ICML 2026 Papers — Page 43
International Conference on Machine Learning · 6554 papers
Position: Assistive Agents Need Accessibility Alignment
Jie Hu (Hunan University), Jiaming Zhang (Hunan University)
Recommendation SystemAutonomous DrivingSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyRobotic IntelligenceAgentic AITextTabularReview/Survey Paper
🎯 What it does: Propose 'accessibility alignment' as the core objective of auxiliary agents, and construct a lifecycle pipeline for design, deployment, and iteration, along with a four-dimensional alignment framework.
Position: Assistive AI requires Personalized Specialists, not Generalists
Homanga Bharadhwaj (Johns Hopkins University)
Federated LearningSafty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextSequentialRetrieval-Augmented Generation
🎯 What it does: Propose and argue that in the field of auxiliary artificial intelligence, the truly valuable assistant is not a general-purpose model, but a personalized expert model that can continuously adapt to specific users, environments, and interaction history after deployment.
Position: Behavioral Systems Require Behavioral Tests
Manuel Cherep (Massachusetts Institute of Technology), Patricia Maes
Explainability and InterpretabilityRobotic IntelligenceAI Code AssistantTransformerLarge Language ModelReinforcement LearningTextSequentialBenchmark
🎯 What it does: Propose shifting the evaluation of AI agents from purely performance metrics to the assessment of behavioral systems, suggesting the construction of a behavioral testing framework through systematic observation, intervention, and interpretation of action sequences.
Position: Benchmarks Cannot Establish Deployment Readiness of Clinical AI
Haoran Zhang (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)
TextBiomedical DataElectronic Health RecordsReview/Survey PaperBenchmark
🎯 What it does: This paper discusses the limitations of benchmarks in the field of clinical artificial intelligence when evaluating the readiness of models for deployment, and argues that benchmark scores should not be considered sufficient evidence for clinical deployment.
Position: Benchmarks for Vision–Language Models in Urban Perception Should Be Reliability-Aware and Negotiated
Rashid Mushkani (Université de Montréal)
Autonomous DrivingExplainability and InterpretabilityComputational EfficiencyRepresentation LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextTabularBenchmark
🎯 What it does: This paper constructs a visual-language evaluation benchmark containing 100 Montreal street scenes, and explores the relationship between evaluation reliability and voting consistency through zero-shot testing on seven VLMs.
Position: Beyond Prediction: Toward Verifiable Physiological Waveform Reasoning with Foundation Models and Agentic LLMs
Xiaoda Wang (Emory University), Carl Yang (Emory University)
Explainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringTime SeriesBiomedical DataElectronic Health RecordsElectrocardiogramRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a verifiable physiological waveform reasoning framework based on foundational models and agent-capable large language models.
Position: Beyond Sensitive Attributes, ML Fairness Should Quantify Structural Injustice via Social Determinants
Zeyu Tang (Stanford University), Sanmi Koyejo (Stanford University)
Recommendation SystemOptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityContrastive LearningTabularBiomedical DataElectronic Health Records
🎯 What it does: Through theoretical analysis and empirical research, the necessity of incorporating social determinants into machine learning fairness evaluation is proposed and verified, and structural injustice that may be triggered by traditional approaches based on sensitive attributes is demonstrated.
Position: Breaking the Dual Curse of Multilingual AI Requires Socio-Technical Guardrails, Not Post-Hoc Alignment Alone
Jason S Lucas, Dongwon Lee (Pennsylvania State University)
Safty and PrivacyExplainability and InterpretabilityKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextMultimodalityReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper systematically reviews 207 studies, revealing that large language models in low-resource languages are both prone to generating harmful content (35%) and struggle to follow instructions (-20%). It also demonstrates that reward models achieve accuracy at random levels in low-resource languages, rendering later alignment ineffective; it proposes a socio-technical framework centered on safe context distillation, community participation in defining harm, and joint evaluation (attack success rate/rejection rate).
Position: Bridge Human Interpretation and Machine Representation With Explicit Specification For Qualitative Data Analysis In LLM Era
Xinyu Pi (University of California San Diego), Hua Shen (New York University Shanghai)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Proposes a four-order two-dimensional (4×4) framework that explicitly distinguishes between the two dimensions of 'meaning construction' and 'modeling' in qualitative analysis, and reveals the shortcomings and preferences of existing LLM-assisted qualitative research in these two dimensions through a systematic audit of 300 related papers.
