peiwen Sun (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageTextAudio
🎯 What it does: This paper proposes a method for generating stereo (dual-channel) spatial audio from text or images, and releases the first large-scale stereo audio dataset with spatial descriptions, BEWO-1M.
Bounds on $L_p$ Errors in Density Ratio Estimation via $f$-Divergence Loss Functions
Yoshiaki Kitazawa (NTT DATA Mathematical Systems Inc)
Tabular
🎯 What it does: By using the variational f-divergence representation, upper and lower bounds for any Lp error in the logarithmic domain are derived, applicable to any Lipschitz continuous density ratio estimator, and independent of the specific f-divergence loss function.
BP-Modified Local Loss for Efficient Training of Deep Neural Networks
REN Lianhai, Qianxiao Li (National University of Singapore)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerImageStochastic Differential Equation
🎯 What it does: This paper proposes a BP-modified local loss method, which significantly improves training performance by utilizing the true backpropagation gradient to correct the local loss gradient while maintaining low memory usage.
BRAID: Input-driven Nonlinear Dynamical Modeling of Neural-Behavioral Data
Parsa Vahidi (University of Southern California), Maryam Shanechi
Recurrent Neural NetworkTime SeriesSequential
🎯 What it does: This paper proposes a deep learning-based BRAID framework for nonlinear dynamic modeling of neuro-behavioral data under measured external inputs.
Brain Bandit: A Biologically Grounded Neural Network for Efficient Control of Exploration
Chen Jiang (Chinese Institute for Brain Research), Ni Ji (Chinese Academy of Medical Sciences and Peking Union Medical College)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: Proposed and analyzed a biologically inspired stochastic continuous Hopfield network (BBN) for efficient exploration in multi-armed bandit and MDP tasks.
Brain Mapping with Dense Features: Grounding Cortical Semantic Selectivity in Natural Images With Vision Transformers
Andrew Luo, Michael J. Tarr (Carnegie Mellon University)
SegmentationRepresentation LearningTransformerContrastive LearningImageMagnetic Resonance Imaging
🎯 What it does: The BrainSAIL method is proposed, which extracts spatially dense semantic embeddings from pre-trained visual Transformers (CLIP, DINO, SigLIP) and combines them with an fMRI encoder to achieve spatial attribution and semantic localization of neural selectivity in higher-order visual cortex.
Brain-inspired $L_p$-Convolution benefits large kernels and aligns better with visual cortex
Jea Kwon (Max Planck Institute), C. Justin Lee (Institute for Basic Science)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an Lp-convolution method based on multivariate p-generalized normal distribution, which overlays learnable sparse masks on the convolution kernel, allowing large convolution kernels to maintain biologically common Gaussian sparse connections, thereby enhancing model performance and robustness.
BrainACTIV: Identifying visuo-semantic properties driving cortical selectivity using diffusion-based image manipulation
Diego Garcia Cerdas, Iris Groen (Goethe University Frankfurt)
GenerationData SynthesisDiffusion modelImageMagnetic Resonance ImagingStochastic Differential Equation
🎯 What it does: This study proposes the BrainACTIV method, which utilizes diffusion models to semantically modulate reference images in order to maximize or minimize the activation of specific brain regions in the human visual cortex.
BrainOOD: Out-of-distribution Generalizable Brain Network Analysis
Jiaxing Xu (Nanyang Technological University), Yiping Ke (Nanyang Technological University)
ClassificationDomain AdaptationExplainability and InterpretabilityGraph Neural NetworkGraphBiomedical DataAlzheimer's Disease
🎯 What it does: The BrainOOD framework is proposed for interpretable feature and structure selection in brain networks to enhance classification performance under multi-site distribution shift (OOD).
BrainUICL: An Unsupervised Individual Continual Learning Framework for EEG Applications
Yangxuan Zhou (Zhejiang University), Gang Pan (Zhejiang University)
Domain AdaptationAnomaly DetectionConvolutional Neural NetworkContrastive LearningTime SeriesBiomedical Data
🎯 What it does: The BrainUICL framework is proposed to achieve unsupervised individual continuous learning of EEG, enabling continuous adaptation and improvement of generalization ability to unseen subjects among continuously emerging new subjects.
