ICLR 2026 Papers — Page 8
International Conference on Learning Representations · 5356 papers
Characterizing Human Semantic Navigation in Concept Production as Trajectories in Embedding Space
Felipe Diego Toro-Hernández (Federal University of ABC), Rodrigo M. Cabral-Carvalho (Federal University of ABC)
RetrievalRepresentation LearningTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: View concept generation as a trajectory in the semantic embedding space, and utilize cumulative embeddings to calculate five kinematic and geometric metrics (distance to next, velocity, acceleration, entropy, distance to centroid) to characterize the human semantic retrieval process.
Characterizing Pattern Matching and Its Limits on Compositional Task Structures
Hoyeon Chang (KAIST AI), Minjoon Seo (KAIST AI)
Explainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Formally define, theoretically analyze, and experimentally verify the pattern matching mechanisms of large language models (LLMs) in combinatorial reasoning tasks. Introduce the concepts of functional equivalence and coverage boundaries, and systematically study the performance of Transformer and Mamba on controlled tasks.
Characterizing the Discrete Geometry of ReLU Networks
Blake B. Gaines (University of Connecticut), Jinbo Bi (University of Connecticut)
Explainability and InterpretabilityImageTabular
🎯 What it does: This paper studies the discrete geometric structure of polyhedral complexes defined by fully connected ReLU networks, providing theoretical upper bounds on the average degree and diameter of the connectivity graph, while proposing a BFS enumeration algorithm based on symbolic sequences to construct the complete complex;
Chart Deep Research in LVLMs via Parallel Relative Policy Optimization
Jiajin Tang (ByteDance), Xing Chen (ByteDance)
OptimizationRepresentation LearningReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: To address the bottlenecks in training and evaluating the ability to deeply analyze charts, this paper proposes a parallel relative policy optimization (PRPO) training framework and an objective evaluation benchmark (MCDR-Bench) based on the principle of error uniqueness.
ChartGalaxy: A Dataset for Infographic Chart Understanding and Generation
Zhen Li (Tsinghua University), Shixia Liu (Tsinghua University)
GenerationData SynthesisRetrievalLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed a dataset named ChartGalaxy containing 2.1 million information charts (including 61,833 real and 1,701,356 synthetic charts), and conducted three tasks based on this dataset: chart understanding, code generation, and example-based generation;
Charts Are Not Images: On the Challenges of Scientific Chart Editing
Li Li (University of Southern California), Yue Zhao (Texas A&M University)
Data SynthesisLarge Language ModelVision Language ModelImageTextBenchmark
🎯 What it does: Proposed and implemented a large-scale scientific figure editing benchmark called FigEdit, defining five task categories: single-step editing, multi-step editing, dialog editing, vision-guided editing, and style transfer, and generating 30,836 real data-driven editing instances based on the Vega/Vega-Lite renderer.
Chasing the Tail: Effective Rubric-based Reward Modeling for Large Language Model Post-Training
Junkai Zhang (University of California, Los Angeles), Lifeng Jin (Scale AI, Inc.)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBiomedical DataBenchmarkFinance Related
🎯 What it does: This paper proposes a rubric-based reward model to address the problem of reward over-optimization during the late training of large language models, and improves the accuracy of the reward model in the high-value tail through an iterative differentiated approach.
ChatInject: Abusing Chat Templates for Prompt Injection in LLM Agents
Hwan Chang (Chung-Ang University), Hwanhee Lee (Chung-Ang University)
Adversarial AttackLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the ChatInject attack, leveraging LLM chat templates and multi-round dialogue structures to achieve indirect prompt injection, bypassing instruction levels and security defenses.
CheckMate! Watermarking Graph Diffusion Models in Polynomial Time
Roberto Gheda (TU Delft), Lydia Y. Chen
GenerationGraph Neural NetworkDiffusion modelGraph
🎯 What it does: Proposes the CheckWate framework, which embeds checkerboard watermarks during sampling in graph diffusion models and achieves polynomial-time verification.
ChemEval: A Multi-level and Fine-grained Chemical Capability Evaluation for Large Language Models
Yuqing Huang (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Large Language ModelPrompt EngineeringMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed and released ChemEval, a multi-level fine-grained evaluation benchmark covering 62 chemical tasks, including text, images, and spectrograms, addressing the gaps in existing benchmarks regarding chemical depth and multi-modal assessment.
Chessformer: A Unified Architecture for Chess Modeling
Daniel Monroe, Ashton Anderson
Explainability and InterpretabilityKnowledge DistillationTransformerSequential
🎯 What it does: Propose Chessformer, a unified transformer architecture for board games, which enhances the strength of chess engines, accurately simulates moves of human players at different skill levels, and achieves model interpretability.
Children's Intelligence Tests Pose Challenges for MLLMs? KidGym: A 2D Grid-Based Reasoning Benchmark for MLLMs
Hengwei Ye (ShanghaiTech University), Zheng Tian (ShanghaiTech University)
ImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Designed and released KidGym, a 2D grid-based multimodal large language model evaluation benchmark, comprising 12 randomly generated layout tasks covering five core capabilities: execution, perceptual reasoning, memory, learning, and planning.