Position: Bridge the AI development-regulation gap through dedicated committees and adaptive legislation
Mansur Ali Khan (University of Washington), Ahmad A Rushdi
Federated LearningSafty and PrivacyExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Quantitatively analyze AI-related legislation in the United States from 2017 to 2025, revealing that the development speed of AI technology far outpaces legislative progress, creating a significant governance gap.
Position: Carbon Footprint Reporting Should Be Routine in Machine Learning Research
Guan-Ming Chiu (National Taiwan University)
Explainability and InterpretabilityComputational EfficiencyReview/Survey Paper
🎯 What it does: Propose and advocate for incorporating carbon footprint reporting into the routine practices of machine learning research, constructing standardized metrics, measurement tools, and a phased adoption process.
Position: Causality Is Key for Interpretability Claims to Generalise
Shruti Joshi (Mila Quebec AI Institute), Dhanya Sridhar (Mila Quebec AI Institute)
Explainability and InterpretabilityLarge Language ModelAuto EncoderContrastive LearningTextReview/Survey Paper
🎯 What it does: Propose combining interpretability research on large language models with causal inference, constructing an interpretability framework centered on causal hierarchy, and clarifying the correspondence between estimands, identifiability, and method evidence.
Position: Certified Correctness in Neural Constraint Reasoning Requires Symbolic Integration
Shufeng Kong (Sun Yat-sen University), Caihua Liu (Foshan University)
OptimizationExplainability and InterpretabilityComputational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerAgentic AIContrastive LearningTextGraphTabularSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes that neural constraint reasoning must prioritize symbolic integration, explains the 'Certification Gap', and designs a Proposer-Verifier-Solver (PVS) multi-agent framework, proving that provably correct solutions can be achieved on Sudoku and other NP problems.
Position: CNNs Don't See Shape — And That Won't Change Without New Architectures
Ali Kayyam (BrainChip Inc.)
ClassificationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningConvolutional Neural NetworkSpiking Neural NetworkTransformerSupervised Fine-TuningAuto EncoderContrastive LearningImage
🎯 What it does: This paper designs minimized, fully controllable image stimuli to compare the cue-conflict and cue-suppression experimental paradigms, directly testing the preference of convolutional neural networks when shape and texture conflict. Based on this, it proposes that data augmentation alone is insufficient to eliminate texture bias, and new network structures are needed.
Position: Code Benchmarks Should Prioritize Rigor, Reliability, and Reproducibility
Jialun Cao (Hong Kong University of Science and Technology), Shing-Chi Cheung (Hong Kong University of Science and Technology)
Review/Survey PaperBenchmark
🎯 What it does: A systematic evaluation of 672 code benchmarks from the past ten years (2014-2025) reveals a mismatch between quality and awareness, and proposes a 55-item HOW2BENCH checklist to standardize the rigor, reliability, and reproducibility of benchmarks.
Position: Collaborative Agentic AI Needs Interoperability Across Ecosystems
Rishi Sharma (EPFL), Anne-Marie Kermarrec (EPFL)
Federated LearningSafty and PrivacyAgentic AI
🎯 What it does: Proposes a minimal interoperability framework called WEB OF AGENTS, aiming to enable collaborative agents across ecosystems;
Position: Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions
Matthew Riemer (Mila, University of Montreal), Guillaume Dumas (Mila, University of Montreal)
TransformerLarge Language ModelPrompt EngineeringTabularFinance RelatedChain-of-Thought
🎯 What it does: Studied the tendency of AI agents with chain-of-thought reasoning capabilities to engage in silent collusion in price decision scenarios, and proposed policy recommendations that behavior authentication should be conducted before allowing them to participate in economic decisions.
Position: Comprehensive AI governance requires addressing non-model capability gains
Arthur Goemans (Google DeepMind), Allan Dafoe (Google DeepMind)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyLarge Language ModelPrompt EngineeringTextReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper analyzes the limitations of model layer governance and proposes the concept of non-model gains along with a governance hierarchy framework.