Breach By A Thousand Leaks: Unsafe Information Leakage in 'Safe' AI Responses
David Glukhov (University of Toronto), Nicolas Papernot (University of Toronto)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper proposes a 'reasoning adversary' threat model and designs a black-box attack based on problem decomposition (Decomposition Attack, DA) to leak dangerous knowledge from constrained LLMs, and constructs a new security assessment framework (Impermissible Information Leakage, IIL).
Breaking Class Barriers: Efficient Dataset Distillation via Inter-Class Feature Compensator
Xin Zhang (Agency for Science Technology and Research), Joey Tianyi Zhou (Agency for Science Technology and Research)
Data SynthesisKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new dataset distillation paradigm called INFER, which utilizes a Universal Feature Compensator (UFC) to extend a single synthetic sample to all categories, and achieves efficient MixUp augmentation through static soft labels, significantly enhancing the diversity and generalization ability of synthetic data.
Breaking Free from MMI: A New Frontier in Rationalization by Probing Input Utilization
Wei Liu (Hunan University of Science and Technology), Ruixuan Li (Hunan University of Science and Technology)
ClassificationExplainability and InterpretabilityGraph Neural NetworkSupervised Fine-TuningReinforcement LearningTextGraph
🎯 What it does: An interpretable framework N2R based on network utilization of input is proposed, replacing the traditional Maximum Mutual Information (MMI) criterion to extract rational explanations from the perspective of how the network utilizes the input.
Breaking Mental Set to Improve Reasoning through Diverse Multi-Agent Debate
Yexiang Liu (Institute of Automation Chinese Academy of Sciences), Tieniu Tan (Institute of Automation Chinese Academy of Sciences)
Large Language ModelPrompt EngineeringTextMultimodalityChain-of-Thought
🎯 What it does: A multi-agent debate framework DMAD is proposed, which enhances the reasoning accuracy of LLM/MLLM by allowing different agents to use different reasoning methods.
Breaking Neural Network Scaling Laws with Modularity
Akhilan Boopathy (Massachusetts Institute of Technology), Ila R Fiete
ClassificationOptimizationImage
🎯 What it does: This study investigates the generalization performance of modular neural networks on high-dimensional compositional tasks and proposes a theoretical model and learning rules.
Breaking the $\log(1/\Delta_2)$ Barrier: Better Batched Best Arm Identification with Adaptive Grids
Tianyuan Jin (National University of Singapore), Dongruo Zhou (Indiana University Bloomington)
Reinforcement Learning
🎯 What it does: This paper proposes batch optimal arm identification algorithms IS-SE and IS-RAGE for multi-armed and linear Bandits, achieving instance-sensitive batch complexity with nearly optimal sample complexity, breaking the traditional log(1/Δ)^2 limit.
Breaking the Reclustering Barrier in Centroid-based Deep Clustering
Lukas Miklautz (University of Vienna), Claudia Plant (University of Vienna)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: A BRB algorithm based on the combination of weight reset and re-clustering is proposed to break through the saturation barrier of clustering objectives and enhance the performance of the Deep Clustering (DC) algorithm.
Bridging and Modeling Correlations in Pairwise Data for Direct Preference Optimization
Yuxin Jiang (Hong Kong University of Science and Technology), Wei Wang (Hong Kong University of Science and Technology)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A BMC framework is proposed, which enhances the relevance of contrastive data by first generating pseudo-winning responses before adversarial training, and then dynamically adjusting token-level rewards during DPO training to achieve finer-grained preference optimization.
Bridging Compressed Image Latents and Multimodal Large Language Models
Chia-Hao Kao (Honda Research Institute USA), Riccardo Leonardi (University of Brescia)
CompressionKnowledge DistillationTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A system for image compression targeting multimodal large language models is proposed, utilizing a lightweight transform-neck to directly map the compressed image latent vectors to the intermediate layers of the visual encoder.
Bridging Context Gaps: Leveraging Coreference Resolution for Long Contextual Understanding
Yanming Liu (Zhejiang University), Xuhong Zhang (Zhejiang University)
TransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: To address coreference ambiguity in long text question answering, a Long Question Coreference Adaptation (LQCA) framework is proposed, which first performs coreference resolution on sub-documents, then calculates mention distances, selects representative mentions, and replaces the original text, thereby improving the understanding and answering quality of LLMs for long texts.
