International Conference on Machine Learning Β· 1032 papers
Beyond Independence: Learning Correlated Views for Variational Incomplete Multi-View Clustering
Zheming Xu (Beijing Jiaotong University), Michael Kampffmeyer (UiT Arctic University of Norway)
CodeOptimizationRepresentation LearningData-Centric LearningTransformerMixture of ExpertsAuto EncoderContrastive LearningMultimodality
π― What it does: This paper proposes a novel variational framework called ACOVA for incomplete multi-view clustering, which can enhance clustering performance by learning cross-view correlations even in the absence of missing view information.
π― What it does: Propose the TACO framework, which utilizes cross-individual and cross-modal topological consistency for self-supervised pre-training on 3D multi-modal medical imaging.
Beyond Logits: Coherent Hallucination Mitigation via Attention Contrastive Decoding
Yujia Chen (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
CodeGenerationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper studies the hallucination problem that occurs during the generation process of large-scale vision-language models (LVLMs), and proposes Attention Contrastive Decoding (ACD) as a training-free plug-in. It transfers the operations of traditional contrastive decoding from the logit layer to the attention layer, and further introduces the Adaptive Subtraction Strategy (ASS) to achieve position-adaptive suppression.
π― What it does: Propose an end-to-end short text clustering framework that utilizes consistency-aware adaptive optimal transport (CAOT) to generate reliable pseudo labels, and constructs semantic similarity between samples through an instance-level attention network, achieving the unification of global structure and local consistency.
π― What it does: Designed a retrieval-enhanced model improvement framework called M-DESIGN, which can quickly approach the optimal neural network architecture through fine-grained structural modifications under a limited evaluation budget.
Beyond Next-Token Alignment: Distilling Multimodal Large Language Models via Token Interactions
Lin Chen (MAIS, Institute of Automation, Chinese Academy of Sciences), Shiming Xiang (MAIS, Institute of Automation, Chinese Academy of Sciences)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: Proposes an AlignβTI framework based on token interaction for knowledge distillation, aiming to compress large-scale multi-modal language models (MLLM) into parameter-efficient small models.
CodeAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Constructed the UOJ-Bench benchmark to evaluate the code generation, code hacking (generating test cases that make opponents' code fail), and code repair (providing minimal patches) capabilities of large language models (LLMs) in competitive programming, and validated it directly using the native judging interface of UOJ.
π― What it does: This paper reinterprets the oversmoothing problem in GNNs through dynamic bifurcation theory, and demonstrates that by using activation functions with stable cubic nonlinearities (such as Sine, tanh) and bifurcation-aware initialization, feature homogenization can be avoided in deep GNNs (up to 64 layers), leading to the emergence of stable non-uniform patterns. Subsequently, polynomial spectral filters are used to further control the topological patterns selected by the model, and the effectiveness of this method is validated on multiple node classification benchmarks.
π― What it does: This paper proposes a dynamic workflow construction framework called SquRL based on reinforcement learning, for the text-to-SQL (Text-to-SQL) task;
Beyond Temperature: Hyperfitting as a Late-Stage Geometric Expansion
Meimingwei Li (LMU Munich), Christian Heumann (LMU Munich)
CodeGenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: This paper reveals that the 'overfitting' process, where LLMs are fine-tuned to extremely low training error, significantly enhances the diversity of generated text and reduces repetition rates, which is attributed to an adaptive vocabulary ranking reordering mechanism;
Beyond Token-level Supervision: Unlocking the Potential of Decoding-based Regression via Reinforcement Learning
Ming Chen (Nanjing University), Chao Qian (Nanjing University)
CodeOptimizationKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringScore-based ModelContrastive LearningTextTabularSequentialBenchmark
π― What it does: This paper proposes a framework called GenRe 2, which combines decoding regression with reinforcement learning, to address the mismatch between traditional token-level supervision based on cross-entropy and continuous numerical targets.
