arXivSub Start free trial

ICLR 2026 Papers — Page 7

International Conference on Learning Representations · 5356 papers

Buckingham $\pi$-Invariant Test‑Time Projection for Robust PDE Surrogate Modeling

Seokki Lee (UNIST), Changwook Jeong (UNIST)

Convolutional Neural NetworkMeshPhysics Related

🎯 What it does: Propose a training-agnostic and model-agnostic Buckingham π-invariant test-time projection method that aligns OOD inputs to the training distribution, improving OOD generalization of PDE surrogate models.

Buffer Matters: Unleashing the Power of Off-Policy Reinforcement Learning in Large Language Model Reasoning

Xu Wan (Zhejiang University), Mingyang Sun (Peking University)

TransformerLarge Language ModelReinforcement LearningImageTextBenchmark

🎯 What it does: Proposed an offline reinforcement learning return verification (RLVR) framework BAPO, aiming to improve the efficiency and effectiveness of post-training for large language models by dynamically reconstructing training batches, re-evaluating difficult samples, and reusing high-quality experiences.

Building a Foundational Guardrail for General Agentic Systems via Synthetic Data

Yue Huang (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)

Data SynthesisSafty and PrivacyLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark

🎯 What it does: Constructed a security protection framework for the pre-execution (planning) phase of LLM agents: including a controllable synthetic risk trajectory generation engine AuraGen, a cross-tool unified adapter + guardian model Safiron, and a benchmark for the planning phase called Pre-Exec Bench.

Building Massively Multimodal Foundation Models with Interaction-aware Mixture-of-Experts

Xing Han (Johns Hopkins University), Suchi Saria (Johns Hopkins University)

Computational EfficiencyRepresentation LearningTransformerMixture of ExpertsMultimodalityBiomedical Data

🎯 What it does: Propose a framework called MERGE that utilizes time-delayed multimodal information interaction to guide sparse expert mixture networks.

Building spatial world models from sparse transitional episodic memories

Zizhan He (McGill University), Pouya Bashivan (McGill University)

Meta LearningRecurrent Neural NetworkTransformerWorld ModelImageSequentialBenchmark

🎯 What it does: Propose an Episodic Spatial World Model (ESWM) that can quickly construct a spatial world model using sparse, non-continuous single-step transition memories, and achieve exploration and navigation without training based on this model.

Bures-Isotropy Alignment: Manifold Learning of Generalized Category Discovery

Luyao Tang (University of Hong Kong), Cheng Chen (University of Hong Kong)

Representation LearningTransformerContrastive LearningImage

🎯 What it does: Propose Bures-Isotropy Alignment (BIA), achieving isotropic alignment of class labels by minimizing the Bures distance to recover the low-rank collapse of the token manifold and enhance general category discovery (GCD) performance.

Bures-Wasserstein Flow Matching for Graph Generation

Keyue Jiang (University College London), Laura Toni (University College London)

GenerationGraph Neural NetworkFlow-based ModelGraph

🎯 What it does: Proposed a graph generation framework called BWFlow based on the Bures-Wasserstein intuition, utilizing Markov Random Field (MRF) representations of graphs to jointly model nodes and edges, constructing optimal transport paths and corresponding velocity fields to achieve flow-matching based graph generation.

BWCache: Accelerating Video Diffusion Transformers through Block-Wise Caching

Hanshuai Cui (Beijing Normal University), Weijia Jia (Beijing Normal University)

GenerationComputational EfficiencyTransformerDiffusion modelVideo

🎯 What it does: Proposes Block-Wise Caching (BWCache), a training-free acceleration method that caches blocks and dynamically reuses intermediate features in Diffusion Transformer (DiT).

ByteFlow: Language Modeling through Adaptive Byte Compression without a Tokenizer

Chunyuan Deng (Rice University), Xian Li (Amazon Science)

TransformerLarge Language ModelText

🎯 What it does: Proposed a hierarchical byte-level language model called ByteFlow Net, which does not use a tokenizer. The model self-learns segmentation during training and dynamically determines token boundaries through the encoding rate compression principle.

Byzantine-Robust Federated Learning with Learnable Aggregation Weights

Javad Parsa (Uppsala University), Mikael Johansson (KTH)

OptimizationFederated LearningConvolutional Neural NetworkImage

🎯 What it does: Proposed FedLAW, a Byzantine-robust federated learning framework that treats aggregation weights as learnable parameters.

C-Evolve: Consensus-based Evolution for Prompt Groups

Tiancheng Li (Zhejiang University), Guo-Jun Qi (Westlake University)

OptimizationLarge Language ModelPrompt EngineeringText

🎯 What it does: Propose the Consensus-Evolve (C-Evolve) evolutionary algorithm, which optimizes system prompts for large language models by aggregating voting results from multiple prompts.

C-Voting: Confidence-Based Test-Time Voting without Explicit Energy Functions

Kenji Kubo (University of Tokyo), Yutaka Matsuo (University of Tokyo)

OptimizationRecurrent Neural NetworkBenchmark

🎯 What it does: Proposed a test-time voting method (C-voting) that does not require an explicit energy function, selecting the most confident prediction from multiple randomly initialized potential trajectories and applying it to recursive models;

Cache What Lasts: Token Retention for Memory-Bounded KV Cache in LLMs

Ngoc Bui (Yale University), Rex Ying (JPMorgan Chase AI Research)

OptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText

🎯 What it does: Proposed a KV cache eviction strategy called TRIM-KV based on retention gates, which dynamically retains the most important tokens under a fixed memory budget.