ChinaTravel: An Open-Ended Travel Planning Benchmark with Compositional Constraint Validation for Language Agents
Jie-Jing Shao (Nanjing University), Yu-Feng Li (Nanjing University)
TransformerLarge Language ModelAgentic AITextTabularBenchmark
🎯 What it does: Proposed ChinaTravel, a China tourism planning benchmark that integrates open-ended natural language queries, compositional DSL constraint verification, and a real-world tourism sandbox environment.
Choices Speak Louder than Questions
Gyeongje Cho (Seoul National University), Jaejin Lee (Seoul National University)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper explores and quantifies the 'choice sensitivity' of large language models in multiple-choice question answering (MCQA), and proposes a new evaluation method called Normalized Probability Shift by the Question (NPSQ) to more accurately assess the model's understanding of the question.
CHROMA: Consistent Harmonization of Multi-View Appearance via Bilateral Grid Prediction
Jisu Shin (Huawei), Eduardo Pérez-Pellitero (Huawei)
Image HarmonizationTransformerNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Proposed a Transformer-based bilateral grid prediction network called CHROMA for globally consistent photometric alignment in multi-view images caused by camera ISP, which can seamlessly integrate into reconstruction pipelines such as 3DGS without scene-specific training.
ChronoEdit: Towards Temporal Reasoning for In-Context Image Editing and World Simulation
Jay Zhangjie Wu (NVIDIA), Huan Ling (NVIDIA)
GenerationKnowledge DistillationDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderImageVideoBenchmark
🎯 What it does: Reformulate the image editing task as two-frame video generation and introduce a brief temporal reasoning phase during inference to achieve physically consistent image editing.
ChronoPlay: A Framework for Modeling Dual Dynamics and Authenticity in Game RAG Benchmarks
Liyang He (Tencent), Shiwei Tong (Tencent)
RetrievalTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the ChronoPlay framework to achieve automated, continuous generation of game RAG benchmarks, addressing dual dynamics of knowledge evolution and player interest drift, as well as authenticity issues.
Chunking the Critic: A Transformer-based Soft Actor-Critic with N-Step Returns
Dong Tian (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
TransformerReinforcement LearningSequential
🎯 What it does: Proposed a Transformer-cascaded Soft Actor-Critic (T-SAC), designing the critic as a sequence-conditional network and training with N-step returns without importance sampling.
CIAR: Interval-based Collaborative Decoding for Image Generation Acceleration
Keming Ye (Zhejiang University), Shengyu Zhang (Zhejiang University)
GenerationComputational EfficiencyTransformerImage
🎯 What it does: Propose the CIAR framework, combining cloud and device-side collaborative autoregressive image generation. The device side identifies trustworthy tokens through self-verification, while the cloud only verifies uncertain tokens, significantly accelerating the generation process.
CIMemories: A Compositional Benchmark For Contextual Integrity In LLMs
Niloofar Mireshghallah (Meta), Kamalika Chaudhuri (Meta)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the CIMemories benchmark, using synthetic user profiles (over 100 attributes) and multi-task contexts to evaluate LLMs' contextual integrity when handling information streams with persistent memories.
CineTrans: Learning to Generate Videos with Cinematic Transitions via Masked Diffusion Models
Xiaoxue Wu (Fudan University), Xinyuan Chen (Shanghai Artificial Intelligence Laboratory)
GenerationDiffusion modelVideoTextMultimodality
🎯 What it does: Propose the CineTrans framework, which utilizes mask control in video diffusion models to generate multi-shot cinematic transition videos
Circuit Insights: Towards Interpretability Beyond Activations
Elena Golimblevskaia (Fraunhofer Heinrich Hertz Institute), Sebastian Lapuschkin (Fraunhofer Heinrich Hertz Institute)
Explainability and InterpretabilityTransformerText
🎯 What it does: Proposes two automated interpretable methods, WeightLens and CircuitLens, based on the internal structure of transformers, for explaining features without relying on large amounts of data or external LLMs.
CircuitNet 3.0: A Multi-Modal Dataset with Task-Oriented Augmentation for AI-Driven Circuit Design
Mingjun Wang (Chinese Academy of Sciences), Huawei Li (Chinese Academy of Sciences)
Data SynthesisOptimizationKnowledge DistillationGraph Neural NetworkTransformerImageTextMultimodalityGraphTabularBenchmark
🎯 What it does: Proposed and released CircuitNet 3.0, a complete and multimodal open-source IC design dataset from RTL to layout, with data augmentation achieved through syntax tree rewriting and task-oriented filtering.
CircuitSense: A Hierarchical MLLM Benchmark Bridging Visual Comprehension and Symbolic Reasoning in Engineering Design Process
Arman Akbari (Northeastern University), Xuan Zhang (Northeastern University)
Data SynthesisTransformerLarge Language ModelImageTextMultimodalityBenchmarkPhysics Related
🎯 What it does: Proposes the CircuitSense benchmark to evaluate the capabilities of multimodal large language models in circuit visual understanding and symbolic reasoning.