Beyond Tokens: Enhancing RTL Quality Estimation via Structural Graph Learning
Yi Liu (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
CodeOptimizationKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextGraph
π― What it does: Proposes a structured self-supervised learning framework called StructRTL based on control data flow graphs (CDFG), aimed at improving RTL design quality estimation;
Beyond Trajectory-Level Attribution: Graph-Based Credit Assignment for Agentic Reinforcement Learning
Xin Cheng (Nanyang Technological University), Bo An (Nanyang Technological University)
CodeOptimizationExplainability and InterpretabilityGraph Neural NetworkTransformerReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelImageTextGraphBenchmark
π― What it does: Construct a unified state transition graph, and achieve fine-grained step-level reward allocation based on the graph's advantage estimation, thereby improving the reinforcement learning training of LLM/VLM agents.
Beyond Two-Stage Training: Cooperative SFT and RL for LLM Reasoning
Liang Chen (Chinese University of Hong Kong), Kam-Fai Wong (Chinese University of Hong Kong)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextBenchmark
π― What it does: Proposes a new training framework called BRIDGE, which leverages the idea of second-order optimization to enable supervised fine-tuning (SFT) to actively supervise reinforcement learning (RL) in order to enhance the reasoning capabilities of large language models (LLMs).
Gongye Liu (Hong Kong University of Science and Technology), Wenhan Luo (Hong Kong University of Science and Technology)
CodeGenerationOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringMixture of ExpertsVision Language ModelDiffusion modelScore-based ModelContrastive LearningImageMultimodality
π― What it does: Proposed DiNa-LRM, a diffusion model-based latent space reward model that performs preference learning directly in noisy states using noise-calibrated Thurstone likelihood;
π― What it does: Proposed a BiasβSpectrum Neural Processes (BSNP) framework for building fast and reliable surrogate models for parameterized PDEs under sparse and irregular observations.
Biases in the Blind Spot: Detecting What LLMs Fail to Mention
IvΓ‘n Arcuschin (Poseidon Research), Oana-Maria Camburu (Imperial College London)
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposed a fully automated, black-box pipeline for detecting implicit biases in large language models that are not verbalized during chain-of-thought reasoning.
π― What it does: Proposed a bidirectional trajectory augmentation framework called BiTrajDiff, which generates forward and backward trajectories on shared anchor states using a bidirectional diffusion model, and concatenates them to produce global trajectories across behavior patterns.
π― What it does: Proposed an unordered quantization reinforcement learning framework to solve black-box combinatorial optimization problems using neural network generators;
π― What it does: Propose a framework called BOCLOAK based on optimal transport for edge editing and node injection attacks targeting social bot detection in graph neural networks.
Break the Block: Dynamic-size Reasoning Blocks for Diffusion Large Language Models via Monotonic Entropy Descent with Reinforcement Learning
Yan Jiang (University of Queensland), Zi Huang (University of Queensland)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningDiffusion modelTextBenchmark
π― What it does: By introducing the b1 framework into the post-training of diffusion large language models (dLLM), the model can learn dynamic-sized reasoning blocks, thereby enhancing the coherence and accuracy of the reasoning process.
π― What it does: A new optimistic actor-critic algorithm (OptAC) is proposed, specifically designed for low-rank Markov decision processes (MDPs), which relies solely on a policy evaluation oracle, avoiding the computationally complex planning or optimization oracles common in previous methods.
Breaking the Factorization Barrier in Diffusion Language Models
Ian Li (University of California, San Diego), Anji Liu (National University of Singapore)
CodeGenerationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelDiffusion modelScore-based ModelContrastive LearningText
π― What it does: Propose the CoDD framework, which couples discrete diffusion language models with tractable probabilistic circuits (Probabilistic Circuits), breaking the traditional independence assumption and allowing the joint distribution of multiple words to be modeled in a single denoising step.
π― What it does: This paper proposes to extract the low-frequency components (i.e., 'learngene') of pre-trained model weights using discrete cosine transform, achieving one-time cross-scale model initialization without training;
Breaking the Simplification Bottleneck in Amortized Neural Symbolic Regression
Paul Saegert (Heidelberg University), Ullrich Koethe
CodeOptimizationComputational EfficiencyRepresentation LearningTransformerReinforcement LearningPrompt EngineeringAuto EncoderContrastive LearningTabularTime SeriesSequentialPhysics Related
π― What it does: Proposes a rule-based matching SIMPLIPY simplification engine and the FLASH-ANSR training framework to significantly improve the speed and quality of expressions in amortized neural symbolic regression.
π― What it does: Proposes the Brep2Shape self-supervised pre-training framework, achieving alignment between abstract parameters and intuitive shapes by mapping B-rep control points to dense spatial points.