Cache-to-Cache: Direct Semantic Communication Between Large Language Models

Tianyu Fu (Tsinghua University), Yu Wang (Chinese University of Hong Kong)

Computational EfficiencyTransformerLarge Language ModelTextBenchmark

🎯 What it does: Propose Cache-to-Cache (C2C) and demonstrate its capability to directly share KV-Cache for semantic communication in multi-LLM systems, improving accuracy and speed.

Cactus: Accelerating Auto-Regressive Decoding with Constrained Acceptance Speculative Sampling

Yongchang Hao (University of Alberta), Lili Mou (University of Alberta)

OptimizationComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: Propose a constraint-based optimization method for accelerating autoregressive decoding called Cactus, which improves upon traditional Speculative Sampling.

CAD-Tokenizer: Towards Text-Based CAD Prototyping via Modality-Specific Tokenization

Ruiyu Wang (University of Toronto), Jiang Bian (Microsoft Research Asia)

GenerationTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderTextMesh

🎯 What it does: Propose a CAD-Tokenizer framework to achieve text-driven unified CAD prototype generation and editing.

cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning

Maksim Kolodiazhnyi (Lomonosov Moscow State University), Danila Rukhovich (Institute of Mechanics Armenia)

GenerationAI Code AssistantLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityPoint Cloud

🎯 What it does: Developed a multi-modal CAD reconstruction model capable of receiving point cloud, image, and text inputs simultaneously, and outputting executable Python CAD code

CAGE: A Framework for Culturally Adaptive Red-Teaming Benchmark Generation

Chaeyun Kim (DATUMO INC), Minwoo Kim (DATUMO INC)

Adversarial AttackLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: Developed a cross-cultural red team benchmark generation framework named CAGE, instantiated as KORSET in Korean.

Calibrated Information Bottleneck for Trusted Multi-modal Clustering

Shizhe Hu (Zhengzhou University), Mingliang Xu (Zhengzhou University)

Representation LearningContrastive LearningMultimodality

🎯 What it does: Propose a multi-modal clustering framework based on calibrated information bottleneck (CLIB), improving the accuracy and reliability of clustering results through a multi-head structure and dynamic pseudo-label screening.

Calibrating Verbalized Confidence with Self-Generated Distractors

Victor Wang (University of Texas at Austin), Elias Stengel-Eskin (University of Texas at Austin)

Explainability and InterpretabilityTransformerLarge Language ModelTextBiomedical Data

🎯 What it does: Propose the DINCO method, which normalizes the verbalized confidence of LLMs through self-generated distractors and NLI weights, combined with self-consistency to calibrate the model's confidence estimates.

CALM: Co-evolution of Algorithms and Language Model for Automatic Heuristic Design

Ziyao Huang (City University of Hong Kong), Jianping Wang (City University of Hong Kong)

OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTabularBenchmark

🎯 What it does: Propose the CALM framework, achieving co-evolution between language models and heuristic search, automatically generating and iteratively optimizing heuristic algorithms for various combinatorial optimization problems.

Cambrian-S: Towards Spatial Supersensing in Video

Shusheng Yang (New York University), Saining Xie (New York University)

TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelWorld ModelVideoMultimodalityBenchmarkChain-of-Thought

🎯 What it does: Propose the concept of spatial supersensing and design a new evaluation benchmark, VSI-SUPER (comprising two tasks: VSR and VSC), construct a large-scale spatial perception training set, VSI-590K, train and evaluate the CambrianS series of models, and introduce latent-frame prediction and surprise-driven memory and segmentation strategies on VSI-SUPER.

Can Language Models Discover Scaling Laws?

Haowei Lin (Peking University), James Zou

Large Language ModelAgentic AIBenchmark

🎯 What it does: Utilize the evolutionary LLM agent SLDAgent to automatically discover and fit scaling laws for various AI models, aiming to predict model performance.

Can Large Language Models Match the Conclusions of Systematic Reviews?

Christopher Polzak (Stanford University), Serena Yeung-Levy (Stanford University)

TransformerLarge Language ModelPrompt EngineeringBiomedical DataBenchmarkChain-of-Thought

🎯 What it does: Construct the MedEvidence benchmark and evaluate the performance of 25 large language models in replicating conclusions from medical systematic reviews.

Can LLMs Move Beyond Short Exchanges to Realistic Therapy Conversations?

Zhengqing Yuan (University of Notre Dame), Yanfang Ye (University of Notre Dame)

Large Language ModelTextBenchmark

🎯 What it does: Propose CareBench-CBT—a clinical validation benchmark containing multi-turn CBT dialogues, question-answering, and case classification—and develop the Hierarchical Therapy Memory (HTM) structured memory framework to address the long conversation context bottleneck.

Can LLMs Reason Soundly in Law? Auditing Inference Patterns for Legal Judgment

Lu Chen (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)

Explainability and InterpretabilityTransformerLarge Language ModelText

🎯 What it does: This paper proposes a reasoning pattern auditing method based on the AND-OR interaction model to evaluate the internal reasoning logic of large language models (LLMs) in legal judgment tasks. By extracting interaction effects from the scoring functions generated by LLMs, it separates reliable and unreliable reasoning patterns and introduces a new metric called the reliable reasoning proportion (s_reliable).

Can LLMs Refuse Questions They Do Not Know? Measuring Knowledge-Aware Refusal in Factual Tasks

Wenbo Pan (City University of Hong Kong), Xiaohua Jia (City University of Hong Kong)

Explainability and InterpretabilityLarge Language ModelText

🎯 What it does: Propose and evaluate a new metric called Refusal Index (RI) to measure the ability of large language models (LLMs) to refuse answers based on knowledge in fact-based question-answering tasks.