Cite Pretrain: Retrieval-Free Knowledge Attribution for Large Language Models
Yukun Huang (Duke University), Bhuwan Dhingra (Duke University)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Train large language models to provide verifiable internal citations when answering questions without relying on retrieval;
CityLens: Evaluating Large Vision-Language Models for Urban Socioeconomic Sensing
Tianhui Liu (Hong Kong University of Science and Technology), Pan Hui (Hong Kong University of Science and Technology)
TransformerVision Language ModelImageTabularBenchmarkChain-of-Thought
🎯 What it does: CityLens proposes a multi-city, multi-indicator benchmark for urban socio-economic perception, evaluating the ability of large vision-language models to predict indicators such as economy, education, crime, transportation, health, and environment using satellite and street view images.
CitySeeker: How Do VLMs Explore Embodied Urban Navigation with Implicit Human Needs?
Siqi Wang (Polytechnic University Shenzhen Research Institute), Haofen Wang (Tongji University)
Vision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the CitySeeker benchmark to evaluate the ability of Vision-Language Models (VLMs) to perform embedded navigation in open urban environments based on implicit human needs, and presents three improvement strategies (Backtracking, Spatial Cognition Enrichment, Memory-Based Retrieval) inspired by human cognition to enhance model performance.
CL-DPS: A Contrastive Learning Approach to Blind Nonlinear Inverse Problem Solving via Diffusion Posterior Sampling
Linfeng Ye (University of Toronto), Konstantinos N. Plataniotis (University of Toronto)
RestorationDiffusion modelContrastive LearningImage
🎯 What it does: Propose CL-DPS, a contrastive learning-based diffusion posterior sampling framework for blind nonlinear inverse problem solving;
CLAP: Unsupervised 3D Representation Learning for Fusion 3D Perception via Curvature Sampling and Prototype Learning
Runjian Chen (University of Hong Kong), Ping Luo (University of Hong Kong)
Autonomous DrivingRepresentation LearningNeural Radiance FieldImageMultimodalityPoint Cloud
🎯 What it does: Proposes CLAP, a joint unsupervised 3D pre-training method based on differentiable rendering;
CLARC: C/C++ Benchmark for Robust Code Search
Kaicheng Wang (University of Southern California), Weihang Wang (University of Southern California)
RetrievalLarge Language ModelTextBenchmark
🎯 What it does: Proposed CLARC, a C/C++ code retrieval benchmark containing compilable real code snippets and automatically generated queries, along with the design of multiple robustness testing setups.
ClarifyVC: Clarifying Ambiguous Commands in Vehicle Control with a Hybrid Data Augmentation Pipeline
Hange Zhou (Hong Kong University of Science and Technology), Yongqi Zhang (Hong Kong University of Science and Technology)
Data SynthesisAutonomous DrivingTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Constructed the ClarifyVC unified framework, including a hybrid data augmentation pipeline (ClarifyVC-Data), reference models based on this data (ClarifyVC-Models), and a three-tier evaluation protocol (ClarifyVC-Eval), achieving clarification and execution of ambiguous commands in vehicle voice control.
CLASH: Evaluating Language Models on Judging High-Stakes Dilemmas from Multiple Perspectives
Ayoung Lee (University of Michigan), Lu Wang (University of Michigan)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Construct and evaluate the CLASH dataset to examine complex scenarios involving value judgments, emotional discomfort, and value transfer from a role-based perspective in high-risk dilemmas.
CLAUSE: Agentic Neuro-Symbolic Knowledge Graph Reasoning via Dynamic Learnable Context Engineering
Yang Zhao (Independent Researcher), Dusit Niyato (Nanyang Technological University)
Computational EfficiencyReinforcement LearningAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a learnable, agent-based neuro-symbolic framework called CLAUSE for dynamically constructing knowledge graph context under constraints of edit, step, and token budgets, thereby improving the accuracy of multi-hop knowledge graph question answering while reducing latency and cost.
CLEAR: Calibrated Learning for Epistemic and Aleatoric Risk
Ilia Azizi (University of Lausanne), Bin Yu (University of California, Berkeley)
Tabular
🎯 What it does: Designed a dual-parameter calibration framework CLEAR to adaptively combine and calibrate the error uncertainties of prior and posterior in regression tasks, generating more reliable conditional prediction intervals.
CLIP Behaves like a Bag-of-Words Model Cross-modally but not Uni-modally
Darina Koishigarina (University of Tubingen), Seong Joon Oh (University of Tubingen)
Representation LearningTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Investigate CLIP's attribute-object binding capability in multi-object scenes, proving that single-modal embeddings already encode binding information, and propose a linear transformation LABCLIP to correct cross-modal alignment, thereby enhancing CLIP's compositional reasoning performance.
CLIP-FMoE: Scalable CLIP via Fused Mixture-of-Experts with Enforced Specialization
Luong Tran (FPT Software AI Center), Van Nguyen (FPT Software AI Center)
ClassificationRetrievalTransformerMixture of ExpertsContrastive LearningImageTextMultimodality
🎯 What it does: Propose CLIP-FMoE, combining Mixture-of-Experts with Fusion Gate to build an expandable CLIP model that maintains zero-shot capability;
Clipped Gradient Methods for Nonsmooth Convex Optimization under Heavy-Tailed Noise: A Refined Analysis
Zijian Liu (New York University)
Optimization
🎯 What it does: Focusing on non-smooth convex optimization problems with heavy-tailed noise, the authors conduct a refined analysis of the gradient clipping (Clipped SGD) method, providing faster high-probability convergence rates. They further prove that the convergence rate in expectation can break existing lower bounds, ultimately achieving optimal expected convergence results that match the lower bounds.