Bridging On-Device and Cloud LLMs for Collaborative Reasoning: A Unified Methodology for Local Routing and Post-Training
Wenzhi Fang (Purdue University), Christopher Brinton
CodeFederated LearningComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
π― What it does: Achieve collaborative reasoning between edge devices and cloud-based LLMs, training a local model to autonomously decide whether to invoke the cloud model during inference.
Bridging the Knowledge-Prediction Gap in LLMs on Multiple-Choice Questions
Yoonah Park (Seoul National University), Yohan Jo (Seoul National University)
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
π― What it does: Investigate the knowledge-prediction gap of LLMs on multiple-choice questions, analyze their geometric structure, and propose the KAPPA intervention method during reasoning to align the knowledge subspace with the prediction subspace.
π― What it does: Propose Argument-induced 3D Point Loss (A3PL) and Grammar-constrained Operator (GCO), achieving three-dimensional geometric supervision for each parameter token in CAD program sequences, and enforcing syntactic constraints during sequence generation, thereby enhancing the accuracy and effectiveness of generated geometry.
Chao Han (Eastern Institute of Technology), Xiaoyu Shen (Eastern Institute of Technology)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
π― What it does: Proposes a dynamic computational allocation framework based on a lightweight feature predictor (LFF), improving traditional greedy routing to achieve efficient inference for large language models.
π― What it does: Proposes BTSP-CAM, a gradient-free binary memory module based on the brain's BTSP mechanism, for sample-free class-incremental learning.
π― What it does: During the iterative sampling process of diffusion models, BudCache is proposed to fix the NFE budget through an offline search caching strategy, thereby achieving predictable inference latency and maximizing the final generation quality.
π― What it does: Proposed a hierarchical attack called BHEA under budget constraints and an adversarial training framework called BHEA-AT using this attack, aiming to enhance the robustness of collaborative multi-agent reinforcement learning.
Building Social World Model with Large Language Models
Haofei Yu (University of Illinois Urbana-Champaign), Jiaxuan You (University of Illinois Urbana-Champaign)
CodeTransformerLarge Language ModelPrompt EngineeringWorld ModelTextTime SeriesBenchmarkFinance RelatedRetrieval-Augmented Generation
π― What it does: Propose the Social World Model (SWM), which leverages large language models to capture the dynamics of social beliefs as they evolve with events;
BYORn: Bootstrap Your Own Responses to Defend Large Vision-Language Models Against Backdoor Attacks
Ivan Sabolic, Sven Loncaric
CodeAdversarial AttackTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Propose BYORn, which detects poisoned samples using low likelihood judgment from pre-trained vision-language models, and dynamically generates clean responses with the model itself during training to achieve robust instruction fine-tuning.
π― What it does: Proposes CADFit, a hybrid optimization framework that uses geometry-driven optimization and program execution to recover editable CAD construction sequences from meshes or images.
π― What it does: Study how to calibrate uncertainty under adversarial attacks in zero-shot CLIP, and propose an adversarial fine-tuning method based on the Dirichlet distribution.
CALM Before the STORM: Unlocking Native Reasoning for Optimization Modeling
Zhengyang Tang (Chinese University of Hong Kong), Benyou Wang (Chinese University of Hong Kong)
CodeOptimizationKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: A local error correction and adaptive training framework called CALM was implemented on large-scale reasoning models (LRM). The model first solves problems on its own, then inserts short prompts at the first error location for local correction. Subsequently, supervised fine-tuning and reinforcement learning were used to obtain the optimized modeling expert STORM 4B.
Can LLM Agents Stick to the Script? Modeling Commitment in Interactive Narratives
Yingpeng Ma (University of Macau), Derek F. Wong (University of Macau)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
π― What it does: Propose the Narrative Commitment Preservation (NCP) task and the NCP-Bench benchmark to evaluate the logical consistency of large language models in interactive storytelling.
Can LLMs Reason Like Automated Theorem Provers for Rust Verification? VCoT-Bench: Evaluating via Verification Chain of Thought
Zichen Xie (University of Virginia), Wenxi Wang (University of Virginia)
CodeExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This study proposes the VCoT-Lift framework and the VCoT-Bench benchmark to elevate the low-level reasoning of SMT solvers into readable verification chains-of-thought, and evaluates the reasoning capabilities of LLMs in the verification of Rust programs.
