CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― 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.
π― 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.
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)
CodeTransformerVision Language ModelImageTabularBenchmarkChain-of-Thought
π― 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.
π― What it does: In unknown environments, we achieve unsupervised learning of fabric dynamics using multi-view videos, constructing the CloDS framework.
Closing the Gap Between Text and Speech Understanding in LLMs
Santiago Cuervo (Universit' e de Toulon), Zakaria Aldeneh (Apple)
CodeDomain AdaptationKnowledge DistillationTransformerLarge Language ModelTextAudio
π― 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)
CodeRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodality
π― 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)
CodeGenerationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageText
π― 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.
Clustering by Denoising: Latent plug-and-play diffusion for single-cell embeddings
Dominik Meier (Cornell Tech), Kyra Gan (Cornell Tech)
CodeClassificationRepresentation LearningDiffusion modelBiomedical Data
π― 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.
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)
CodeTransformerLarge Language ModelTextBenchmarkPhysics Related
π― 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.
π― 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)
CodeFederated LearningMultimodalityBenchmark
π― 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)
CodeGenerationVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
π― 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)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
π― 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)
CodeAI Code AssistantTransformerLarge Language ModelAgentic AIVision Language ModelBenchmark
π― 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.
Code2Bench: Scaling Source and Rigor for Dynamic Benchmark Construction
Zhe Zhang (Beihang University), Hailong Sun (Beihang University)
CodeLarge Language ModelTextGraphBenchmark
π― 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)
CodeClassificationRepresentation LearningTransformerAuto EncoderContrastive LearningBiomedical Data
π― 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.
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)
CodeComputational EfficiencyMixture of ExpertsText
π― What it does: Proposes CodeQuant, a unified quantization and clustering framework specifically designed for low-precision Mixture-of-Experts (MoE) models;
CoDi: Subject-Consistent and Pose-Diverse Text-to-Image Generation
Zhanxin Gao (Nanjing University), Ying Tai (Nanjing University)
CodeGenerationVision Language ModelDiffusion modelMultimodality
π― 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.
Letian Peng (University of California San Diego), Jingbo Shang (University of California San Diego)
CodeGenerationExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextSequential
π― 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)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageText
π― 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;
CogMoE: Signal-QualityβGuided Multimodal MoE for Cognitive Load Prediction
Aamir Bader Shah (University of Houston), Xin Fu (University of Houston)
CodeClassificationTransformerMixture of ExpertsMultimodalityTime Series
π― 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.
CogniMap3D: Cognitive 3D Mapping and Rapid Retrieval
Feiran Wang (University of Illinois Chicago), Yan Yan (University of Illinois Chicago)
CodeDepth EstimationRetrievalOptimizationConvolutional Neural NetworkSimultaneous Localization and MappingOptical FlowVideoPoint Cloud
π― 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.
CodeExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
π― 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.
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)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
π― 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)
CodeRobotic IntelligenceLarge Language ModelAgentic AITextTabularSequentialBenchmark
π― What it does: Created and released the Collaborative Gym (Co-Gym) framework for implementing and evaluating human-robot collaboration in dual-control task environments.
CoMAS: Co-Evolving Multi-Agent Systems via Interaction Rewards
Xiangyuan Xue (Chinese University of Hong Kong Shanghai Artificial Intelligence Laboratory), LEI BAI
CodeOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― 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)
CodeOptimizationKnowledge DistillationTransformerMixture of ExpertsBenchmark
π― 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.
COMI: Coarse-to-fine Context Compression via Marginal Information Gain
Jiwei Tang (Tsinghua University), Bo Zheng (Alibaba)
CodeComputational EfficiencyLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― 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)
CodeData-Centric LearningLarge Language ModelAgentic AITextRetrieval-Augmented Generation
π― 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;
π― 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)
CodeTransformerLarge Language ModelAgentic AIPrompt Engineering
π― 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)
CodeClassificationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― 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.
Completing Missing Annotation: Multi-Agent Debate for Accurate and Scalable Relevant Assessment for IR Benchmarks
Minjeong Ban (Korea Advanced Institute of Science and Technology), Hwanjun Song (Korea Advanced Institute of Science and Technology)
CodeRetrievalTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
π― What it does: Design and implement a multi-agent debate framework called DREAM for automated relevance assessment in information retrieval, and construct the BRIDGE benchmark based on this framework to fill missing text fragments.