CLoD-GS: Continuous Level-of-Detail via 3D Gaussian Splatting
Zhigang Cheng (Tsinghua University), Peng Pan (Tsinghua University)
Computational EfficiencyGaussian Splatting
🎯 What it does: Introduce a continuous level of detail (CLoD) mechanism for 3D Gaussian Splattering (3DGS), achieving gradual detail transitions by learning distance attenuation parameters on each Gaussian primitive, and maintaining high-quality rendering from arbitrary viewpoints through virtual distance scaling training.
CloDS: Visual-Only Unsupervised Cloth Dynamics Learning in Unknown Conditions
Yu-Liang Zhan, Hao Sun (Renmin University of China)
GenerationData SynthesisGraph Neural NetworkGaussian SplattingVideo
🎯 What it does: In unknown environments, we achieve unsupervised learning of fabric dynamics using multi-view videos, constructing the CloDS framework.
Closed-form $\ell_r$ norm scaling with data for overparameterized linear regression and diagonal linear networks under $\ell_p$ bias
Shuofeng Zhang (University of Oxford), Ard A. Louis (University of Oxford)
Optimization
🎯 What it does: This paper studies the scaling laws of various ℓ_r norms of the minimal ℓ_p interpolator in overparameterized linear regression and diagonal linear networks as the sample size changes, and provides closed-form high-probability predictions;
Closing the Gap Between Text and Speech Understanding in LLMs
Santiago Cuervo (Universit' e de Toulon), Zakaria Aldeneh (Apple)
Domain AdaptationKnowledge DistillationTransformerLarge Language ModelTextAudio
🎯 What it does: Studied and quantified the understanding gap between large language models (LLMs) under text and speech inputs, and proposed a sample-efficient training framework (SALAD) that combines cross-modal distillation with active data selection to migrate text LLMs to the speech domain.
Closing the Modality Gap Aligns Group-Wise Semantics
Eleonora Grassucci (Sapienza University of Rome), Danilo Comminiello (Sapienza University of Rome)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a novel contrastive learning loss to bridge the 'modality gap' between multimodal data, thereby enhancing group semantic structure.
Closing the Safety Gap: Surgical Concept Erasure in Visual Autoregressive Models
Xinhao Zhong (Harbin Institute of Technology), Ke Xu (Tsinghua University)
GenerationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageText
🎯 What it does: Proposed a concept elimination framework called VARE for visual autoregressive models (VAR), and designed S-VARE based on it, achieving precise concept elimination through filtered cross-entropy loss and retention loss.
CLUE: Conflict-guided Localization for LLM Unlearning Framework
Hang Chen (Xi'an Jiaotong University), Wenya Wang (Nanyang Technological University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Fine-grained localization of forgetting and retention nodes in LLMs is achieved through circuit discovery and Boolean satisfiability solving, followed by a two-stage fine-tuning approach to realize efficient and controllable forgetting.
Clustering by Denoising: Latent plug-and-play diffusion for single-cell embeddings
Dominik Meier (Cornell Tech), Kyra Gan (Cornell Tech)
ClassificationRepresentation LearningDiffusion modelBiomedical Data
🎯 What it does: Proposed a plug-and-play denoising framework called DICE that utilizes diffusion models in a low-dimensional latent space for clustering and cell type identification in single-cell RNA-seq data.
CLUTCH: Contextualized Language model for Unlocking Text-Conditioned Hand motion modelling in the wild
Balamurugan Thambiraja (Technical University of Darmstadt), Justus Thies (Technical University of Darmstadt)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelAuto EncoderVideoText
🎯 What it does: Propose the CLUTCH model, which utilizes large language models to achieve text-controlled 3D hand motion generation and description, and constructs the 32K real-world hand motion 3D-HIW dataset through a two-stage automated annotation pipeline.
CMPhysBench: A Benchmark for Evaluating Large Language Models in Condensed Matter Physics
Weida Wang (Shanghai Artificial Intelligence Laboratory), Hongming Weng (Beijing National Laboratory for Condensed Matter Physics and Institute of Physics, Chinese Academy of Sciences)
TransformerLarge Language ModelTextBenchmarkPhysics Related
🎯 What it does: Proposed CMPhysBench benchmark containing 520 open-ended computational problems in condensed matter physics to evaluate LLMs' reasoning and computational capabilities in this domain.
CMT-Benchmark: A Benchmark for Condensed Matter Theory Built by Expert Researchers
Haining Pan (Rutgers University), Eun-Ah Kim (Cornell University)
TextBenchmarkPhysics Related
🎯 What it does: Created the CMT-Benchmark, containing 50 research-level condensed matter theory problems authored by experts, and developed an automated machine grading pipeline capable of parsing and evaluating answers.