π― What it does: Proposed a microcanonical Langevin dynamics sampler (SMILE) that can efficiently operate under small-batch gradient noise, and further improved its robustness through gradient noise preprocessing and adaptive scheduling based on energy variance.
π― What it does: Proposes a recursive self-improvement recommendation framework, RSIR, which allows the model to continuously enhance the training set and improve performance by generating high-fidelity interaction sequences without relying on external data or teacher models.
CodeOptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
π― What it does: Systematically sweep the Ξ² parameter in Direct Preference Optimization (DPO) to explore its impact on different capabilities (reasoning, arithmetic, formatting, empathy, etc.), revealing traps such as the decoupling of proxy metrics from true capabilities, path dependence, and probe cluster structures.
Capacitated Fair-Range Clustering: Hardness and Approximation Algorithms
Ameet Gadekar (CISPA Hemholtz Center for Information Security), Suhas Thejaswi (Aalto University)
CodeOptimizationTabularBenchmark
π― What it does: The study investigates and proposes a clustering problem that combines capacity constraints with fair range (Capacitated Fair-Range k-Clustering), and provides its theoretical complexity and approximation algorithms.
CARE: Class-Adaptive Expert Consensus for Reliable Learning with Long-Tailed Noisy Labels
Mengke Li (Shenzhen University), Hui Huang (Shenzhen University)
CodeClassificationTransformerMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Proposed a Class-Adaptive Rectification with Experts (CARE) framework, which uses multi-modal experts (text, image, original labels) to adaptively correct long-tailed noisy labels through a class-adaptive Top-K consensus mechanism and achieve long-tail calibration;
CAT-Q: Cost-efficient and Accurate Ternary Quantization for LLMs
Shigeng Wang (Intel Labs China), Anbang Yao (Intel Labs China)
CodeCompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
π― What it does: This paper proposes a post-training ternary quantization method called CAT-Q, which can compress large language models into 1.58-bit weight quantized models without the need for large-scale training data or retraining.
Causal discovery for time series with endogenous context variables
Oana-Iuliana Popescu (University of Potsdam), Jakob Runge (University of Potsdam)
CodeExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTime SeriesPhysics Related
π― What it does: This paper addresses the problem of causal discovery in time series with endogenous context variables, proposing two adaptive testing algorithms based on PCMCI+ (PAC-PCMCI+ and SAC-PCMCI+), which can recover context-specific causal graphs without assuming context exogeneity.
π― What it does: Propose a new autoregressive video diffusion model distillation process called Causal Forcing. First, train an autoregressive teacher model using teacher forcing. Then, perform ODE distillation with this teacher to ensure frame-level injectivity. Finally, further improve the model performance through asymmetric DMD.
Edwin V. Bonilla (CSIRO), Daniel M. Steinberg (CSIRO)
CodeOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningDrug DiscoveryReinforcement Learning from Human FeedbackGraph Neural NetworkMixture of ExpertsGraphTabularBiomedical DataBenchmark
π― What it does: Propose a Bayesian framework named CaPE that actively approximates the DAG posterior distribution by utilizing step-by-step feedback from experts on local causal relationships.
π― What it does: Propose an unsupervised table anomaly detection method called CausalAno, which uses causal GAN to learn the causal structure of normal data, and performs anomaly scoring based on Mahalanobis distance in the discriminator's latent space.
Causal-JEPA: Learning World Models through Object-Level Latent Masking
Heejeong Nam (Brown University), Randall Balestriero (Brown University)
CodeAutonomous DrivingRepresentation LearningRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelAuto EncoderContrastive LearningWorld ModelImageVideoSequential
π― What it does: Propose C-JEPA, an object-level masked joint embedding prediction (JEPA) world model, which learns dynamic representations without using reconstruction loss by leveraging frozen object-aware encoders (such as VideoSAUR / SAVi).
CausalX: A Unified and Causally-Interpretable Plug-and-Play Model for Multi-modal Spatio-Temporal Forecasting
Shiqi Zhang (Zhejiang University of Technology), Cong Bai (Zhejiang University of Technology)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyRecurrent Neural NetworkGraph Neural NetworkTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningMultimodalityGraphTime Series
π― What it does: Propose the CausalX model, construct a dynamic causal heuristic graph, and achieve interpretability and performance improvement in multi-modal spatiotemporal prediction.