CompoDistill: Attention Distillation for Compositional Reasoning in Multimodal LLMs
Jiwan Kim (KAIST), Chanyoung Park (KAIST)
CodeKnowledge DistillationTransformerSupervised Fine-TuningVision Language ModelMultimodality
π― What it does: This paper addresses the problem of knowledge distillation failing to effectively enhance visual perception capabilities in multimodal large language models (MLLMs), proposing the CompoDistill framework. Through two modules, visual attention alignment (VAT) and teacher adapter grasping (TAF), it significantly improves the student model's performance in compositional reasoning tasks while maintaining visual question answering (VQA) performance.
Compositional Generalization from Learned Skills via CoT Training: A Theoretical and Structural Analysis for Reasoning
Xinhao Yao (Renmin University of China), Yong Liu (Renmin University of China)
CodeTransformerTextChain-of-Thought
π― What it does: This paper explains how Chain-of-Thought (CoT) training enhances the compositional generalization ability of large language models in both in-distribution (ID) and out-of-distribution (OOD) scenarios through theoretical derivation and structural analysis, and verifies its effectiveness on controllable synthetic data and real mathematical problems.
π― What it does: Propose a novel transformer architecture called Abduction Transformer, which achieves combinatorial generalization capability for various discrete abstract reasoning tasks by leveraging non-parametric mixture distribution latent spaces, information theory regularization, and gradient search during testing.
Computer Agent Arena: Toward Human-Centric Evaluation and Analysis of Computer-Use Agents
Bowen Wang (The University of Hong Kong), Tao Yu (The University of Hong Kong)
CodeReinforcement Learning from Human FeedbackLarge Language ModelAgentic AITextMultimodalityBenchmark
π― What it does: Developed an open-source platform named COMPUTER AGENT ARENA for anonymous side-by-side, human-centric evaluation of Computer-Using Agents (CUA) in real-world cloud-diverse environments;
ComputerRL: Scaling End-to-End Online Reinforcement Learning for Computer Use Agents
Hanyu Lai (Tsinghua University), Jie Tang (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
π― What it does: Built a complete computer usage intelligent agent framework named COMPUTERRL, capable of performing complex tasks through API and GUI dual-mode interaction in a desktop environment, achieving end-to-end online reinforcement learning.
Concept-based Adversarial Attack: a Probabilistic Perspective
Andi Zhang (University of Warwick), Samuel Kaski (University of Manchester)
CodeAdversarial AttackLarge Language ModelSupervised Fine-TuningDiffusion modelImage
π― What it does: Proposed and implemented a concept-level adversarial attack framework that generates diverse adversarial samples from concept distributions rather than single images.
Confident and Adaptive Generative Speech Recognition via Risk Control
Amit Damri (Tel Aviv University), Bracha Laufer-Goldshtein (Tel Aviv University)
CodeRecognitionLarge Language ModelAudio
π― What it does: Propose an adaptive generative speech recognition error correction framework based on risk control, dynamically determining the required number of N-best options for each input to achieve efficient and reliable post-processing.
π― What it does: Propose a method named CBDS-IMVC, which utilizes confidence block diagonal structure and invariant graph completion techniques to reconstruct missing multi-view data and achieve clustering.
π― What it does: Built a unified contrastive learning framework ConRep4CO, which converts multi-type graph decision problems into SAT, using the corresponding SAT graph as positive samples and other SAT graphs as negative samples to pre-train cross-domain generalizable graph representations.
π― What it does: This paper proposes the Consistency-driven Calibration and Matching (ConCM) framework to address knowledge conflicts in Few-Shot Class Incremental Learning (FSCIL). It eliminates prototype bias through memory-aware prototype calibration and achieves dual consistency between features and structure via dynamic structural matching.
David Woodruff, Samson Zhou (Texas A&M University)
CodeOptimizationComputational EfficiencyTabular
π― What it does: This paper proposes and studies the problem of consistent low-rank approximation under row-stream input, presenting two algorithms that maintain (1+Ξ΅) approximation accuracy while achieving lower recurrence cost than traditional online algorithms.
Consistent Text-to-Image Generation via Scene De-Contextualization
Song Tang (University of Shanghai for Science and Technology), Xiatian Zhu (University of Surrey)
CodeGenerationTransformerImageTextMultimodality
π― What it does: This paper proposes a training-agnostic scene decontextualization method called SDeC, which maintains identity consistency in text-to-image generation without prior knowledge of all scenes.