Can SAEs reveal and mitigate racial biases of LLMs in healthcare?

Hiba Ahsan (Northeastern University), Byron C Wallace

Explainability and InterpretabilityTransformerLarge Language ModelAuto EncoderBiomedical DataElectronic Health RecordsChain-of-Thought

🎯 What it does: This paper uses a sparse autoencoder (SAE) to interpret the internal activations of the Gemma-2 LLM, revealing that the model associates Black identity with negative concepts such as crime and drugs in clinical text. The causal impact of this association on prediction outcomes is demonstrated through steering of the corresponding SAE latent variables, while the effectiveness of mitigating bias by zeroing out these latent variables is evaluated.

Can Small Training Runs Reliably Guide Data Curation? Rethinking Proxy-Model Practice

Jiachen T. Wang (Princeton University), Prateek Mittal (Princeton University)

Hyperparameter SearchData-Centric LearningTransformerText

🎯 What it does: Studied the reliability of small proxy models in data curation decisions, and proposed using extremely low learning rates to train proxy models to improve transferability to large-scale models.

Can Speech LLMs Think while Listening?

Yi-Jen Shih (University of Texas at Austin), Mike Seltzer

Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-ThoughtAudio

🎯 What it does: Perform chain-of-thought (CoT) fine-tuning on the multi-stream speech large language model (Moshi), and design a 'think-while-listening' mechanism that allows the model to start generating reasoning trajectories before the user's question is fully spoken, thereby reducing response latency.

Can Transformers Really Do It All? On the Compatibility of Inductive Biases Across Tasks

Damien Teney (Idiap Research Institute), Simon Lucey (Adelaide University)

OptimizationNeural Architecture SearchTransformerTextSequential

🎯 What it does: This study proposes a method that replaces the nonlinear activations (GeLU, softmax) in Transformers with learnable piecewise linear functions, subsequently optimizes the architecture in two stages, and uses this method to evaluate the transferability between different tasks.

Can Vision-Language Models Answer Face to Face Questions in the Real-World?

Reza Pourreza (Qualcomm AI Research), Roland Memisevic (Qualcomm AI Research)

TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmarkAudio

🎯 What it does: Proposed the Qualcomm Interactive Video Dataset (QIVD) and conducted baseline experiments for real-time, context-aware audio-visual question answering tasks, verifying the shortcomings of existing large models in real-time interaction;

Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective.

Ruichuan An, Jiang Bian (Microsoft Research Asia)

ClassificationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark

🎯 What it does: Proposes AesEval-Bench, a comprehensive benchmark for graphic design aesthetic evaluation, encompassing four dimensions, twelve indicators, and three quantifiable tasks (aesthetic judgment, region selection, and precise localization). Subsequently, the authors conduct systematic evaluations of various proprietary, open-source, and reasoning-enhanced visual language models (VLMs), finding that they achieve approximately 72–73% accuracy in aesthetic judgment but perform significantly worse in region selection and precise localization, particularly with IoU below 0.2 for localization. To improve performance, the authors construct the AesEval-Train training set, generating task labels and reasoning paths through two methods: 'human-in-the-loop VLM annotation' and 'metric-associated reasoning.' They then perform full-parameter fine-tuning on Qwen2.5-VL-7B, significantly enhancing performance across all three tasks.

Can we generate portable representations for clinical time series data using LLMs?

Zongliang Ji (University of Toronto), Rahul G Krishnan

Domain AdaptationSafty and PrivacyRepresentation LearningTransformerLarge Language ModelTime SeriesBiomedical DataElectronic Health Records

🎯 What it does: This study investigates whether large language models (LLMs) can generate portable clinical time series data representations for effective deployment of prediction models across different hospitals.

Can You Hear Me Now? A Benchmark for Long-Range Graph Propagation

Luca Miglior (University of Pisa), Davide Bacciu (University of Pisa)

Hyperparameter SearchDrug DiscoveryProtein Structure PredictionGraph Neural NetworkGraphBiomedical DataBenchmark

🎯 What it does: This paper proposes the ECHO benchmark, which includes three synthetic tasks (single-source shortest path, node extreme distance, graph diameter) and two chemical tasks (atomic partial charge prediction, molecular total energy prediction), to rigorously evaluate the long-range information propagation capability of graph neural networks.

Cancer-Myth: Evaluating Large Language Models on Patient Questions with False Presuppositions

Wang Bill Zhu (University of Southern California), Robin Jia (University of Southern California)

TransformerLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmark

🎯 What it does: This paper proposes and constructs the Cancer-Myth dataset to evaluate the response capabilities of large language models (LLMs) on cancer patient questions containing incorrect premises;

Cannistraci-Hebb Training on Ultra-Sparse Spiking Neural Networks

Yuan Hua (Tsinghua University), Hong Chen (Tsinghua University)

ClassificationComputational EfficiencySpiking Neural NetworkImageVideo

🎯 What it does: Proposed a four-stage dynamic sparse training framework called CH-SNN for constructing ultra-sparse spiking neural networks.