Position: Creating High-Fidelity Synthetic Training Data Should Employ Multi-level Optimization
Pengtao Xie (University of California San Diego), Ruiyi Zhang (University of California San Diego)
Image TranslationSegmentationGenerationData SynthesisDomain AdaptationOptimizationTransformerLarge Language ModelPrompt EngineeringDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningImageTextBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
🎯 What it does: Propose a multi-layer optimization (MLO) framework that jointly integrates data generation, automatic annotation, domain adaptation, and data selection, aiming to produce high-fidelity synthetic training data with the goal of improving downstream model performance.
Position: Current Benchmarking Hinders Real Progress in Deep Learning for Time Series Forecasting
Valentina Moretti (IDSIA, Università della Svizzera italiana), Andrea Cini (IMOS Lab, EPFL)
Hyperparameter SearchRecurrent Neural NetworkTransformerTime SeriesBenchmark
🎯 What it does: Evaluate the effectiveness of benchmark methods for deep learning time series forecasting models, revealing the impact of key design dimensions on performance
Position: Current Model Cards Are Insufficient for Downstream Governance of Open-Weight Foundation Models
Sungwon Chae (Seoul National University), Sangchul Park (Seoul National University)
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelTextReview/Survey Paper
🎯 What it does: A systematic analysis of model cards, acceptable use policies (AUP), and licenses of the 500 most downloaded open-weight foundation models (OWFM) on Hugging Face, and based on the findings, a three-tier governance framework is proposed (information layer: model cards; normative layer: AUP; legal layer: specialized licenses).
Position: Deciphering the Functions of DNAs, RNAs, and Proteins Should Consider Multi-Modal Large Language Models
Pengtao Xie (University of California San Diego), Bernhard Palsson (University of California San Diego)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose a multi-modal large language model framework that utilizes protein, DNA, RNA sequence and structural information to generate free-text functional descriptions and supports interactive dialogue;
Position: Deployed Reinforcement Learning should be Continual
Parnian Behdin (Alberta Machine Intelligence Institute), Golnaz Mesbahi (Alberta Machine Intelligence Institute)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyMeta LearningReinforcement LearningPrompt EngineeringContrastive LearningWorld ModelTabularTime SeriesSequentialReview/Survey PaperBenchmarkFinance Related
🎯 What it does: The paper argues that continuous learning after deployment in the real world is essential, viewing measurable deployment as a continuous reinforcement learning problem, and elaborating it through a historical process framework and multi-industry case studies.
Position: Digital Agents Require Unified Agent-Native Environments
Yiran Wu (Pennsylvania State University), Qingyun Wu (Pennsylvania State University)
Robotic IntelligenceAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a unified Agent-Native Computer framework required for digital agents, and implemented the AgentVM environment.
Position: Don't Just "Fix it in Post'': A Science of AI Must Study Learning Dynamics
Stella Biderman (EleutherAI), Naomi Saphra (Boston University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: This paper proposes a framework for building AI science, emphasizing the need to understand and improve model behavior from the perspective of training dynamics rather than post-hoc fixes.
Position: Early-Stage Quality Assurance in Annotation Pipelines Is More Cost-Effective Than Late-Stage Validation
Sunil Kothari (Centific AI Research), Tao Liu (Centific AI Research)
Anomaly DetectionData-Centric LearningContrastive LearningVideo
🎯 What it does: Propose viewing the timing of quality assurance (T0 pre-annotation, T1 post-annotation, T2 post-review) as a key design variable in the annotation pipeline, and demonstrate that early QA is more cost-effective through theoretical models, empirical investigations, and pilot experiments.
Position: Embodied AI Requires a Privacy-Utility Tradeoff
Xiaoliang Fan (Xiamen University), Cheng Wang (Xiamen University)
Safty and PrivacyRobotic IntelligenceReinforcement LearningSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Propose the SPINE framework, treating privacy as a dynamic control signal throughout the Embodied AI lifecycle, and experimentally validate it in navigation tasks.
Position: Epistemic Uncertainty Estimation Methods are Fundamentally Incomplete
Sebastian Jimenez, Willem Waegeman (Ghent University)
TabularReview/Survey Paper
🎯 What it does: This paper addresses the method of decomposing uncertainty into aleatoric and epistemic uncertainty in supervised learning, pointing out that existing second-order distribution-based estimates have fundamental defects both theoretically and practically, especially neglecting estimation bias and capturing only partial variance contributions.