π― What it does: Introduce MaskGRPO to achieve reinforcement learning optimization for multimodal discrete diffusion models, combining modality-specific importance estimation and sampling methods.
Constrained Decoding of Diffusion LLMs with Context-Free Grammars
Niels MΓΌndler (ETH Zurich), Martin Vechev (ETH Zurich)
CodeGenerationAI Code AssistantLarge Language ModelDiffusion modelText
π― What it does: Proposes the first restricted decoding method for diffusion language models (DLMs), enabling the generation of words in arbitrary order while maintaining syntactic constraints and supporting context-free grammar (CFG) constraints.
Contamination Detection for VLMs Using MultiβModal Semantic Perturbations
Jaden Park (University of Wisconsin-Madison), Yong Jae Lee (University of Wisconsin-Madison)
CodeAnomaly DetectionLarge Language ModelVision Language ModelDiffusion modelMultimodality
π― What it does: Addressing the test set leakage issue in Vision-Language Models, we propose a detection method based on multimodal semantic perturbations, and evaluate model generalization differences under varying pollution levels.
π― What it does: Proposed Content-Aware Mamba for Image Compression (CMIC), which enhances the rate-distortion performance of image compression by incorporating content-adaptive token ordering and global prior hints into the Mamba state space model.
Context and Diversity Matter: The Emergence of In-Context Learning in World Models
Fan Wang (Shenzhen Institute of Artificial Intelligence and Robotics for Society), Yu Kang (University of Science and Technology of China)
CodeTransformerAuto EncoderWorld ModelImage
π― What it does: Investigated in-context learning (ICL) in world models, theoretically derived and empirically validated two mechanisms: environment recognition (ER) and environment learning (EL), and proposed an efficient long-context linear attention world model, L2World, demonstrating its cross-environment adaptability.
Xingyuan Hua (Tsinghua University), Ju Ren (Tsinghua University)
CodeOptimizationTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
π― What it does: Propose M2CL, a context learning method designed for multi-agent discussion (MAD), which dynamically generates and evolves context instructions for each LLM during the discussion process.
Context Tokens are Anchors: Understanding the Repeat Curse in dMLLMs from an Information Flow Perspective
Qiyan Zhao (Shanghai Jiao Tong University), Da-Han Wang (Xiamen University of Technology)
CodeComputational EfficiencyTransformerLarge Language ModelVision Language ModelDiffusion modelMultimodality
π― What it does: This paper studies the repeat curse problem that occurs during cache acceleration in diffusion-based multimodal LLM (dMLLM) from the information flow perspective, and proposes the CoTA method to alleviate it.
π― What it does: Propose ContextGen, a multi-instance generation framework based on Diffusion Transformer, capable of achieving precise layout control and identity consistency simultaneously.
ContextIF: Enhancing Instruction-Following through Context Reward
Yule Zhong (East China Normal University), Guoxiu He (East China Normal University)
CodeReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
π― What it does: Propose ContextIF, which automatically generates high-quality context (constraint summary + examples) for any instruction using reinforcement learning, thereby enhancing the instruction-following capability of LLMs.
ContextPRM: Leveraging Contextual Coherence for multi-domain Test-Time Scaling
Haotian Zhang (Beihang University), Xianglong Liu (Beihang University)
CodeDomain AdaptationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes ContextPRM, a Process Reward Model based on context coherence evaluation, to enhance test-time generalization performance in multi-domain reasoning.
π― What it does: Proposed a prediction-based dual-branch model called CS-LSTMs for detecting subtle anomalies and chronic trend anomalies in univariate time series;
Continual Low-Rank Adapters for LLM-based Generative Recommender Systems
Hyunsik Yoo (University of Illinois Urbana-Champaign), Hanghang Tong (University of Illinois Urbana-Champaign)
CodeRecommendation SystemTransformerLarge Language ModelText
π― What it does: This study investigates continuous learning methods for generative recommendation systems based on large language models, proposing a PESO algorithm that incorporates proximal regularization into a single LoRA adapter to address issues of catastrophic forgetting in traditional single adapters and the excessive rigidity of cumulative adapters.