CMT: Mid-Training for Efficient Learning of Consistency, Mean Flow, and Flow-Map Models
Zheyuan Hu (Sony AI), Stefano Ermon (Stanford University)
GenerationComputational EfficiencyDiffusion modelFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Propose an intermediate training (CMT) framework to insert a lightweight, trajectory-consistent initialization step between diffusion model pre-training and flow mapping (Consistency/Mean Flow) post-training;
Co-LoRA: Collaborative Model Personalization on Heterogeneous Multi-Modal Clients
Minhyuk Seo (KU Leuven), Tinne Tuytelaars (KU Leuven)
Federated LearningMultimodalityBenchmark
🎯 What it does: This paper proposes FedMosaic, addressing data and model heterogeneity in multi-modal federated learning, enhancing personalization and generalization capabilities.
Co-occurring Associated REtained concepts in Diffusion Unlearning
Miso Kim (Dongguk University), Woojin Lee (Dongguk University)
GenerationVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: For the concept removal task in diffusion models, the ReCARE framework is proposed, which automatically constructs and utilizes the CARE-set (i.e., the co-occurring concept vocabulary that needs to be retained) and introduces retention loss and erasure loss during the removal process, thereby maintaining the generation capability of beneficial co-occurring concepts while removing the target concept; simultaneously, the CARE score metric is proposed to quantify the retention degree of co-occurring concepts.
Co-rewarding: Stable Self-supervised RL for Eliciting Reasoning in Large Language Models
Zizhuo Zhang (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Proposed a self-supervised reinforcement learning framework called Co-Rewarding to enhance the reasoning capabilities of large language models and prevent training collapse;
CoAct-1: Computer-using Multi-agent System with Coding Actions
Linxin Song (University of Southern California), Caiming Xiong (Salesforce)
AI Code AssistantTransformerLarge Language ModelAgentic AIVision Language ModelBenchmark
🎯 What it does: Proposed and implemented CoAct-1, a multi-agent system that integrates traditional GUI operations with code execution, dynamically assigning subtasks to GUI Operator or Programmer via Orchestrator to accomplish complex computer usage tasks.
Coarse-to-Fine Learning of Dynamic Causal Structures
Dezhi Yang (Shandong University), Guoxian Yu (Shandong University)
OptimizationRepresentation LearningConvolutional Neural NetworkScore-based ModelTime Series
🎯 What it does: This paper proposes the DyCausal framework for learning fully dynamic time series causal structures from coarse to fine;
CoCoDiff: Correspondence-Consistent Diffusion Model for Fine-grained Style Transfer
Wenbo Nie (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: Developed a training-free, low-cost fine-grained style transfer framework called CoCoDiff, which leverages pre-trained latent diffusion models to explore pixel-level semantic correspondence relationships and guides generation through cycle consistency, achieving structure-preserving style transfer.
CoDA: Agentic Systems for Collaborative Data Visualization
Zichen Chen (University Of California Santa Barbara), Jinsung Yoon (Google)
Data-Centric LearningAI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the CoDA multi-agent system, which automatically generates high-quality data visualization code from natural language queries.
CoDA: From Text-to-Image Diffusion Models to Training-Free Dataset Distillation
Letian Zhou (National University of Singapore), Xinchao Wang (National University of Singapore)
Knowledge DistillationDiffusion modelAuto EncoderImageText
🎯 What it does: Propose the CoDA framework, which utilizes an offline text-to-image diffusion model to achieve training-free Dataset Distillation.
Code Aesthetics with Agentic Reward Feedback
Bang Xiao (Microsoft Research Asia), Furu Wei (Microsoft Research Asia)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIImageTextBenchmark
🎯 What it does: This paper introduces the concept of code aesthetics, constructs the AesCode-358K dataset and OpenDesign benchmark, and enhances the executability and aesthetic quality of visual code generation through an agentic reward framework combined with the GRPO algorithm for large language models (LLMs).
Code Driven Planning with Domain-Adaptive Selector
Zikang Tian (State Key Lab of Processors Institute of Computing Technology Chinese Academy of Sciences), Yunji Chen (State Key Lab of Processors Institute of Computing Technology Chinese Academy of Sciences)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: Propose the CoPiC framework, which generates multiple planning programs using LLMs and selects the plan that best aligns with long-term rewards through a domain adaptation selector, thereby reducing LLM query costs and improving planning quality.
Code World Models for General Game Playing
Wolfgang Lehrach (Google DeepMind), Kevin Patrick Murphy (Google DeepMind)
TransformerLarge Language ModelAuto EncoderWorld ModelTextSequential
🎯 What it does: Translate natural language rules and game trajectories into an executable 'Code World Model' (CWM) using large language models (LLMs), then use this model for efficient planning and gameplay.
Code2Bench: Scaling Source and Rigor for Dynamic Benchmark Construction
Zhe Zhang (Beihang University), Hailong Sun (Beihang University)
Large Language ModelTextGraphBenchmark
🎯 What it does: Proposed the CODE2BENCH framework, which adopts a dual expansion approach (dynamically acquiring real-world repository code and high-coverage Property-Based Testing) to construct the CODE2BENCH-2509 benchmark for evaluating large language models' code generation capabilities.
CodeBrain: Bridging Decoupled Tokenizer and Multi-Scale Architecture for EEG Foundation Model
Jingying Ma (National University of Singapore), Mengling Feng (National University of Singapore)
ClassificationRepresentation LearningTransformerAuto EncoderContrastive LearningBiomedical Data
🎯 What it does: Designed and trained a two-stage EEG foundation model called CodeBrain, leveraging a time-frequency separated tokenizer and a multi-scale architecture to achieve interpretable and efficient EEG representation learning.