CausalXRL: Explainable Reinforcement Learning through Causal Graph Reasoning
Yanming Zhang (Stony Brook University), Klaus Mueller (Stony Brook University)
CodeExplainability and InterpretabilityGraph Neural NetworkReinforcement LearningContrastive LearningGraphTabular
π― What it does: Propose the CausalXRL framework, which uses causal graph reasoning to provide interpretability for model-free reinforcement learning;
Bo Peng (Shanghai Jiao Tong University), Chaochao Lu (Shanghai Artificial Intelligence Laboratory)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerScore-based ModelAuto EncoderContrastive LearningGraphTabularBenchmark
π― What it does: Proposes CauScale, a neural network architecture capable of efficiently performing causal structure learning on large-scale graphs with thousands of nodes.
CauSciBench: Can LLMs Automate Causal Inference in Real-World Scientific Research?
Sawal Acharya (Jinesis Lab, University of Toronto & Vector Institute), Zhijing Jin (Jinesis Lab, University of Toronto & Vector Institute)
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes CauSciBench, a benchmark for evaluating large language models across the complete causal inference workflow (variable selection, method decision-making, implementation, and result interpretation).
Causes and Consequences of Representational Similarity in Machine Learning Models
Zeyu Michael Li (Duke University), Emily Wenger (Duke University)
CodeRepresentation LearningAdversarial AttackConvolutional Neural NetworkTransformerMixture of ExpertsContrastive LearningImageText
π― What it does: Studied the causal impact of dataset overlap and task overlap in training data on the representational similarity of machine learning models, and systematically evaluated the similarity and susceptibility to attacks across different models (such as ResNet, ViT, nanoGPT, Llama, UNet, etc.).
CCLRec: Consensus-driven Contrastive Learning for LLM-enhanced Graph Recommendation
Ting Guo (North University of China), Pinle Qin (North University of China)
CodeRecommendation SystemGraph Neural NetworkTransformerLarge Language ModelContrastive LearningMultimodalityGraph
π― What it does: Propose the CCLRec framework, which deeply integrates large language models (LLMs) with graph neural networks (GNNs), leveraging the consistency between structural and semantic views to generate high-confidence positive and negative samples;
Certified Circuits: Stability Guarantees for Mechanistic Circuits
Alaa Anani (Max Planck Institute for Informatics), Jonas Fischer (Max Planck Institute for Informatics)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImageText
π― What it does: Proposes the Certified Circuits framework, which provides provable stability guarantees for 'circuits' in neural networks under edits on concept datasets, resulting in more reliable and interpretable subnetworks.
Chain-of-Thought Reasoning In The Wild Is Not Always Faithful
IvΓ‘n Arcuschin (Poseidon Research), Arthur Conmy
CodeExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringWorld ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This paper systematically evaluates the reliability of large language models in generating Chain-of-Thought (CoT) through experiments on natural language question answering and mathematical problems, finding that models often produce unfaithful reasoning chains without explicit bias prompts, leading to inconsistencies between the final answer and the reasoning process.
π― What it does: A lightweight and pluggable channel adapter, ChaTSFM, is proposed to enable pre-trained time series foundation models (TSFM) to capture spatial dependencies among multivariate variables in a zero-shot setting, thereby improving the accuracy of multivariate time series forecasting.
Characterizing the Predictive Impact of Modalities with Supervised Latent-Variable Modeling
Divyam Madaan (New York University), Kyunghyun Cho (New York University)
CodeExplainability and InterpretabilityRepresentation LearningAuto EncoderContrastive LearningImageMultimodalityTabularBiomedical DataElectronic Health RecordsAudio
π― What it does: Proposes PRIMO, a supervised latent variable model for quantifying the impact of missing modalities on predictions in multimodal learning, and trains and infers under both complete and missing modalities.
π― What it does: The CINOC framework is proposed by modeling control policies as neural operators and end-to-end training with a differentiable PDE solver, achieving PDE control for multi-agent systems with variable sensor/actuator configurations.