π― What it does: Propose a Continuous Token Chain-of-Thought (CoT2) that enables models to parallelly track multiple reasoning trajectories within a single inference path, significantly enhancing reasoning efficiency and accuracy.
Continuum Transformers Perform In-Context Learning by Operator Gradient Descent
Yash Patel (University of Michigan), Ambuj Tewari (University of Michigan)
CodeOptimizationRepresentation LearningMeta LearningTransformerPhysics Related
π― What it does: Investigated the theoretical mechanisms of continuous Transformers in context-free learning, proving their equivalence to gradient descent in operator RKHS.
π― What it does: Proposes Contraction Homology (CH), novel Backward PH, Forward-Backward (FB) persistence, and 'Hourglass persistence,' extending them to simplicial and cell complexes; designs an efficient algorithm embeddable into end-to-end differentiable graph neural networks.
CodeSafty and PrivacyTransformerLarge Language ModelTextChain-of-Thought
π― What it does: Proposed and quantified the concept of 'control tax' to measure the operational costs and safety benefits generated by integrating AI supervision mechanisms (such as monitoring models) into practical systems.
π― What it does: Proposes a controllable first-frame guided video editing framework utilizing mask-aware LoRA fine-tuning, enabling flexible and fine-grained control over the entire video editing process without modifying the model architecture.
Controllable Logical Hypothesis Generation for Abductive Reasoning in Knowledge Graphs
Yisen Gao (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningGraph
π― What it does: Studied controlled inductive reasoning in knowledge graphs, proposing the CtrlHGen framework to generate logical hypotheses that comply with semantic and structural constraints.
ConvT3: Structured State Kernels for Convolutional State Space Models
Jaeyoung Hong (SolverX), Minseon Gwak (University of Massachusetts Amherst)
CodeGenerationConvolutional Neural NetworkVideoPhysics Related
π― What it does: This paper proposes the ConvT3 model, which expands the state kernel of ConvSSM from 1Γ1 to 3Γ3 to more comprehensively capture spatiotemporal dynamics.
π― What it does: Propose Cooperative Sheaf Neural Networks (CSNN), a neural network based on a cellular layered structure with directed graph orientation, enabling nodes to adaptively decide on cooperative behaviors for information transmission and reception;
Copy-Paste to Mitigate Large Language Model Hallucinations
Yongchao Long (Tianjin University of Technology), Shenda Hong (Peking University)
CodeRetrievalExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: Propose a 'Copy-Paste' generation paradigm, directly copying fragments from retrieved contexts to generate answers, to enhance contextual faithfulness in Retrieval-Augmented Generation (RAG) systems; and construct a two-stage training framework: Stage-1 generates high-repetition-rate candidate answers through three prompts (CP-Order, CP-Link, CP-Refine); Stage-2 performs Direct Preference Optimization (DPO) using automatically generated high-repetition-rate preference dialogues, resulting in CopyPasteLLM.
CoRA: Boosting Time Series Foundation Models for Multivariate Forecasting through Correlation-aware Adapter
Hanyin Cheng (East China Normal University), Chenjuan Guo (East China Normal University)
CodeComputational EfficiencyTransformerSupervised Fine-TuningContrastive LearningTime Series
π― What it does: Propose a lightweight CoRA plugin that significantly enhances multivariate time series forecasting performance during the fine-tuning stage by leveraging the internal representations and prediction results of TSFM.
CORE: Concept-Oriented Reinforcement for Bridging the DefinitionβApplication Gap in Mathematical Reasoning
Zijun Gao (University of Illinois Urbana Champaign), Ben Zhou (Arizona State University)
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: Proposes the CORE framework, enhancing the concept recognition and application capabilities of LLMs in mathematical reasoning through concept-oriented reinforcement learning.
Cost-of-Pass: An Economic Framework for Evaluating Language Models
Mehmet Hamza Erol (Stanford University), James Zou (Stanford University)
CodeComputational EfficiencyTransformerText
π― What it does: This paper introduces the 'cost-of-pass' metric, which combines the accuracy of language models with inference costs to measure the economic efficiency of models. Based on this metric, it defines frontier cost, tracks technological progress, analyzes the contributions of model families, and evaluates the economic value of inference technologies.
π― What it does: Proposes a chain-of-thought (CoT) distillation framework called COT-EVO based on evolutionary algorithms, aiming to generate high-quality, domain-specific scientific reasoning paths and train smaller models.