Canonical Tree Cover Neural Networks for Expressive and Invariant Graph Learning

Michael Ito (University of Michigan), Jenna Wiens (University of Michigan)

Representation LearningDrug DiscoveryProtein Structure PredictionRecurrent Neural NetworkGraph Neural NetworkGraphBiomedical DataBenchmark

🎯 What it does: Proposed Canonical Tree Cover Neural Networks (CTNNs), which utilize multiple minimum spanning trees to cover the graph in order to obtain expressive and reversible representations that preserve graph distances;

Capability-Based Scaling Trends for LLM-Based Red-Teaming

Alexander Panfilov (ELLIS Institute), Jonas Geiping (ELLIS Institute)

Safty and PrivacyAdversarial AttackLarge Language ModelTextBenchmark

🎯 What it does: This paper investigates the scaling trends of LLM-based red team attacks with the capability gap through large-scale experiments (600+ attacker-target combinations);

Capacity-Aware Inference: Mitigating the Straggler Effect in Mixture of Experts

Shwai He (University of Maryland), Ang Li (University of Maryland)

OptimizationComputational EfficiencyMixture of ExpertsTextMultimodality

🎯 What it does: By introducing capacity-aware Token Drop and Expanded Drop during the MoE inference phase, dynamically limiting expert load and utilizing the idle capacity of underloaded experts, thus alleviating the Straggler effect, improving inference efficiency while maintaining model performance.

CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning

Long Xing (University of Science and Technology of China), Dahua Lin (Shanghai AI Laboratory)

GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality

🎯 What it does: By applying Reinforcement Learning with Verifiable Rewards (RLVR) to open-ended image captioning tasks, the CapRL framework was constructed, adopting a decoupled two-stage VQA mechanism. The framework uses the accuracy of language models on caption-based multiple-choice QA as a verifiable reward to guide the caption generation model to produce richer and more accurate descriptions.

CAPSUL: A Comprehensive Human Protein Benchmark for Subcellular Localization

Yicheng Hu, Fuli Feng (University Of Science And Technology Of China)

ClassificationGraph Neural NetworkTransformerContrastive LearningBiomedical DataBenchmark

🎯 What it does: This paper constructs the CAPSUL benchmark by integrating the AlphaFold2 3D structures of human proteins, FoldSeek 3D tokens, and fine-grained 20-cell compartment localization labels (including experimentally validated levels), and evaluates multiple sequence-and-structure dual-modal models on this dataset.

Captain Cinema: Towards Short Movie Generation

Junfei Xiao (Johns Hopkins University), Lu Jiang (ByteDance Seed)

GenerationData SynthesisDiffusion modelVideoText

🎯 What it does: Achieved end-to-end generation from text to short films by first generating keyframe storyboards and then synthesizing complete films between keyframes.

Capturing Visual Environment Structure Correlates with Control Performance

Jiahua Dong (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)

Representation LearningRobotic IntelligenceConvolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelContrastive LearningImageMultimodality

🎯 What it does: This paper proposes a visual representation prediction proxy task based on a unified low-dimensional state encoding, and evaluates its predictive ability on downstream control strategy performance in various simulated and real-world environments.

CAR-LoRA: Training Compression-Aware and Robust LoRA Adapters for Evolving LLMs

Rana Shahroz (University of North Carolina at Chapel Hill), Charles Fleming (Cisco)

OptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText

🎯 What it does: Proposes the CAR-LoRA framework, which enables the LoRA adapter to be compatible with various quantization, pruning, and layer skipping techniques through random compression simulation during a single training process, maintaining performance as the model evolves.

CARD: Towards Conditional Design of Multi-agent Topological Structures

Tongtong Wu (Monash University), Gholamreza Haffari (Monash University)

Graph Neural NetworkLarge Language ModelReinforcement LearningAgentic AIText

🎯 What it does: Designed a multi-agent LLM framework called CARD that dynamically generates communication topologies based on environmental conditions and supports adaptive topology adjustment during both training and inference phases.

CardioComposer: Leveraging Differentiable Geometry for Compositional Control of Anatomical Diffusion Models

Karim Kadry (Massachusetts Institute of Technology), Elazer R Edelman

GenerationDiffusion modelAuto EncoderBiomedical Data

🎯 What it does: Utilizing differentiable geometric constraints to generate controllable three-dimensional multi-class anatomical label maps in unconditional diffusion models according to shape, size, and position

CaRe-BN: Precise Moving Statistics for Stabilizing Spiking Neural Networks in Reinforcement Learning

Zijie Xu (Peking University), Zhaofei Yu (Peking University)

Spiking Neural NetworkReinforcement LearningImage

🎯 What it does: Proposed a batch normalization method called CaRe-BN specifically designed for spiking neural networks (SNN) in reinforcement learning scenarios, addressing the gradient instability caused by inaccurate estimation under non-stationary distributions in traditional BN.

CARE: Covariance-Aware and Rank-Enhanced Decomposition for Enabling Multi-Head Latent Attention

Zhongzhu Zhou (University of Sydney), Shuaiwen Leon Song (University of Sydney)

Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: Proposed a posterior transfer method called CARE, which combines low-rank decomposition based on activation covariance with hierarchical adaptive rank allocation, converting traditional multi-head attention (MHA/GQA) into multi-head latent attention (MLA) while maintaining the KV cache size and improving model accuracy.

CARE: Towards Clinical Accountability in Multi-Modal Medical Reasoning with an Evidence-Grounded Agentic Framework

Yuexi Du (Microsoft Research Asia), Yan Lu (Microsoft Research Asia)

Explainability and InterpretabilityReinforcement LearningAgentic AIVision Language ModelMultimodalityBiomedical DataBenchmark

🎯 What it does: Proposed the CARE framework, decomposing medical visual question answering into three steps: entity proposal, reference segmentation, and evidence-driven reasoning, while introducing a dynamic coordinator.