Position: EU AI Act's Research Exemptions Can Break the Publication Norms of Major AI Conferences
Alina Wernick (University Tübingen), Kristof Meding (University Tübingen)
Safty and PrivacyExplainability and InterpretabilityReview/Survey Paper
🎯 What it does: This paper conducts an in-depth analysis of the applicability of the EU AI Act (AI Act) in academic research, revealing conflicts between legal texts and AI research practices (especially model and system releases), and proposes actionable response strategies.
Position: Evaluating LLMs in Finance Requires Explicit Bias Consideration
Yaxuan Kong (University of Oxford), Stefan Zohren (University of Oxford)
TransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperFinance RelatedRetrieval-Augmented Generation
🎯 What it does: This paper identifies five common biases in the evaluation of financial domain LLMs (prospective bias, survivorship bias, narrative bias, objective bias, cost bias), and designs a structured validity framework (Structural Validity Framework) and an evaluation checklist to standardize experiments and deployment validation of financial LLMs.
Position: Evaluation of ECG Representations Must Be Fixed
Zachary Evan Berger (Massachusetts Institute of Technology), Collin Stultz (Massachusetts Institute of Technology)
Anomaly DetectionRepresentation LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningTime SeriesBiomedical DataElectrocardiogramReview/Survey PaperBenchmark
🎯 What it does: Reviewed and evaluated benchmark practices for 12-lead electrocardiogram (ECG) representation learning, proposed a broader set of clinical tasks and assessment best practices, and conducted experiments on six downstream tasks using five representative pretraining methods, comparing their performance with a random encoder baseline.
Position: Evaluation of ML Resource Utilization Requires Model Life Cycle Assessment
Jared Fernandez (Carnegie Mellon University), Emma Strubell (Carnegie Mellon University)
Explainability and InterpretabilityComputational EfficiencyReview/Survey Paper
🎯 What it does: Propose applying Life Cycle Assessment (LCA) to the development and deployment of machine learning models, constructing a unified functional unit and evaluation process to quantify resource consumption and environmental impact across the entire lifecycle, including hardware manufacturing, training, and inference.
Position: Every Ground Truth is a Human Construction, not an Objective Truth
Charlotte Högberg (Lund University), Kiri L. Wagstaff (Oregon State University)
ImageTextMultimodalityTabularReview/Survey Paper
🎯 What it does: This paper argues through theoretical reasoning and case analysis that ground truth is a socio-technical construct co-constructed by humans and technology, rather than an objective natural fact, and proposes concepts and practical recommendations such as situational reliability, co-construction, and documentation.
Position: Explainability Research Must Prioritize Foundations over Ad-hoc Methods
Michal Moshkovitz (Google Research), Jennifer Wortman Vaughan (Microsoft Research)
Explainability and InterpretabilityLarge Language ModelTextReview/Survey Paper
🎯 What it does: This paper reviews and systematically analyzes the current state of XAI research, pointing out the contradiction between method development and foundational shortcomings, summarizing four core challenges (definition, attributes, evaluation, operability), and providing a practical checklist.
Position: Explanation Stability Is a Property of the Model–Method Pair, Not the Model
Kabilan Elangovan (Singapore Health Services and Singapore Eye Research Institute), Daniel Shu Wei Ting (Singapore Health Services and Singapore Eye Research Institute)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Evaluated the explanation stability of different deep learning models during transfer learning and fine-tuning, using two gradient attribution methods.
Position: Express Your Doubts — Probabilistic World Modeling Should Not Be Based on Token *logprobs*
Eitan Wagner (Hebrew University of Jerusalem), Omri Abend (Hebrew University of Jerusalem)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringWorld ModelText
🎯 What it does: This paper argues that the output probability of traditional language models based on token logprob cannot reliably estimate the probability of real-world events, and advocates using 'second-order prediction' to explicitly report event probabilities.
Position: Fairness Failure in Generative Models is an Evaluation Problem
Mariia Vladimirova (Criteo AI Lab), Thibaut Issenhuth (Criteo AI Lab)
GenerationData SynthesisFederated LearningSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringDiffusion modelTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Diagnose the current state of fairness evaluation for generative models and propose a novel auditable reporting standard called Fairness Card, aiming to unify and clarify evaluation protocols, thereby improving the comparability and reproducibility of fairness assessments.