Bridging Information Asymmetry in Text-video Retrieval: A Data-centric Approach
Zechen Bai (National University of Singapore), Mike Zheng Shou (National University of Singapore)
RetrievalData-Centric LearningTransformerLarge Language ModelVideoTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a data-centric text enhancement framework that, during the training phase, covers all video scenes through event-aware segmentation and image-generated subtitles. In the retrieval phase, it utilizes large language models to generate diverse queries and employs Farthest Query Sampling (FQS) for efficient query filtering, significantly improving text-video retrieval performance.
Bridging Jensen Gap for Max-Min Group Fairness Optimization in Recommendation
Chen Xu (Renmin University of China), Tat-Seng Chua (National University of Singapore)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: This study investigates the Jensen gap problem caused by mini-batch training based on the group maximum-minimum fairness (MMF) objective in recommendation systems, and proposes the FairDual algorithm to reduce this gap through dual optimization methods.
Bridging the Data Provenance Gap Across Text, Speech, and Video
Shayne Longpre, Jad Kabbara
VideoTextMultimodalityAudio
🎯 What it does: Conduct a large-scale longitudinal audit of publicly available text, speech, and video datasets to assess their sources, licenses, terms of use, and geographical/language diversity.
Bridging the Gap between Database Search and \emph{De Novo} Peptide Sequencing with SearchNovo
Jun Xia (Westlake University), Stan Z. Li (Westlake University)
Transformer
🎯 What it does: By combining database search with de novo sequence inference, the identification accuracy of peptide sequences is significantly improved by retrieving the most similar spectra and integrating reference sequences into the Transformer model.
Bridging the Gap Between f-divergences and Bayes Hilbert Spaces
Linus Lach (University of Augsburg), Yarema Okhrin (University of Augsburg)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A locally non-convex pseudo f-divergence framework is proposed, which is associated with the metric of the Bayes Hilbert space, thereby enabling sampling of high-dimensional posterior distributions.
Bridging the Gap between Variational Inference and Stochastic Gradient MCMC in Function Space
Mengjing Wu (Australian Artificial Intelligence Institute), Jie Lu (Australian Artificial Intelligence Institute)
OptimizationImageTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A mixed inference method called FVIMC is proposed, which simultaneously uses function variational inference (fVI) and function MCMC (fMCMC) in function space, achieving posterior approximation through alternating V-Stage, M-Stage, and T-Stage.
Bridging the Semantic Gap Between Text and Table: A Case Study on NL2SQL
Lin Long (Zhejiang University), Junbo Zhao (Zhejiang University)
TransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextTabular
🎯 What it does: Proposes the TNT framework, which utilizes table-text multimodal representation to help LLM better understand table semantics.
BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval
Hongjin SU, Tao Yu (University of Hong Kong)
RetrievalLarge Language ModelTextBenchmark
🎯 What it does: A BRIGHT retrieval benchmark has been constructed, containing real queries that require deep reasoning to retrieve relevant documents.
Bringing NeRFs to the Latent Space: Inverse Graphics Autoencoder
Antoine Schnepf (Criteo AI Lab), Valerie Gouet-Brunet
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldAuto EncoderImagePoint Cloud
🎯 What it does: Proposed the Inverse Graphics Autoencoder (IG-AE) and latent NeRF training pipeline, which transfers NeRF to a 3D-aware latent space, thereby accelerating training and rendering while improving view synthesis quality.
Broaden your SCOPE! Efficient Multi-turn Conversation Planning for LLMs with Semantic Space
Zhiliang Chen (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A framework called SCOPE is proposed for efficient dialogue planning in real-time multi-turn conversations using LLMs, without the need for additional LLM inference.
Broadening Target Distributions for Accelerated Diffusion Models via a Novel Analysis Approach
Yuchen Liang (Ohio State University), Ness Shroff (Ohio State University)
GenerationOptimizationDiffusion modelTabularOrdinary Differential Equation
🎯 What it does: This paper proposes a Hessian-accelerated DDPM sampler and provides accelerated convergence guarantees under a broader range of target distributions (smooth, finite variance, Gaussian mixture).
BTBS-LNS: Binarized-Tightening, Branch and Search on Learning LNS Policies for MIP
Hao Yuan (Lenovo Research), Junchi Yan (Shanghai Jiao Tong University)
OptimizationGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: This paper proposes a learning-based LNS (BTBS-LNS) that combines binary constraint shrinking, a graph attention model, and an additional branching strategy to solve general MIP problems.