CaReBench: A Fine-grained Benchmark for Video Captioning and Retrieval

Yifan Xu (Nanjing Univerisity), Limin Wang (Shanghai AI Laboratory)

GenerationRetrievalSupervised Fine-TuningVision Language ModelContrastive LearningVideoTextMultimodalityBenchmark

🎯 What it does: Propose CAREBENCH, a fine-grained video captioning and retrieval benchmark, and build a unified CARE model that achieves video captioning and retrieval through two-stage supervised fine-tuning;

CARL: Camera-Agnostic Representation Learning for Spectral Image Analysis

Alexander Baumann (Siemens AG), Slobodan Ilic (Siemens AG)

SegmentationAutonomous DrivingConvolutional Neural NetworkTransformerContrastive LearningImageMultimodalityBiomedical Data

🎯 What it does: Proposes CARL, a camera-agnostic spectral image representation learning framework that learns general features across RGB, MS, and HSI cameras, and achieves cross-modal knowledge transfer through self-supervised pre-training.

CARL: Preserving Causal Structure in Representation Learning

Yulong Li (Mohamed bin Zayed University of Artificial Intelligence), Imran Razzak (Mohamed bin Zayed University of Artificial Intelligence)

Representation LearningContrastive LearningMultimodalityBiomedical Data

🎯 What it does: Proposes a cross-modal causal structure preservation framework, CARL, for learning representations in multi-modal data that can both align and maintain causal graph structures.

CARPRT: Class-Aware Zero-Shot Prompt Reweighting for Vision-Language Model

Ruijiang Dong (University Of Melbourne), Masashi Sugiyama (Riken Center For Advanced Intelligence Project)

ClassificationPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark

🎯 What it does: In zero-shot classification for black-box vision-language models (VLMs), a training-agnostic class-aware prompt reweighting method (CARPRT) is proposed, which automatically infers prompt weights for each class by leveraging only the similarity scores between unlabeled images and pre-trained VLMs, thereby enhancing zero-shot inference performance.

Carré du champ flow matching: better quality-generalisation tradeoff in generative models

Jacob Bamberger (Institute of Artificial Intelligence Medical University of Vienna), Adam Gosztolai (Institute of Artificial Intelligence Medical University of Vienna)

GenerationDiffusion modelFlow-based ModelImagePoint CloudTabularTime SeriesStochastic Differential Equation

🎯 What it does: Propose a novel method called Carr' e du champ flow matching (CDC-FM) within the Flow Matching framework, introducing geometric regularization through spatially varying anisotropic noise to balance generation quality and generalization, reducing memorization of training samples.

Cartridges: Lightweight and general-purpose long context representations via self-study

Sabri Eyuboglu (Stanford University), Christopher Re (Stanford University)

Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: Propose CARTRIDGE, a compact key-value cache obtained through offline training, to replace the original context window in large language models; and design the SELF-STUDY method, which utilizes model self-dialogue to generate synthetic data and trains a cache with general reasoning capabilities through context distillation;

Cascadia: An Efficient Cascade Serving System for Large Language Models

Youhe Jiang (Hong Kong University of Science and Technology), Binhang Yuan (University of Cambridge)

OptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark

🎯 What it does: Proposes CASCADIA—a cascading inference service framework for large language models that jointly optimizes model deployment, parallelism strategies, and request routing to achieve a balance between low latency, high throughput, and target answer quality;

CASteer: Cross-Attention Steering for Controllable Concept Erasure

Tatiana Gaintseva (Queen Mary University of London), Ismail Elezi (Huawei Noah's Ark)

GenerationPrompt EngineeringDiffusion modelImage

🎯 What it does: Propose a training-free method called CASteer, which dynamically suppresses and eliminates unwanted concepts (including abstract and concrete concepts) in the cross-attention layers of diffusion models by leveraging precomputed steering vectors, achieving controllable concept elimination;

Cat-PO: Cross-modal Adaptive Token-rewards for Preference Optimization in Truthful Multimodal LLMs

Zhixiao Zheng (University Of Science And Technology Of China), Zhendong Mao (University Of Science And Technology Of China)

OptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningMultimodalityBenchmark

🎯 What it does: This paper proposes calculating visual-related rewards for each generated token in a multimodal large language model and embedding these rewards into the direct preference optimization (DPO) loss to finely reduce hallucinations and enhance model truthfulness.

Catalog-Native LLM: Speaking Item-ID dialect with Less Entanglement for Recommendation

Reza Shirkavand (University of Maryland), Michelle Gong (Roblox)

Recommendation SystemTransformerLarge Language ModelMixture of ExpertsTextSequential

🎯 What it does: Propose a Mixture-of-Experts LLM model (IDIOMoE) that treats product IDs as local dialects in a language space, achieving recommendation by splitting FFN layers into text experts and product experts with static Token-Type routing at each layer.

CatalystBench: A Comprehensive Multi-Task Benchmark for Advancing Language Models in Catalysis Science

Xueqing Chen (Chinese Academy of Sciences), Cheng-Lin Liu (University of Chinese Academy of Sciences)

TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark

🎯 What it does: Proposed the CatalystBench multi-task benchmark and trained a specialized CatalystLLM model for catalytic science based on it

Catching the Details: Self-Distilled RoI Predictors for Fine-Grained MLLM Perception

Yuheng Shi (University of Sydney), Chang Xu (University of Sydney)

SegmentationKnowledge DistillationTransformerVision Language ModelMultimodality

🎯 What it does: Proposed and implemented SD-RPN, a self-distillation based RoI prediction framework, achieving high-precision, unlabeled fine-grained visual perception by generating pseudo labels from internal attention of MLLM.

CaTs and DAGs: Integrating Directed Acyclic Graphs with Transformers for Causally Constrained Predictions

Matthew James Vowels (Kivira Health), Sina Akbari (University Of Cambridge)

Explainability and InterpretabilityTransformerTabularElectronic Health RecordsBenchmark

🎯 What it does: Propose two structures: Causal Transformer (CaT) and Causal Fully Connected Network (CFCN). In Transformer, DAG masks and learnable query vectors γ are used to enforce the model to follow causal structures, supporting causal inference with multi-dimensional embeddings.