CounselBench: A Large-Scale Expert Evaluation and Adversarial Benchmarking of Large Language Models in Mental Health Question Answering
Yahan Li, Ruishan Liu (University Of Southern California)
CodeAdversarial AttackTransformerLarge Language ModelTextBenchmark
π― What it does: Constructed a large-scale mental health Q&A benchmark COUNSELBENCH, integrating expert-evaluated LLM responses and adversarial questionnaires;
Count Counts: Motivating Exploration in LLM Reasoning with Count-based Intrinsic Rewards
Xuan Zhang (Fudan University), Yuan Qi (Fudan University)
CodeLarge Language ModelReinforcement LearningText
π― What it does: Propose the MERCI algorithm, which promotes exploration and diversity in multi-step reasoning by designing a counting-based intrinsic reward mechanism for the LLM inference process;
Counterfactual Reasoning for Retrieval-Augmented Generation
Huaiyu Qin (Renmin University of China), Yunhai Wang (Renmin University of China)
CodeGenerationRetrievalExplainability and InterpretabilityTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Propose the CF-RAG framework, achieving causal reasoning in retrieval-augmented generation through counterfactual exploration and parallel arbitration, thereby escaping the relevance trap.
Coupled Transformer Autoencoder for Disentangling Multi-Region Neural Latent Dynamics
Ram Dyuthi Sristi (UC San Diego), Gal Mishne (UC San Diego)
CodeTransformerAuto EncoderTime SeriesBiomedical Data
π― What it does: Propose Coupled Transformer Autoencoder (CTAE), which learns nonlinear, non-stationary, long-range temporal latent variables in multi-region neural recordings through Transformer encoder/decoder, while orthogonally separating shared and private dynamics within the same framework.
CP-Agent: ContextβAware Multimodal Reasoning for Cellular Morphological Profiling under Chemical Perturbations
Yuxin Zhang (University of Hong Kong), Kevin Kin-Man Tsia (University of Hong Kong)
CodeExplainability and InterpretabilityDrug DiscoveryLarge Language ModelAgentic AIVision Language ModelContrastive LearningMultimodalityBiomedical Data
π― What it does: Proposed CP-Agent, an intelligent agent combining context-aware CLIP with a multimodal large language model to generate explainable reports on drug-induced cell morphology changes.
CPQS-Tuning: A Model Self-Perception-Based Data Filtering Algorithm for Efficient Instruction Fine-Tuning
Yi Ren (Nanjing University), Diandong Liu (Shaanxi University of Science & Technology)
CodeComputational EfficiencyData-Centric LearningConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
π― What it does: This paper proposes a self-perception data filtering method called CPQS-Tuning based on the hidden states of large language models (LLMs) to enhance the efficiency and effectiveness of instruction fine-tuning.
CodeAI Code AssistantLarge Language ModelReinforcement LearningAgentic AITextBenchmark
π― What it does: Designed and implemented USACOArena, an interactive coding arena based on ACM-ICPC, integrating programming tasks with a unified credit economy, requiring agents to make decisions under a limited budget.
Critique-RL: Training Language Models For Critiquing Through Two-Stage Reinforcement Learning
Zhiheng Xi (Fudan University), Xuanjing Huang (Fudan University)
CodeReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
π― What it does: Proposed and implemented a two-stage reinforcement learning framework called Critique-RL for training critical language models that can evaluate and provide constructive feedback.
π― What it does: Propose a unified continuous-time multi-view prediction framework named CRONOS, capable of predicting target volumes at any arbitrary time from multiple historical scans in 3D medical sequences.
CodeRepresentation LearningVision Language ModelAuto EncoderMultimodality
π― What it does: Study the geometric structure of shared embedding spaces in vision-language models, propose the Iso-Energy hypothesis, and extract cross-modal concepts through aligned sparse autoencoders.
CrossPL: Systematic Evaluation of Large Language Models for Cross Programming Language Interoperating Code Generation
zhanhang xiong, Wenhai Wang (Zhejiang University)
CodeGenerationTransformerLarge Language ModelTextBenchmark
π― What it does: Created the CrossPL benchmark to evaluate LLMs' ability in cross-language interoperability (IPC and FFI) code generation, and systematically assessed the performance of 20 LLMs on this benchmark.