Budgeted Online Continual Learning by Adaptive Layer Freezing and Frequency-based Sampling
Minhyuk Seo (Seoul National University), Jonghyun Choi (Seoul National University)
Computational EfficiencyLarge Language ModelImageMultimodality
🎯 What it does: A new online continual learning method is proposed under limited computation and storage budgets, balancing high performance and low resource consumption.
Build-A-Scene: Interactive 3D Layout Control for Diffusion-Based Image Generation
Abdelrahman Eldesokey (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Proposes Build-A-Scene, an interactive 3D layout control method based on diffusion models, employing a multi-stage generation process, Dynamic Self-Attention (DSA), and a consistent 3D translation strategy;
Building Interactable Replicas of Complex Articulated Objects via Gaussian Splatting
Yu Liu (Tsinghua University), Siyuan Huang (Peking University)
Pose EstimationRobotic IntelligenceGaussian SplattingPoint CloudMesh
🎯 What it does: Utilizing 3D Gaussian splatting to construct a digital twin of an interactive multi-part complex robotic arm object, completing part mesh reconstruction and joint parameter estimation.
Building Math Agents with Multi-Turn Iterative Preference Learning
Wei Xiong (University of Illinois Urbana-Champaign), Tianqi Liu (Google Deepmind)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This study investigates the combination of a tool-integrated mathematical reasoning LLM with direct preference learning, proposing a multi-round direct preference learning framework (M-DPO, M-KTO) and an online iterative Gibbs sampling training method to enhance the model's performance in mathematical reasoning tasks.
Building, Reusing, and Generalizing Abstract Representations from Concrete Sequences
Shuchen Wu (Max Planck Institute for Biological Cybernetics), Eric Schulz (Helmholtz Munich)
CompressionRepresentation LearningTextSequential
🎯 What it does: This paper proposes a non-parametric hierarchical variable learning model (HVM) that can automatically learn chunks and abstract variables from continuous sequences, achieving a more compact sequence representation.
Bundle Neural Network for message diffusion on graphs
Jacob Bamberger (University of Oxford), Michael M. Bronstein
Graph Neural NetworkGraphStochastic Differential Equation
🎯 What it does: Proposes the Bundle Neural Networks (BuNN) architecture, which learns graph neural networks by diffusing information over flat vector bundles;
C-CLIP: Multimodal Continual Learning for Vision-Language Model
Wenzhuo Liu (University of Chinese Academy of Sciences), Qi Tian (Huawei)
RetrievalDomain AdaptationTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: A multi-modal continuous learning framework named C-CLIP is proposed, which improves the catastrophic forgetting problem of visual-language models when continuously receiving new domain data.
Cached Multi-Lora Composition for Multi-Concept Image Generation
Xiandong Zou (Imperial College London), Yiren Zhao (Imperial College London)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A training-free multi-LoRA combination framework called CMLoRA is proposed, which determines the injection order of LoRA by analyzing the high-frequency/low-frequency features of LoRA in the frequency domain, and employs a caching mechanism to enhance stability, addressing the semantic conflict issue in multi-concept image generation.
Cafe-Talk: Generating 3D Talking Face Animation with Multimodal Coarse- and Fine-grained Control
Hejia Chen (Kuaishou Technology), Shuai Li
GenerationData SynthesisTransformerDiffusion modelVideoTextMultimodalityAudio
🎯 What it does: Developed Cafe-Talk, a 3D talking facial animation model based on diffusion Transformer, which supports both coarse-grained (speaking style, emotion labels, and intensity) and fine-grained (AU action units) multimodal control, and provides a text AU detector for natural language input.
CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences
Ziran Qin (Shanghai Jiao Tong University), Jianguo Li (Ant Group)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: The CAKE framework is proposed, which combines adaptive KV cache allocation with hierarchical priority, hierarchical cascading management, and designs an eviction metric that considers dynamic time and space attention to efficiently manage KV caches in long text contexts.
Calibrating Expressions of Certainty
Peiqi Wang (Massachusetts Institute of Technology), Polina Golland (Massachusetts Institute of Technology)
TextMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Proposes modeling confidence expressions as probability distributions and defines a new ECE estimation and reliability graph based on this.
Calibrating LLMs with Information-Theoretic Evidential Deep Learning
Yawei Li (LMU Munich), Mina Rezaei (LMU Munich)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes an Information Bottleneck-based Evidence Deep Learning (IB-EDL) method, which utilizes information bottleneck regularization to reduce the issues of overconfidence and poor calibration that arise when fine-tuning large language models on small datasets.