CaTS: Calibrated Test-Time Scaling for Efficient LLM Reasoning

Chengsong Huang (Washington Univeristy in St. Louis), Jiaxin Huang (Washington Univeristy in St. Louis)

Computational EfficiencyTransformerLarge Language ModelTextChain-of-Thought

🎯 What it does: This study proposes a self-calibrated dynamic sampling framework called CaTS, which adaptively controls multiple sampling methods such as Best-of-N and Self-Consistency during inference using reliable confidence scores obtained from training, thereby improving the inference accuracy of LLMs under the same sampling budget.

CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data

Shifeng Xie (Universit e Paris Cit e), Ievgen Redko (Huawei Noah Ark Lab)

ClassificationData SynthesisTime Series

🎯 What it does: Proposed the CAUKER framework for generating high-quality synthetic data for time series classification;

Causal Discovery in the Wild: A Voting-Theoretic Ensemble Approach

Vy Vo (Monash University), Mingming Gong (University of Melbourne)

OptimizationTabular

🎯 What it does: Propose a structure integration framework based on voting theory to aggregate the output structures of multiple causal discovery algorithms, thereby enhancing the stability and uncertainty estimation of the causal graph.

Causal Discovery via Quantile Partial Effect

Yikang Chen (East China Normal University), Dehui du

GraphTabular

🎯 What it does: Investigated the identifiability of Quantile Percentile Effect (QPE) in causal discovery, and proposed bivariate and multivariate causal inference algorithms based on QPE.

Causal Imitation Learning under Expert-Observable and Expert-Unobservable Confounding

Daqian Shao (University of Oxford), Marta Kwiatkowska (University of Oxford)

Reinforcement Learning

🎯 What it does: Proposes a generic causal imitation learning framework that can simultaneously handle both observable and unobservable hidden confounding variables by experts;

Causal Interpretation of Neural Network Computations with Contribution Decomposition

Joshua Brendan Melander (Stanford University), Stephen Baccus (Stanford University)

Explainability and InterpretabilityConvolutional Neural NetworkTransformerAuto EncoderImageBiomedical Data

🎯 What it does: Propose the CODEC (Contribution Decomposition) framework, which uses sparse autoencoders to decompose the causal contributions of neural network hidden layer neurons to outputs into interpretable sparse patterns, thereby enabling analysis and control of network behavior.

Causal Score Conditioning for Multi-Resolution Latent Systems

Xuechun Li (Johns Hopkins University), Susu Xu (Johns Hopkins University)

Explainability and InterpretabilityDiffusion modelScore-based ModelGraphTime SeriesStochastic Differential Equation

🎯 What it does: Proposed a Score-based Variational Graphical Diffusion Model (SVGDM) to achieve joint inference of causal systems under multi-resolution and heterogeneous observations.

Causal Structure Learning in Hawkes Processes with Complex Latent Confounder Networks

Songyao Jin (University of California San Diego), Biwei Huang (University of California San Diego)

Explainability and InterpretabilityTime Series

🎯 What it does: This paper studies causal structure learning in partially observable multivariate Hawkes processes, focusing on identifying potential confounding sub-processes and recovering the full causal graph without prior knowledge of the number and location of latent sub-processes.

Causal-Steer: Disentangled Continuous Style Control without Parallel Corpora

Qingsong Wang (Zhejiang University), Jingyuan Chen (Zhejiang University)

Representation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText

🎯 What it does: Proposes the Causal-Steer framework, which utilizes LoRA as a causal intervention tool to directly learn and extract pure style vectors from non-parallel data, achieving continuous, bidirectional, and multi-attribute style control of LLMs through modulation of activation layers.

Causality ≠ Invariance: Function and Concept Vectors in LLMs

Gustaw Opielka, Claire E Stevenson

Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText

🎯 What it does: Studied two types of vectors in large language models: functional vectors (FV) and concept vectors (CV), and compared their performance across different input formats, languages, and question-answer types.

Causally Robust Reward Learning from Reason-Augmented Preference Feedback

Minjune Hwang (University of Southern California), Erdem Biyik

Domain AdaptationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: Propose a preference reward learning framework ReCouPLe that leverages natural language reasoning to eliminate causal confusion caused by low-confidence binary preference feedback and achieve zero-shot transfer across tasks.

Cautious Optimizers: Improving Training with One Line of Code

Kaizhao Liang (University of Texas at Austin), qiang liu

OptimizationImageText

🎯 What it does: Propose a 'cautious optimizer' mechanism implementable with a single line of code in any momentum-based optimizer, utilizing a mask to ensure the update direction aligns in sign with the gradient, thereby accelerating training and improving stability.

Cautious Weight Decay

Lizhang Chen (University of Texas at Austin), qiang liu

ClassificationOptimizationConvolutional Neural NetworkTransformerImageText

🎯 What it does: Proposed a CWD method that can be directly applied to any gradient-based optimizer, applying weight decay only when parameter update directions are consistent, while maintaining the original objective unchanged;

CDBridge: A Cross-omics Post-training Bridge Strategy for Context-aware Biological Modeling

Chang Yu, Stan Z. Li (Westlake University)

TransformerBiomedical DataBenchmark

🎯 What it does: Constructed a two-stage post-training bridging framework, CDBridge, which achieves context-aware mapping from DNA to protein expression using pre-trained DNA and protein models without requiring full retraining.