Circle-RoPE: Cone-like Decoupled Rotary Positional Embedding for Vision-Language Models
Chengcheng Wang (University of Sydney), Kai Han (Huawei Noah's Ark Lab)
CodeComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: Proposes Circle-RoPE, a method based on conical decoupled rotational position encoding, which can eliminate cross-modal relative position information bias caused by the unified indexing of RoPE in vision-language models. It quantifies and ensures PTD = 0 through the Per-Token Distance (PTD) metric, thereby maintaining the spatial structure within images while achieving isometric decoupling between text and images; meanwhile, it introduces Alternating Geometry Encoding (AGE), which alternates between Circle-RoPE and traditional M-RoPE across different Transformer layers to balance cross-modal alignment and image local feature extraction.
CLASP: Online learning algorithms for Convex Losses And Squared Penalties
Ricardo N. Ferreira (NOVA School of Science and Technology), Claudia Soares
CodeOptimization
π― What it does: Proposes the CLASP (Convex Losses And Squared Penalties) framework, aiming to minimize the accumulated loss and squared constraint violations, addressing online convex optimization problems with dynamic constraints.
CodeGenerationData SynthesisTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelAuto EncoderTabularTime SeriesPhysics Related
π― What it does: Proposed ClimateAR, a self-attention generative model for probabilistic climate forecasting, combining an aligned tokenizer and multi-scale conditional mechanisms;
CLINIC : Evaluating Multilingual Trustworthiness in Language Models for Healthcare
Akash Ghosh (Indian Institute of Technology Patna), Chirag Agarwal (University of Virginia)
CodeSafty and PrivacyExplainability and InterpretabilityAdversarial AttackDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented Generation
π― What it does: Proposed the CLINIC multilingual reliability benchmark for systematically evaluating the reliability of medical language models in multilingual environments.
CLIP Tricks You: Training-free Token Pruning for Efficient Pixel Grounding in Large Vision-Language Models
Sangin Lee (Sejong University), Yukyung Choi (Sejong University)
CodeSegmentationComputational EfficiencyTransformerPrompt EngineeringVision Language ModelContrastive LearningImageVideoText
π― What it does: Developed a training-agnostic, text-guided visual token pruning method called LiteLVLM for efficient pixel-level localization inference.
Clipping Bottleneck: Stabilizing RLVR via Stochastic Recovery of Near-Boundary Signals
Shuo Yang (Peking University), Jingren Zhou (Alibaba)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
π― What it does: Propose the Near-boundary Stochastic Rescue (NSR) method, which restores the learning signals that were clipped during RLVR training by randomly retaining tokens near the boundary that exceed the hard clipping threshold, thereby improving the model's training stability and convergence performance.
Clipping Low-Probability Tokens in SFT Yields a Generalizable Initialization for RL
Tian-Shuo Liu (Nanjing University), Yang Yu (Nanjing University)
CodeData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
π― What it does: This paper proposes trimming low-probability (off-policy) tokens during the supervised fine-tuning (SFT) phase to reduce forgetting of prior knowledge, thereby obtaining a more general and exploratory RL initialization; and validates the effectiveness of this method through theoretical and experimental analysis.
Cluster-Aware Causal Mixer for Online Anomaly Detection in Multivariate Time Series
Md Mahmuddun Nabi Murad (University of South Florida), Yasin Yilmaz (University of South Florida)
CodeAnomaly DetectionRecurrent Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningGaussian SplattingTabularTime Series
π― What it does: Propose a cluster-aware causal mixer (CCM-TAD) for online anomaly detection in multivariate time series data, combining clustering embedding, causal mixing layers, and serialized anomaly scoring;
Clustering as Reasoning: A $k$-Means Interpretation of Chain-of-Thought Graph Learning
Xuanting Xie (University of Electronic Science and Technology of China), Yuan Fang (Singapore Management University)
CodeExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkTransformerPrompt EngineeringContrastive LearningTextGraphChain-of-Thought
π― What it does: Propose a unified KCOT framework that combines chain-of-thought (CoT) with graph neural networks (GNN), and theoretically equates Transformer blocks to k-means clustering, thereby interpreting CoT as an iterative process of 'clustering + updating'.
CoCoEmo: Composable and Controllable Human-Like Emotional TTS via Activation Steering
Siyi Wang (University of Melbourne), Ting Dang (University of Melbourne)
CodeGenerationData SynthesisTransformerMixture of ExpertsTextAudio
π― What it does: Proposes a hybrid emotion control framework based on activation steering, which can achieve quantitative interpolation of emotions and synthesis of text-emotion mismatch without retraining the model.