CameraCtrl: Enabling Camera Control for Video Diffusion Models
Hao He (Chinese University of Hong Kong), Ceyuan Yang (Shanghai Artificial Intelligence Laboratory)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: A plugin module named CameraCtrl has been developed, allowing video diffusion models to precisely control the camera perspective in videos based on given camera trajectories, and it can seamlessly integrate with various benchmark models (such as AnimateDiff, SVD) and other control methods (such as SparseCtrl).
CAMEx: Curvature-aware Merging of Experts
Viet Dung Nguyen, Linh Duy Tran (Viettel Group)
Mixture of ExpertsImageText
🎯 What it does: This paper proposes CAMEx, a sparse expert merging protocol that considers parameter curvature and implements a dynamic merging architecture.
Can a Large Language Model be a Gaslighter?
Wei Li (National University of Singapore), Yang You (National University of Singapore)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper studies the risks of large language models (LLMs) in gaslighting behavior, constructs a dataset of gaslighting conversations and safe conversations, and proposes the DeepCoG two-stage framework along with prompt/fine-tuning attacks and safety alignment methods.
Can a MISL Fly? Analysis and Ingredients for Mutual Information Skill Learning
Chongyi Zheng (Princeton University), Benjamin Eysenbach (Princeton University)
Robotic IntelligenceReinforcement LearningContrastive LearningSequential
🎯 What it does: This paper demonstrates through theoretical and experimental research that the current strongest unsupervised skill learning method, METRA, can be explained within the mutual information (MI) framework, and based on this, proposes a simpler contrastive learning + Successor Features (CSF) algorithm.
Can Generative AI Solve Your In-Context Learning Problem? A Martingale Perspective
Andrew Jesson (Columbia University), David Blei (Columbia University)
GenerationData SynthesisImageTextTabular
🎯 What it does: A new method is proposed to evaluate whether conditional generative models (CGMs) can solve the in-context learning (ICL) problem—generative predictive p-value.
Can In-context Learning Really Generalize to Out-of-distribution Tasks?
Qixun Wang (Peking University), Yisen Wang (Peking University)
ClassificationRetrievalOptimizationTransformerLarge Language ModelText
🎯 What it does: This study investigates the in-context learning (ICL) mechanism of Transformers in out-of-distribution (OOD) tasks, evaluating whether it can truly learn new tasks and exploring its algorithmic selection behavior under abstract label classification and multi-task pre-training.
Can Knowledge Editing Really Correct Hallucinations?
Baixiang Huang (Emory University), Kai Shu (Illinois Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: A HallUEditBench benchmark dataset was constructed, containing 9 domains, 26 topics, and over 6,000 real hallucination examples. The performance of seven knowledge editing methods was evaluated from five dimensions: Efficacy, Generalization, Portability, Locality, and Robustness.
Can Large Language Models Understand Symbolic Graphics Programs?
Zeju Qiu (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextBenchmark
🎯 What it does: A benchmark called SGP-Bench is proposed to evaluate the semantic understanding of large language models for symbolic graphic programs (SVG and CAD), and the benchmark is used to verify the models' reasoning and consistency. Additionally, Symbolic Instruction Tuning (SIT) is introduced to fine-tune open-source LLMs, enhancing their understanding of symbolic graphic programs and general reasoning capabilities.
Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
Chenglei Si (Stanford University), Tatsunori Hashimoto (Stanford University)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This study recruited over 100 NLP experts to conduct a large-scale blind evaluation experiment to assess the performance of LLMs in generating research ideas and compare them with ideas from human experts.
Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book?
Seth Aycock (University of Amsterdam), Khalil Sima'an (University of Amsterdam)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: By decomposing the grammar books of extremely low-resource languages, this paper explores whether LLMs can utilize grammatical descriptions for translation, comparing the performance of long-context LLMs with fine-tuned small MT models, and introducing typological feature prompts to evaluate their performance on tasks such as grammatical judgment and stemming prediction.
Can LLMs Separate Instructions From Data? And What Do We Even Mean By That?
Egor Zverev (Institute of Science and Technology Austria), Christoph H. Lampert (Institute of Science and Technology Austria)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This study investigates the separation of instructions and data in single-turn contexts for large language models (LLMs). It proposes a formal definition of separation, a computable empirical proxy metric, and releases the SEP dataset specifically for evaluating this metric. Subsequently, it assesses the separation and practicality of several mainstream LLMs and explores mitigation strategies such as prompt engineering, prompt optimization, and low-rank fine-tuning.