CDE: Curiosity-Driven Exploration for Efficient Reinforcement Learning in Large Language Models

Runpeng Dai (Tencent), Dong Yu (University of North Carolina at Chapel Hill)

TransformerLarge Language ModelReinforcement LearningTextBenchmark

🎯 What it does: This paper proposes the Curiosity-Driven Exploration (CDE) framework, which provides exploration rewards for RLVR training of large language models by leveraging the actor's perplexity and the variance of a multi-head critic, addressing issues of premature convergence and entropy collapse.

CE-Nav: Flow-Guided Reinforcement Refinement for Cross-Embodiment Local Navigation

Kai Yang (Alibaba Group), Mu Xu (Alibaba Group)

Robotic IntelligenceReinforcement LearningFlow-based ModelTabular

🎯 What it does: Proposed CE-Nav, a two-stage (offline imitation learning + online reinforcement learning) cross-morphology local navigation framework; first train a morphology-agnostic high-level velocity planner (VelFlow) using multimodal data generated by a classical planner, then freeze this expert and use online RL to train a lightweight dynamic adapter, enabling new robots to quickly adapt to their dynamics and controllers;

CellAgent: LLM-Driven Multi-Agent Framework for Natural Language-Based Single-Cell Analysis

Yihang Xiao (Northwestern Polytechnical University), Jiajie Peng (Northwestern Polytechnical University)

TransformerLarge Language ModelAgentic AIBiomedical Data

🎯 What it does: Developed CellAgent, a multi-agent framework based on large language models (LLMs), enabling natural language interaction to achieve end-to-end automated analysis of single-cell RNA-seq and spatial transcriptomics.

CellDuality: Unlocking Biological Reasoning in LLMs with Self-Supervised RLVR

Yuhang Chen (University of North Carolina at Chapel Hill), Tianlong Chen (University of North Carolina at Chapel Hill)

Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningBiomedical DataBenchmark

🎯 What it does: Proposed a self-supervised reinforcement learning-based single-cell biological inference framework called CellDuality.

Celo2: Towards Learned Optimization Free Lunch

Abhinav Moudgil (Mila), Eugene Belilovsky (Mila)

OptimizationMeta LearningReinforcement LearningImageText

🎯 What it does: Propose Celo2, a meta-learned optimizer with extremely low computational cost (only 4.5 GPU hours), which can transfer from small-scale 8×8 image classification tasks to large-scale pre-training (GPT-3 1.3B, ViT ImageNet) and reinforcement learning (Atari PPO) domains while maintaining stability and outperforming traditional AdamW and previous VeLO.

CerebraGloss: Instruction-Tuning a Large Vision-Language Model for Fine-Grained Clinical EEG Interpretation

Wei Gu (Shanghai Jiao Tong University), Wei-Long Zheng (Shanghai Jiao Tong University)

ClassificationRecognitionObject DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelBiomedical DataBenchmark

🎯 What it does: Developed CerebraGloss, an instruction-tuned vision-language model specifically designed for fine-grained clinical EEG interpretation;

Certified Evaluation of Model-Level Explanations for Graph Neural Networks

Sayan Saha (Indian Statistical Institute), Sanghamitra Bandyopadhyay (Indian Statistical Institute)

Explainability and InterpretabilityGraph Neural NetworkGraph

🎯 What it does: Proposed a theoretical framework and practical metrics for evaluating layer explanation methods in graph neural network (GNN) models, addressing the limitations of relying solely on classification scores;

CERTIFIED VS. EMPIRICAL ADVERSARIAL ROBUSTNESS VIA HYBRID CONVOLUTIONS WITH ATTENTION STOCHASTICITY

Joy Dhar (Indian Institute of Technology Ropar), Nayyar Zaidi (Deakin University)

Adversarial AttackConvolutional Neural NetworkImageBiomedical Data

🎯 What it does: Propose a hybrid convolution and attention stochasticity (HyCAS) network, combining a 1-Lipschitz deterministic backbone with two internal randomization modules to achieve dual improvements in ℓ2 certificates and ℓ∞ empirical robustness.

Certifying the Full YOLO Pipeline: A Probabilistic Verification Approach

Zongxin Liu (Key Laboratory of System Software Chinese Academy of Sciences Institute of Software Chinese Academy of Sciences), Lijun Zhang (Key Laboratory of System Software Chinese Academy of Sciences Institute of Software Chinese Academy of Sciences)

Object DetectionAdversarial AttackConvolutional Neural NetworkImage

🎯 What it does: Propose a three-step probabilistic verification framework to evaluate the robustness of YOLO detection networks against object disappearance attacks, covering the entire detection pipeline including NMS processing;

CFO: Learning Continuous-Time PDE Dynamics via Flow-Matched Neural Operators

Xianglong Hou (University of Pennsylvania), Paris Perdikaris (University of Pennsylvania)

Flow-based ModelBenchmarkPhysics RelatedOrdinary Differential Equation

🎯 What it does: Developed the Continuous Flow Operator (CFO) framework, which directly learns the time-right-hand side of PDEs using flow matching, enabling neural operator learning for continuous-time PDEs without requiring backpropagation through ODE solvers during training.

CFT-RAG: An Entity Tree Based Retrieval Augmented Generation Algorithm With Cuckoo Filter

Zihang Li (Peking University), Tong Yang (Peking University)

RetrievalTextBiomedical DataRetrieval-Augmented Generation

🎯 What it does: This paper proposes the entity tree-based retrieval-augmented generation algorithm CFT-RAG, which uses an improved Cuckoo Filter to accelerate the retrieval process of Tree-RAG.