π― What it does: Propose and implement CODA-BENCH, a data-intensive Linux sandbox benchmark built within the Kaggle ecosystem that simultaneously evaluates code generation and data discovery capabilities, containing 1,009 tasks;
Code2Video: A Code-centric Paradigm for Educational Video Creation
Yanzhe Chen (National University of Singapore), Mike Zheng Shou (National University of Singapore)
CodeGenerationExplainability and InterpretabilityData-Centric LearningAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelDiffusion modelVideoTextMultimodalityBenchmark
π― What it does: Propose Code2Video, a multi-agent framework based on executable Manim code, for generating structured and interpretable educational videos.
CoEvol-NO: State and Coordinate Co-Evolution with an Error-Driven Predictor-Corrector Paradigm for Neural Operator Transformer
Jianqiao Zeng (Fudan University), Junchi Yan (Shanghai Jiao Tong University)
CodeOptimizationComputational EfficiencyTransformerPoint CloudMeshGraphBenchmarkPhysics Related
π― What it does: Designed CoEvol-NO, a linear complexity neural operator that achieves co-evolution of potential states and grid coordinate sequences through a predictor-corrector framework.
Coevolutionary Continuous Discrete Diffusion: Make Your Diffusion Language Model a Latent Reasoner
Cai Zhou (Massachusetts Institute of Technology), Dinghuai Zhang (Microsoft Research)
CodeGenerationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelMixture of ExpertsDiffusion modelScore-based ModelTextSequentialStochastic Differential EquationOrdinary Differential Equation
π― What it does: Proposed and implemented the Coevolutionary Continuous Discrete Diffusion (CCDD) language model, which jointly integrates continuous diffusion and discrete diffusion within the same network, achieving implicit reasoning and stronger expressiveness.
Cold-Start Personalization via Bayesian Adaptive Questioning
Avinandan Bose (Meta Superintelligence Labs), Asli Celikyilmaz (Meta Superintelligence Labs)
CodeRecommendation SystemReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringWorld ModelText
π― What it does: Propose CAPEn, a cold start personalization framework, which achieves efficient question selection through offline learning of preference associations followed by online Bayesian inference.
π― What it does: Propose a generative super-resolution method ASASR that combines Sobolev space for noise inpainting and adversarial Sobolev alignment, addressing the distortion of high-frequency details in traditional generative priors.
Ricardo Olmedo (Max Planck Institute for Intelligent Systems), Moritz Hardt (Max Planck Institute for Intelligent Systems)
CodeOptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextFinance RelatedRetrieval-Augmented Generation
π― What it does: This study explores the possibility of computational arbitrage in the AI model market, demonstrating how arbitrageurs can effectively allocate inference budgets across different model providers to create competitive products.
Concept Removal for Frontier Image Generative Models
Aditya Kumar (CISPA Helmholtz Center for Information Security), Franziska Boenisch (CISPA Helmholtz Center for Information Security)
CodeGenerationSafty and PrivacyExplainability and InterpretabilityTransformerPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningImageTextMultimodality
π― What it does: This paper proposes a concept elimination framework called BLOCK based on a transcoder, which directly replaces the bottleneck layer of the text-to-generator in state-of-the-art text-image generation models (such as SD3.5, Flux, Infinity-2B/8B), thereby achieving concept elimination without modifying the backend network.
Condition Number Based Low-Bit Quantization for Image Super-Resolution
Kai Liu (Shanghai Jiao Tong University), Linghe Kong (Shanghai Jiao Tong University)
CodeSuper ResolutionTransformerImage
π― What it does: Designed a low-bit quantization method called CondiQuant based on the condition number, for post-training quantization of image super-resolution models.
Conditional Coverage Diagnostics for Conformal Prediction
Sacha Braun (Inria), Francis Bach (Inria)
CodeAnomaly DetectionExplainability and InterpretabilityData-Centric LearningSupervised Fine-TuningContrastive LearningTabularTime SeriesBenchmark
π― What it does: This paper proposes a new framework that transforms conditional coverage evaluation into a classification task, and quantifies the failure of conditional coverage in conformal prediction by constructing a metric called 'Excess Risk of Target Coverage' (ERT).