Can LLMs Solve Longer Math Word Problems Better?
Xin Xu (Hong Kong University of Science and Technology), Yang Wang (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study investigates the ability of LLMs to solve mathematical word problems in longer contexts, constructs the E-GSM dataset, and proposes the CoLeG metric.
Can LLMs Understand Time Series Anomalies?
Zihao Zhou (University of California San Diego), Rose Yu (University of California San Diego)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringMultimodalityTime SeriesChain-of-Thought
🎯 What it does: The system evaluates the understanding and performance of LLM in time series anomaly detection, conducting experiments under various settings such as zero/few-shot, chain reasoning, and text/image input to validate six key hypotheses.
Can Neural Networks Achieve Optimal Computational-statistical Tradeoff? An Analysis on Single-Index Model
Siyu Chen (Yale University), Tianhao Wang (Toyota Technological Institute at Chicago)
Tabular
🎯 What it does: This paper studies whether neural networks can achieve the optimal statistical-computational trade-off when learning Gaussian single-index models. A unified gradient-based algorithm is proposed for training two-layer neural networks, which can adapt to various loss and activation functions in polynomial time.
Can One Modality Model Synergize Training of Other Modality Models?
Jae-Jun Lee (Ulsan National Institute of Science and Technology), Sung Whan Yoon (Ulsan National Institute of Science and Technology)
Domain AdaptationRepresentation LearningTransformerPrompt EngineeringImageTextMultimodalityAudio
🎯 What it does: This paper proposes a method to facilitate the training of one modality model using the representation of another modality model (which may be incomplete or mismatched), achieving cross-modal collaborative learning in the absence of paired supervision.
Can Reinforcement Learning Solve Asymmetric Combinatorial-Continuous Zero-Sum Games?
Yuheng Li (William and Mary), Haipeng Chen (William and Mary)
OptimizationReinforcement LearningGraph
🎯 What it does: This paper defines a new class of asymmetric combinatorial-continuous zero-sum games (ACCES) and proposes an algorithm based on the Combined Continuous Dual Oracle (CCDO) and its reinforcement learning implementation (CCDO-RL) to solve the Nash equilibria of these games.
Can Textual Gradient Work in Federated Learning?
Minghui Chen (University of British Columbia), Xiaoxiao Li (University of British Columbia)
OptimizationFederated LearningSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the FedTextGrad framework, which utilizes text gradients for iterative optimization of prompts in a federated learning environment.
Can Transformers Do Enumerative Geometry?
Baran Hashemi (Technical University Munich), Alessandro Giacchetto (ETH Zurich)
TransformerTabular
🎯 What it does: Using Transformer to predict ψ-type crossing numbers and reproducing Virasoro constraints through self-learning.
Can Video LLMs Refuse to Answer? Alignment for Answerability in Video Large Language Models
Eunseop Yoon (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningVideoText
🎯 What it does: A framework of 'answerability alignment' is proposed, enabling video large language models to judge and reject questions that exceed the scope of video information.
Can Watermarked LLMs be Identified by Users via Crafted Prompts?
Aiwei Liu (Tsinghua University), Xuming Hu (Hongkong University of Science and Technology)
Large Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a unified black-box detection method called Water-Probe, which can sample the same LLM multiple times through carefully designed prompts and compare the distribution differences of different watermark keys for similar prompts, thereby determining whether the LLM has embedded watermarks; it also introduces the Water-Bag scheme to enhance the undetectability of watermarks.
Can Watermarks be Used to Detect LLM IP Infringement For Free?
Zhengyue Zhao (University of Wisconsin-Madison), Chaowei Xiao (University of Wisconsin-Madison)
TransformerLarge Language ModelText
🎯 What it does: This paper explores how to use watermarking technology of large language models (LLMs) to detect intellectual property infringement of models, proposing a new detection method called LIDet.
Can We Ignore Labels in Out of Distribution Detection?
Hong Yang (Rochester Institute of Technology), Travis Desell (Rochester Institute of Technology)
Anomaly DetectionContrastive LearningImageBenchmark
🎯 What it does: This paper demonstrates, through information-theoretic analysis and experimental research, the existence of 'label blindness' in unsupervised/self-supervised OOD detection methods without the use of labels, which leads to inevitable failure when ID and OOD share features. It proposes the Adjacent OOD evaluation task to validate this theory.