CryoNet.Refine: A One-step Diffusion Model for Rapid Refinement of Structural Models with Cryo-EM Density Map Restraints
Fuyao Huang (State Key Laboratory of Membrane Biology-Membrane Structure and Artificial Intelligence Biology Branch), Qiangfeng Cliff Zhang (Tsinghua University)
CodeOptimizationProtein Structure PredictionTransformerDiffusion modelBiomedical Data
π― What it does: Developed a first-order diffusion model, CryoNet.Refine, to rapidly approximate experimental cryo-EM density maps from initial atomic models and improve geometric structures.
CodeProtein Structure PredictionGaussian SplattingBiomedical Data
π― What it does: Propose a self-consistent cryo-EM homogeneous reconstruction method called cryoSplat, which enables physically accurate three-dimensional potential reconstruction from raw particle images through direct initialization.
Lixuan Guo (Stony Brook University), Chenyu You (Stony Brook University)
CodeComputational EfficiencyRepresentation LearningSupervised Fine-TuningAuto EncoderContrastive LearningImageTextBiomedical Data
π― What it does: Propose CSRv2, a training framework for extremely sparse embeddings (kβ€4), achieving efficient and high-quality sparse embeddings through progressive k-annealing, supervised sparse contrastive learning, and optional full model fine-tuning.
CTBench: Cryptocurrency Time Series Generation Benchmark
Yihao Ang (National University of Singapore), Zhiyong Huang (National University of Singapore)
CodeGenerationData SynthesisDiffusion modelFlow-based ModelAuto EncoderGenerative Adversarial NetworkTime SeriesBenchmarkFinance Related
π― What it does: Designed CTBench, a benchmark framework for cryptocurrency time series generation, providing data, dual-task evaluation, and multidimensional financial metrics.
CTC-DRO: Robust Optimization for Reducing Language Disparities in Speech Recognition
Martijn Bartelds (Stanford University), Karen Livescu (Toyota Technological Institute at Chicago)
CodeRecognitionOptimizationTransformerAudio
π― What it does: Propose a group distribution robust optimization algorithm called CTC-DRO for multi-voice recognition, addressing the subgroup imbalance issue caused by the incomparability of CTC loss in traditional Group DRO.
π― What it does: Propose a pluggable plugin called CUPID, which can jointly estimate the aleatoric and epistemic uncertainties of deep learning models without modifying or retraining the base model.
CurES: From Gradient Analysis to Efficient Curriculum Learning for Reasoning LLMs
Yongcheng Zeng (Institute of Automation, Chinese Academy of Sciences), Jun Wang (University College London)
CodeComputational EfficiencyLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: Propose an adaptive curriculum learning framework called CurES based on gradient analysis and Bayesian posterior inference to enhance the training efficiency of reasoning large language models (LLMs).
π― What it does: This paper proposes E2H Reasoner, a post-training method for large language models (LLMs) based on curriculum reinforcement learning (CRL), which enhances the reasoning ability of language models through task scheduling from easy to difficult.
Customizing Visual Emotion Evaluation for MLLMs: An Open-vocabulary, Multifaceted, and Scalable Approach
Daiqing Wu (Chinese Academy of Sciences), Yu ZHOU
CodeLarge Language ModelVision Language ModelImageMultimodalityBenchmark
π― What it does: Propose the Emotion Statement Judgment task and its automated annotation pipeline INSETS, constructing and publicly releasing the MVEI benchmark to evaluate the ability of multimodal large language models (MLLM) in visual emotion understanding.
Cyber-Zero: Training Cybersecurity Agents without Runtime
Terry Yue Zhuo (Monash University), Zijian Wang (Meta Superintelligence Labs)
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
π― What it does: Proposed the CYBER-ZERO framework, which generates high-quality proxy trajectories without a runtime environment by leveraging publicly available CTF writing to train LLM performance in cybersecurity tasks.
CyclicReflex: Improving Reasoning Models via Cyclical Reflection Token Scheduling
Chongyu Fan (Michigan State University), Sijia Liu (Michigan State University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Proposed a training-free decoding strategy called CyclicReflex, which dynamically balances thinking and decision-making by periodically adjusting reflection tokens during the reasoning process.
π― What it does: Proposes CylinderSplat, a dual-branch feedforward panoramic 3D Gaussian profile framework capable of generating realistic novel view images from single-view or sparse-view inputs.