CodeGenGuard: A Watermark for Code Generation Models
Borui Yang (Shanghai Jiao Tong University), Xinghao Jiang (Shanghai Jiao Tong University)
GenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextSequential
🎯 What it does: Proposed CodeGenGuard, a digital watermarking framework for code generation models to verify model ownership.
CodeQuant: Unified Clustering and Quantization for Enhanced Outlier Smoothing in Low-Precision Mixture-of-Experts
Xiangyang Yin (Courant Institute of Mathematical Sciences New York University), Sai Qian Zhang (Courant Institute of Mathematical Sciences New York University)
Computational EfficiencyMixture of ExpertsText
🎯 What it does: Proposes CodeQuant, a unified quantization and clustering framework specifically designed for low-precision Mixture-of-Experts (MoE) models;
CodeSense: a Real-World Benchmark and Dataset for Code Semantic Reasoning
Monoshi Kumar Roy (Iowa State University), Wei Le (Iowa State University)
Large Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the CodeSense benchmark by collecting code extracted from 744 real-world Python, C, and Java projects, designed multi-level fine-grained code semantic reasoning tasks, and automatically generated corresponding execution traces and annotations.
CoDi: Subject-Consistent and Pose-Diverse Text-to-Image Generation
Zhanxin Gao (Nanjing University), Ying Tai (Nanjing University)
GenerationVision Language ModelDiffusion modelMultimodality
🎯 What it does: Proposes a training-free two-stage framework named CoDi, which utilizes optimal transport (OT) in the early denoising stage to achieve identity feature transfer, and adopts selective cross-image attention in the later stage for detail refinement, thereby maintaining subject consistency while preserving pose and layout diversity.
Codified Finite-state Machines for Role-playing
Letian Peng (University of California San Diego), Jingbo Shang (University of California San Diego)
GenerationExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextSequential
🎯 What it does: Propose Codified Finite-State Machine (CFSM) and its probabilistic version (CPFSM), which automatically extract key states from text-based role profiles using LLM and generate executable state transition code to achieve explainable tracking and generation of role states;
CoEmoGen: Towards Semantically-Coherent and Scalable Emotional Image Content Generation
Kaishen Yuan (Hong Kong University of Science and Technology), Yutao Yue (Hong Kong University of Science and Technology)
GenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Proposed the CoEmoGen framework to achieve Emotion-Driven Image Content Generation (EICG), capable of generating images with semantically clear and emotionally credible content based on specified emotional categories;
CoFact: Conformal Factuality Guarantees for Language Models under Covariate Shift
Zirui Hu (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
Domain AdaptationExplainability and InterpretabilityTransformerText
🎯 What it does: Designed and implemented a conformal prediction framework called CoFact, which provides statistical guarantees on the factualness of text generated by large language models (LLMs) without relying on the exchangeability assumption required by traditional methods, and is applicable to scenarios with continuous covariate shift.
CogFlow: Bridging Perception and Reasoning through Knowledge Internalization for Visual Mathematical Problem Solving
Shuhang Chen (Zhejiang University), Hangjie Yuan (Zhejiang University)
Reinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed a three-stage cognitive heuristic framework called COGFLOW to address the drift between perception and reasoning in visual math reasoning;
CogMoE: Signal-Quality–Guided Multimodal MoE for Cognitive Load Prediction
Aamir Bader Shah (University of Houston), Xin Fu (University of Houston)
ClassificationTransformerMixture of ExpertsMultimodalityTime Series
🎯 What it does: Proposed the CogMoE framework, utilizing signal quality guided Mixture-of-Experts (MoE) to achieve multimodal cognitive load prediction. First, time-frequency synchronization and recovery are applied to eliminate signal alignment and missing data issues, followed by dynamic routing to experts specifically designed for high-fidelity, noisy, and missing scenarios based on estimated signal quality.
CogniLoad: A Synthetic Natural Language Reasoning Benchmark With Tunable Length, Intrinsic Difficulty, and Distractor Density
Daniel Kaiser (Integreat Norwegian Centre for knowledge-driven machine learning), Benjamin Ricaud (Integreat Norwegian Centre for knowledge-driven machine learning)
Data SynthesisLarge Language ModelTextBenchmark
🎯 What it does: Propose CogniLoad, a controllable synthetic natural language reasoning benchmark based on Cognitive Load Theory (CLT);
CogniMap3D: Cognitive 3D Mapping and Rapid Retrieval
Feiran Wang (University of Illinois Chicago), Yan Yan (University of Illinois Chicago)
Depth EstimationRetrievalOptimizationConvolutional Neural NetworkSimultaneous Localization and MappingOptical FlowVideoPoint Cloud
🎯 What it does: Proposed the CogniMap3D framework to achieve 3D scene understanding and reconstruction from dynamic videos, with capabilities for persistent spatial memory and fast retrieval.
Cognitive models can reveal interpretable value trade-offs in language models
Sonia Krishna Murthy, Tomer Ullman (Harvard University)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper quantifies LLM behavior in polite language scenarios using the Rational Speech Acts (RSA) cognitive model, investigating the impact of reasoning budget, system prompts, and post-training on model trade-offs between values.