Can We Talk Models Into Seeing the World Differently?
Paul Gavrikov (Offenburg University), Janis Keuper (Tübingen AI Center)
TransformerPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This study investigates the performance of visual-language models (VLM) on texture/shape bias and explores the impact of natural language prompts on this bias.
Can We Trust Embodied Agents? Exploring Backdoor Attacks against Embodied LLM-Based Decision-Making Systems
Ruochen Jiao (Northwestern University), Qi Zhu (Northwestern University)
Adversarial AttackRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringVision Language ModelTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: The BALD framework is proposed to conduct three types of backdoor attacks (word injection, scene manipulation, knowledge injection) on LLM-based embodied decision-making systems, and evaluations are performed on various LLMs and simulation platforms.
Capability Localization: Capabilities Can be Localized rather than Individual Knowledge
Xiusheng Huang (Chinese Academy of Sciences), Kang Liu (Chinese Academy of Sciences)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study evaluates the effectiveness of existing single knowledge localization methods through fidelity and reliability experimental systems and proves their ineffectiveness; subsequently, it proposes the 'Commonality Neuron Localization (CNL)' method, which locates shared neurons in similar datasets and verifies through fine-tuning, erasure, and cross-data experiments that these neurons are a collection of model capabilities, significantly enhancing or weakening the model's performance on different tasks.
CapeX: Category-Agnostic Pose Estimation from Textual Point Explanation
Matan Rusanovsky (Tel Aviv University), Shai Avidan (Tel Aviv University)
Pose EstimationGraph Neural NetworkTransformerLarge Language ModelImageTextGraph
🎯 What it does: A category-agnostic pose estimation method called CapeX is proposed, which utilizes a structured graph composed of text descriptions (text-graph) to locate key points in a query image without the need for supporting images.
CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation
Jie Liu (University of Amsterdam), Efstratios Gavves (University of Amsterdam)
OptimizationRobotic IntelligenceTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes the Cooperative Plan Optimization (CaPo) framework, which allows large language model-driven embodied multi-agent systems to generate long-term meta-plans through multi-round discussions before executing tasks, and dynamically adjust based on progress during execution.
Captured by Captions: On Memorization and its Mitigation in CLIP Models
Wenhao Wang (CISPA), Franziska Boenisch (CISPA)
Representation LearningData-Centric LearningTransformerLarge Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the CLIPMem metric, which quantifies the degree of memorization of the CLIP model, and analyzes the impact of different sample types (such as mislabeled and outlier samples) on memorization through experiments, subsequently proposing a memorization mitigation strategy in the text modality.
Capturing the Temporal Dependence of Training Data Influence
Jiachen T. Wang (Princeton University), Ruoxi Jia (Virginia Tech)
Tabular
🎯 What it does: This study investigates the time dependence of file impact during the training process, proposing a Trajectory-Specific Leave-One-Out (TSLOO) method and Data Value Embedding (DVEmb), which can quickly estimate the influence of samples on the final model at different training moments.
CarbonSense: A Multimodal Dataset and Baseline for Carbon Flux Modelling
Matthew Fortier (Mila Quebec AI Institute), Christopher Pal (Mila Quebec AI Institute)
TransformerMultimodalityTime SeriesBenchmarkAgriculture Related
🎯 What it does: The CarbonSense multimodal carbon flux dataset and the baseline model EcoPerceiver are proposed, and their effectiveness in global carbon flux prediction is validated.
CARTS: Advancing Neural Theorem Proving with Diversified Tactic Calibration and Bias-Resistant Tree Search
Xiao-Wen Yang (Nanjing University), Yu-Feng Li (Nanjing University)
Large Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Designed and implemented the CARTS algorithm, which enhances the performance of first-order tree search in neural theorem proving through diversification strategies for calibration and bias-resistant value functions.
CAT-3DGS: A Context-Adaptive Triplane Approach to Rate-Distortion-Optimized 3DGS Compression
Yu-Ting Zhan (National Yang Ming Chiao Tung University), Wen-Hsiao Peng (National Yang Ming Chiao Tung University)
CompressionOptimizationGaussian SplattingPoint Cloud
🎯 What it does: A 3D Gaussian Splatting compression framework CAT-3DGS based on PCA-guided multi-scale triplane and context-adaptive encoding is proposed.