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageTextMultimodality
π― What it does: By automating concept discovery and spatial localization, ConEx generates interpretable concept heatmaps and combines gradients to achieve both local and global explanations for classification.
Shuhang Lin (Rutgers University), Dimitris N. Metaxas (Rutgers University)
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
π― What it does: Proposed the Conformal Path Reasoning (CPR) framework, which achieves trustworthy reasoning for knowledge graph question answering through path-level non-conformity measures;
ConFu: Contemplate the Future for Better Speculative Sampling
Zongyue Qin (University of California Los Angeles), Yizhou Sun (University of California Los Angeles)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
π― What it does: Propose ConFu, a reasoning framework that improves the quality of draft models by 'thinking about the future,' combining pause (think) tokens, soft prompts, dynamic MoE mechanisms, and robust training;
π― What it does: Proposed a semi-consumptive method called COPEx for handling dynamic constraints in Bayesian Experimental Design (BED), combining offline pre-trained design policies and posterior networks, and performing online planning through multi-step scenario trees during runtime;
Zhong Li (Great Bay University), Zaiwen Wen (Peking University)
CodeOptimizationLarge Language ModelPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
π― What it does: This paper proposes a new benchmark, MIPLIB-NL, with large scale and complete structure, by reverse constructing natural language descriptions and executable code from real industrial MILP instances in MIPLIB 2017.
π― What it does: Propose the Context Forcing framework, which achieves long-term consistent autoregressive video generation by guiding the student model with a long-context teacher.
Contextual Rollout Bandits for Reinforcement Learning with Verifiable Rewards
Xiaodong Lu (Beihang University), deqing wang
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmark
π― What it does: To address the noise and myopia issues in trajectory sampling within Reinforcement Learning with Verifiable Rewards (RLVR), this paper proposes an adaptive scheduling framework based on contextual bandits (CBS), which dynamically selects high-value trajectories at each training step and efficiently reuses historical trajectories.
π― What it does: This paper studies an auditable buffer-aggregation process for continuous differential privacy release in federated and streaming learning systems, addressing participant changes with single insertions/deletions (single-edit adjacency).
π― What it does: Proposes a continuous semantic segmentation framework called JASCL that simultaneously adapts to changes in classes, domains, and supervision, achieving continuous learning under conditions of limited labeled and abundant unlabeled data.
π― What it does: Proposed a general framework called ConOrd that combines contrastive learning with ordinal learning, performing ordinal contrastive learning on all sample pairs using soft weights;
π― What it does: This paper studies the frequency domain vulnerability of CLIP on adversarial samples, and proposes a test-time defense method called CSR based on this;
π― What it does: Proposed a personalized federated learning framework called pFedCCG based on control collaborative geometry, which avoids consistency collapse and self-clustering through static similarity templates, objective alignment projection, and collaborative intensity scheduling.
π― What it does: Propose an energy-based Transformer framework named Controlled Dynamics Attractor Transformer (CDAT), which combines mixed von Mises-Fisher (Mo-vMF) attention energy, Hopfield refinement energy, and CANN-inspired excitation-inhibition modulation to achieve controllable attractor dynamics; achieves state-of-the-art performance on graph classification and graph anomaly detection tasks.
Conversation for Non-verifiable Learning: Self-Evolving Large Language Models through Meta-Evaluation
Yuan Sui (National University of Singapore), Bryan Hooi (National University of Singapore)
CodeMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose a multi-agent self-dialogue framework called CoNL, which enhances both generation and evaluation capabilities through critique and revision among agents, particularly for unverifiable tasks.
Siqi Zeng (University of Illinois Urbana-Champaign), Xue Feng (Meta)
CodeOptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelContrastive LearningTextMultimodality
π― What it does: This paper proposes a convex dataset valuation method based on gradient space, used to select the most valuable auxiliary datasets during the post-training phase of LLMs.
Convex Low-resource Accent-Robust Language Detection in Speech Recognition
Miria Feng (Stanford University), Mert Pilanci (Stanford University)
CodeRecognitionOptimizationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkContrastive LearningAudio
π― What it does: Propose the Convex Language Detection (CLD) framework, achieving robust language detection in low-resource and accent-diverse speech recognition tasks through convex optimization;