CGSA: Class-Guided Slot-Aware Adaptation for Source-Free Object Detection

Boyang Dai (University of Hong Kong), Yizhou Yu (University of Hong Kong)

Object DetectionDomain AdaptationTransformerContrastive LearningImage

🎯 What it does: In the source-agnostic domain adaptation object detection task, the CGSA framework is proposed, achieving object-level decomposition and semantic guidance on target domain images by integrating Hierarchical Slot Awareness and Class-Guided Slot Contrast into the DETR detector, thereby enhancing adaptation performance.

Chain-of-Context Learning: Dynamic Constraint Understanding for Multi-Task VRPs

Shuangchun Gui (Singapore Management University), Zhiguang Cao (Singapore Management University)

OptimizationTransformerReinforcement LearningGraphSequential

🎯 What it does: Propose a multi-task vehicle routing planning framework based on Chained Context Learning (CCL), which achieves progressive node state updates through dynamic constraint reshaping and trajectory-sharing node re-embedding.

ChainGPT: Dual-Reasoning Model with Recurrent Depth and Multi-Rank State Updates

Yunao Zheng (Beijing University of Posts and Telecommunications), Chen Wei (Li Auto Inc)

TransformerLarge Language ModelTextChain-of-Thought

🎯 What it does: Propose the ChainGPT two-layer reasoning model: within layers, multi-step LoRA state updates (RWKV-Product) and state-guided sparse attention (SGSA) are adopted, while between layers, a recursive depth mechanism is used to achieve multi-round implicit reasoning, thereby enhancing the reasoning depth and efficiency of large language models.

ChainMPQ: Interleaved Text-Image Reasoning Chains for Mitigating Relation Hallucinations

Yike Wu (University of Queensland), Yujun Cai (University of Queensland)

Prompt EngineeringVision Language ModelMultimodalityChain-of-Thought

🎯 What it does: Propose a training-agnostic multi-perspective QA interactive text-image reasoning chain (ChainMPQ), which first extracts theme and object keywords to enhance visual attention, constructs multi-angle sub-questions, and gradually propagates text and visual memories within the reasoning chain, significantly reducing relational hallucinations in large vision-language models.

CHAMMI-75: Pre-training multi-channel models with heterogeneous microscopy images

Vidit Agrawal (Morgridge Institute for Research), Juan C. Caicedo (Boston University)

Domain AdaptationRepresentation LearningTransformerContrastive LearningBiomedical DataBenchmark

🎯 What it does: Constructed the CHAMMI-75 multi-channel microscope image dataset and trained a general-purpose cell morphology model, MorphEm, using self-supervised learning.

Change Point Localization and Inference in Dynamic Multilayer Networks

Fan Wang (University of Melbourne), OSCAR HERNAN MADRID PADILLA

GraphTime Series

🎯 What it does: Study the change point localization and inference for offline dynamic multi-layer random dot product graphs (D-MRDPG), proposing a two-stage algorithm: Seed Binary Segmentation + Low-Rank Tensor Estimation, and proving its consistency and limiting distribution.

Channel-Aware Mixed-Precision Quantization for Efficient Long-Context Inference

Chengxi Liao (Hong Kong University of Science and Technology), Zeyi Wen (Hong Kong University of Science and Technology)

Computational EfficiencyTransformerText

🎯 What it does: Proposes ChanMix, a channel-aware mixed-precision quantization framework tailored for KV cache;

Characteristic Root Analysis and Regularization for Linear Time Series Forecasting

Zheng Wang (Bosch (China) Investment Co., Ltd.), Tobias Schlagenhauf (Robert Bosch GmbH)

Time SeriesBenchmark

🎯 What it does: Systematically study linear time series forecasting models, analyze the role of characteristic roots in noisy and noise-free scenarios, and propose two root reconstruction strategies.

Characterization and Learning of Causal Graphs with Latent Confounders and Post-treatment Selection from Interventional Data

Gongxu Luo (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)

Biomedical Data

🎯 What it does: Proposes a causal structure learning method addressing potential confounding variables and post-treatment selection bias, defining fine-grained interactive Markov equivalence classes (FI-Markov equivalence) and corresponding graphical representations (F-PAG). Subsequently, designs a complete and feasible algorithm F-FCI to simultaneously identify causal relationships, latent confounders, and selection bias from observational and interventional data.

Characterizing and Mitigating Reasoning Drift in Large Language Models

Yufeng Zhang (Institute of Automation, Chinese Academy of Sciences), Jinqiao Wang (Institute of Automation, Chinese Academy of Sciences)

Explainability and InterpretabilityTransformerTextBenchmarkChain-of-Thought

🎯 What it does: Analyzes functional transfer during the reasoning process of large language models, identifies and defines the 'reasoning drift' problem, and proposes a lightweight activation guidance method (RAAS) during inference to suppress drift and improve reasoning accuracy.

Characterizing and Optimizing the Spatial Kernel of Multi Resolution Hash Encodings

Tianxiang Dai (Stanford University), Jonathan Fan (Stanford University)

OptimizationComputational EfficiencyNeural Radiance FieldImagePoint CloudMeshBenchmark

🎯 What it does: Analyze the spatial behavior of multi-resolution hash encoding (MHE) from a physical system perspective, and quantify its effective resolution and variance through the point spread function (PSF);

Characterizing Deep Research: A Benchmark and Formal Definition

Abhinav Java (Microsoft Research), Amit Sharma (Microsoft Research)

Large Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation

🎯 What it does: This paper formally defines the Deep Research (DR) task as requiring both high search volume and non-trivial reasoning, and proposes a statement-based evaluation framework along with the public benchmark LiveDRBench (100 scientific and current affairs-related queries) to objectively assess DR system performance.