Catastrophic Failure of LLM Unlearning via Quantization
Zhiwei Zhang (Pennsylvania State University), Suhang Wang (Pennsylvania State University)
Large Language ModelText
🎯 What it does: The paper investigates the catastrophic failure of machine unlearning after quantization in LLMs and proposes a gradient significance-based unlearning framework with a large learning rate (SURE) to prevent quantization from recovering forgotten knowledge.
CATCH: Channel-Aware Multivariate Time Series Anomaly Detection via Frequency Patching
Xingjian Wu (East China Normal University), Bin Yang (Hong Kong University of Science and Technology)
Anomaly DetectionTransformerTime Series
🎯 What it does: A multivariate time series anomaly detection framework named CATCH is proposed, which achieves joint detection of point anomalies and subsequence anomalies through frequency domain patching and a channel fusion module.
CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models
Zheng Chong (Sun Yat-sen University), Xiaodan Liang (Sun Yat-sen University)
Image TranslationGenerationDiffusion modelAuto EncoderImage
🎯 What it does: CatVTON is proposed, a virtual try-on system that achieves clothing and person image stitching in the spatial dimension using a simplified VAE+UNet structure.
Cauchy-Schwarz Regularizers
Sueda Taner (ETH Zurich), Christoph Studer (ETH Zurich)
OptimizationImage
🎯 What it does: A differentiable, automatically scaling, and non-false stationary regularization function constructed using the Cauchy-Schwarz inequality, referred to as the CS regularizer, is proposed. It can induce various structural properties (discrete value vectors, feature vectors, orthogonal column matrices, etc.) and is applied to tasks such as solving underdetermined linear equations and neural network weight quantization.
Causal Concept Graph Models: Beyond Causal Opacity in Deep Learning
Gabriele Dominici (Università della Svizzera italiana), Marc Langheinrich (Università della Svizzera italiana)
Explainability and InterpretabilityGraph Neural NetworkReinforcement LearningImage
🎯 What it does: A Causal Concept Graph Models (Causal CGMs) architecture is designed to address the causal ambiguity of deep learning models by learning interpretable causal graphs and high-dimensional concept representations, making the decision-making process traceable.
Causal Discovery via Bayesian Optimization
Bao Duong (Deakin University), Thin Nguyen (Deakin University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: The DrBO method is proposed, which uses Bayesian optimization to find high-scoring DAGs, achieving sample-efficient causal graph learning.
Causal Effect Estimation with Mixed Latent Confounders and Post-treatment Variables
Yaochen Zhu (University of Virginia), Jundong Li (University of Virginia)
Auto Encoder
🎯 What it does: This paper studies the causal estimation bias caused by the mixing of post-treatment variables in observational data and proposes a new identifiable variational autoencoder (CiVAE) to address this bias, achieving an unbiased average treatment effect (ATE).
Causal Graph Transformer for Treatment Effect Estimation Under Unknown Interference
Anpeng Wu (Zhejiang University), Kun Zhang (Zhejiang University)
Graph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes CauGramer, a causal graph transformer model for estimating treatment effects under unknown interference graphs; it aggregates L-order neighbor information through cross-attention and combines representation balancing with minimax moment conditions to achieve joint estimation of direct effects, peer effects, and total effects.
Causal Graphical Models for Vision-Language Compositional Understanding
Fiorenzo Parascandolo (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
GenerationRetrievalTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a visual-language model (VLM) autoregressive training method called COGT based on causal graph models (CGM). It utilizes a dependency parser to pre-construct causal relationships between words, guiding semi-parallel word prediction, thereby enhancing the understanding and retrieval capabilities for language compositional tasks.
Causal Identification for Complex Functional Longitudinal Studies
Andrew Ying
Time SeriesBiomedical DataElectronic Health RecordsStochastic Differential Equation
🎯 What it does: A non-parametric causal identification framework for continuous-time infinite-dimensional functional longitudinal data is proposed;
Causal Information Prioritization for Efficient Reinforcement Learning
Hongye Cao (Nanjing University), Yang Gao (Nanjing University)
Reinforcement Learning
🎯 What it does: A Causal Information Prioritization (CIP) framework is proposed and implemented, which discovers causal relationships between state-reward and action-reward through causal structure, generates synthetic transitions by counterfactual state feature exchange, and enhances the sample efficiency of reinforcement learning by combining causal weighting with mutual information objectives.