CoLA: Co-Calibrated Logit Adjustment for Long-Tailed Semi-Supervised Learning
Qian Shao (Zhejiang University), Jian Wu (Zhejiang Key Laboratory Of Medical Imaging Artificial Intelligence)
ClassificationMeta LearningImage
🎯 What it does: Proposes a co-calibration logit adjustment method called CoLA for long-tailed semi-supervised learning, addressing bias in pseudo-label generation and distribution estimation errors;
COLD-Steer: Steering Large Language Models via In-Context One-step Learning Dynamics
Kartik Sharma (Georgia Institute of Technology), Rakshit Trivedi (Massachusetts Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes COLD-Steer, a training-free and sample-efficient activation layer intervention framework that dynamically approximates gradient updates through first-order learning on context examples, enabling direct regulation of LLM's intermediate representations during inference to achieve desired behaviors; two efficient implementations are provided: Unit Kernel (COLD-Kernel) and Finite Difference (COLD-FD);
Collaborative Gym: A Framework for Enabling and Evaluating Human-Agent Collaboration
Yijia Shao (Stanford University), Diyi Yang (Stanford University)
Robotic IntelligenceLarge Language ModelAgentic AITextTabularSequentialBenchmark
🎯 What it does: Created and released the Collaborative Gym (Co-Gym) framework for implementing and evaluating human-robot collaboration in dual-control task environments.
CollectiveKV: Decoupling and Sharing Collaborative Information in Sequential Recommendation
Jingyu Li (Sun Yat-sen University), Pengwen Dai (Sun Yat-sen University)
Recommendation SystemSequential
🎯 What it does: Propose the CollectiveKV mechanism, which significantly compresses the cache and reduces inference latency by sharing KV caches across users in sequential recommendation.
CoLLMLight: Cooperative Large Language Model Agents for Network-Wide Traffic Signal Control
Zirui Yuan (Hong Kong University of Science and Technology (Guangzhou)), Hao Liu (Hong Kong University of Science and Technology (Guangzhou))
Autonomous DrivingOptimizationLarge Language ModelReinforcement LearningAgentic AIGraphChain-of-Thought
🎯 What it does: Propose the CoLLMLight framework, achieving network-level traffic signal control through asynchronous collaborative decision-making.
Color3D: Controllable and Consistent 3D Colorization with Personalized Colorizer
Yecong Wan (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
Image TranslationGenerationSupervised Fine-TuningDiffusion modelNeural Radiance FieldGaussian SplattingImageVideo
🎯 What it does: Proposed the Color3D framework, which achieves controllable, colorful, and consistent 3D colorization from monochrome inputs to static or dynamic 3D scenes by fine-tuning a personalized colorizer on a single key view and then consistently propagating colors to all other views and time frames.
COMAL: A Convergent Meta-Algorithm for Aligning LLMs with General Preferences
Yixin Liu (Yale University), Arman Cohan (Yale University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes COMAL, a meta-algorithm for aligning large language models (LLMs) under a general preference framework. It constructs a two-player zero-sum game model and achieves alignment by iteratively solving KL-regularized zero-sum subgames.
CoMAS: Co-Evolving Multi-Agent Systems via Interaction Rewards
Xiangyuan Xue (Chinese University of Hong Kong Shanghai Artificial Intelligence Laboratory), LEI BAI
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Introduce the CoMAS framework, enabling LLM agents to self-evolve through mutual discussion and evaluation in multi-agent systems without requiring external rewards.
Combination-of-Experts with Knowledge Sharing for Cross-Task Vehicle Routing Problems
Zikang Yu (Sun Yat-sen University), Jiahai Wang (Sun Yat-sen University)
OptimizationKnowledge DistillationTransformerMixture of ExpertsBenchmark
🎯 What it does: Proposes the CoEKS model to achieve zero-shot generalization across tasks in vehicle routing problems (VRP), capable of handling any combination of constraints.
Combinatorial Bandit Bayesian Optimization for Tensor Outputs
Jingru Huang (Tsinghua University), Chen Zhang (Tsinghua University)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a Bayesian optimization framework for tensor outputs (TOBO) and further extends it to combinatorial bandit Bayesian optimization for partially observable tensor outputs (TOCBBO).
Combinatorial Rising Bandits
Seockbean Song (POSTECH), Jungseul Ok (POSTECH)
OptimizationReinforcement Learning
🎯 What it does: Proposed the Combinatorial Rising Game Framework (CRB) and designed the Combinatorial UCB algorithm CRUCB that can handle the enhancement effects of base arms;
CoMem: Compositional Concept-Graph Memory for Vision–Language Adaptation
Heng Zhou (National University Of Singapore), Jiawei Yao (National University Of Singapore)
RetrievalDomain AdaptationGraph Neural NetworkMixture of ExpertsContrastive LearningMultimodality
🎯 What it does: We propose the COMEM framework, which achieves continuous visual-language learning using concept graph memory and feature space replay.
ComGS: Efficient 3D Object-Scene Composition via Surface Octahedral Probes
Jian Gao (Nanjing University), Yao Yao (Nanjing University)
Image HarmonizationGenerationDiffusion modelGaussian SplattingImagePoint Cloud
🎯 What it does: Propose the ComGS framework to achieve real-time 3D object-scene synthesis across multiple stages (reconstruction → editing → rendering), significantly enhancing shadow realism and visual harmony.
COMI: Coarse-to-fine Context Compression via Marginal Information Gain
Jiwei Tang (Tsinghua University), Bo Zheng (Alibaba)
Computational EfficiencyLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Designed and implemented the COMI framework, which employs a coarse-to-fine hierarchical context compression approach. It utilizes Marginal Information Gain (MIG) to measure the relevance and redundancy of context units, dynamically allocating compression rates and performing weighted merging during compression.
CoMind: Towards Community-Driven Agents for Machine Learning Engineering
Sijie Li (Peking University), Yiming Yang (Carnegie Mellon University)
Data-Centric LearningLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Proposed a community knowledge-sharing-based ML engineering evaluation framework MLE-Live and a multi-agent system CoMind, capable of automating machine learning engineering workflows in a simulated Kaggle community;
Command-V: Training-Free Representation Finetuning Transfer
Barry Wang (Carnegie Mellon University), Daphne Ippolito (Carnegie Mellon University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a no-training, no-gradient backpropagation cross-model behavior transfer method ⌘V that directly migrates pre-trained adapters to differently structured LLMs.
Common Corpus: The Largest Collection of Ethical Data for LLM Pre-Training
Pierre-Carl Langlais (PleIAs), Ivan P. Yamshchikov (PleIAs)
Safty and PrivacyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Created a 'Common Corpus' open multilingual corpus with approximately 2 trillion tokens, providing complete sources, licenses, and quality processing tools.
Communication-Efficient Decentralized Optimization via Double-Communication Symmetric ADMM
Jinrui Huang, Runxiong Wu (University of Science and Technology of China)
OptimizationFederated LearningTabular
🎯 What it does: Propose a double-communication symmetric ADMM algorithm (DS-ADMM), which performs two rounds of neighbor communication in each iteration to more efficiently solve decentralized composite optimization problems without centralized coordination.
Compactness and Consistency: A Conjoint Framework for Deep Graph Clustering
Wei Ju (Sichuan University), Jiancheng Lv (Sun Yat-sen University)
Representation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: Proposed the CoCo framework for unsupervised deep graph clustering, which learns node representations from both local and global perspectives, then eliminates redundancy and performs consistency learning through low-rank reconstruction, ultimately achieving more compact and semantically rich embeddings.
Comparing AI Agents to Cybersecurity Professionals in Real-World Penetration Testing
Justin W Lin, Daniel E. Ho (Stanford University)
TransformerLarge Language ModelAgentic AIPrompt Engineering
🎯 What it does: This study compared the penetration testing performance of ten cybersecurity professionals with six existing AI agents and the self-developed multi-agent framework ARTEMIS on real-world enterprise networks;
Comparing the learning dynamics of in-context learning and fine-tuning in language models
Basile Confavreux (University College London), Andrew M Saxe (University College London)
ClassificationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Compared the learning dynamics and internal representations of in-context learning (ICL) and supervised fine-tuning (SFT) in pre-trained language models for 2D linear classification tasks using the same data, sequence, and task.
COMPASS: Robust Feature Conformal Prediction for Medical Segmentation Metrics
Matt Y. Cheung (Rice University), Guha Balakrishnan (Rice University)
SegmentationConvolutional Neural NetworkImageBiomedical DataUltrasound
🎯 What it does: This paper proposes the COMPASS framework, which constructs conformal prediction intervals based on downstream metrics (e.g., target area) by applying linear perturbations in the low-dimensional sensitive subspace of neural network intermediate representations.
CompassNav: Steering From Path Imitation to Decision Understanding In Navigation
LinFeng Li (East China Normal University), Xuelong Li (Institute of Artificial Intelligence TeleAI China Telecom)
OptimizationRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelTextMultimodality
🎯 What it does: Propose the CompassNav framework, achieving a transition from path imitation to decision understanding, and training a high-performance goal navigation model on a large-scale LVLM.
ComPhy: Composing Physical Models with end-to-end Alignment
Alessandro Trenta (University of Pisa), Davide Bacciu (University of Pisa)
Physics Related
🎯 What it does: Proposes ComPhy, a modular, end-to-end aligned framework for solving multi-equation partial differential equation (PDE) systems.
Complementing Self-Consistency with Cross-Model Disagreement for Uncertainty Quantification
Kimia Hamidieh, Marzyeh Ghassemi (MIT)
Explainability and InterpretabilityLarge Language ModelText
🎯 What it does: Proposes a total uncertainty (TU) evaluation method that combines model internal consistency (aleatoric) and cross-model semantic inconsistency (epistemic), quantifying the reliability of generated results through black-box access to LLM outputs.
Completed Hyperparameter Transfer across Modules, Width, Depth, Batch and Duration
Bruno Kacper Mlodozeniec (Apple), marco cuturi (Apple)
OptimizationHyperparameter SearchTransformerTextStochastic Differential Equation
🎯 What it does: Proposed a new complete(d) parameterization rule to enable hyperparameter transfer across width, depth, batch size, and training duration, and demonstrated that hyperparameter optimization at the module level can also transfer.