Position: Federated Learning is a Lens towards a Democratized Future for the Scaling Law Era
Harry H. Jiang (Carnegie Mellon University), Carlee Joe-Wong (Carnegie Mellon University)
Federated LearningSafty and PrivacyData-Centric LearningReview/Survey Paper
🎯 What it does: This paper discusses the advantages of federated learning (FL) in the areas of data, privacy, and computational resources, proposing that FL can serve as an entry point for achieving a decentralized, collaborative, and accountable machine learning ecosystem.
Position: From Crowdsourcing to Crowd-LLM-Sourcing and LLM-Sourcing
Jiyi Li (Hokkaido University)
Recommendation SystemOptimizationFederated LearningExplainability and InterpretabilityData-Centric LearningLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes two major paradigms, 'Crowd-LLM-Sourcing' and 'LLM-Sourcing', systematically organizing and aligning traditional crowdsourcing mechanisms with large language model (LLM) technologies, and explaining how to design processes such as quality control, task allocation, and result aggregation at their intersection.
Position: GenAI Systems Should Implement Contribution-Aware Revenue Sharing for Data Providers
Gengrui Zhang (Concordia University)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelGenerative Adversarial NetworkTextFinance Related
🎯 What it does: Proposes a framework for distributing revenue from generative AI systems based on contribution, and elaborates on its necessity and implementation ideas.
Position: Generative Distributional Integrity against Backdoor Attacks
Shuaibiao Han (University of Science and Technology of China), Wenjie Ruan (University of Science and Technology of China)
GenerationSafty and PrivacyTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTextMultimodality
🎯 What it does: Proposes a security framework for generative models based on distributional integrity, systematically analyzing the propagation mechanisms of generated backdoors in the model supply chain and synthetic data loops, and proposes a three-layer defense scheme including parameter-level auditing, weight smoothing, and cross-modal geometric verification.
Position: Generative Engine Optimization Creates Underexamined Risks, Governance Must Target Concentration, Disclosure, and Academic Blind Spots
Yizhu Wen (Indiana University Bloomington), Hanqing Guo (Indiana University Bloomington)
OptimizationSafty and PrivacyExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes a general pipeline called Generative Engine Optimization (GEO), systematically analyzes its impact on LLM answer engines, and compares academic and industrial practices and risks.
Position: Generative Models Erode Human Temporal Learning Through Market Selection
Wenjun Cao (Independent Researcher)
Large Language ModelGenerative Adversarial NetworkTextReview/Survey PaperFinance Related
🎯 What it does: The paper explores the structural risks that modern generative models pose to knowledge and cultural production at the current sub-AGI capability level, defining human temporal learning (HTL) as knowledge accumulated through continuous engagement with problems. The outputs of generative models increasingly resemble the surface characteristics of HTL-intensive work, making it costly to verify whether outputs reflect genuine human learning, leading to a collapse in value.
Position: Genomic Model Research Must Move Beyond Anecdotal Evaluation of Interpretability Methods
Shasha Zhou (University of Exeter), Ke Li (University of Exeter)
Explainability and InterpretabilityDrug DiscoveryTransformerBiomedical DataReview/Survey PaperBenchmark
🎯 What it does: Through systematic mapping of 3,575 genome IML papers and rigorous benchmark evaluation of TF binding prediction, this paper points out that existing explanation methods lack consistency, credibility, and biological validity, and proposes a hierarchical evaluation criterion;
Position: Good Embodied Reward Models Need Bad Behavior Data
Thomas Tian (University of California Berkeley), Andrea Bajcsy (Carnegie Mellon University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageVideoTextSequentialRetrieval-Augmented Generation
🎯 What it does: This paper analyzes three mainstream body reward models and demonstrates that they produce excessive rewards when facing failures, low-quality actions, or safety violations. It proposes that incorporating real or synthetic 'bad' robot data during training and evaluation can significantly improve the alignment of reward models with human preferences.
Position: Graph Condensation Needs a Reset—Move Beyond Full-dataset Training and Model-Dependence
Mridul Gupta (Indian Institute of Technology Delhi), Sayan Ranu (Indian Institute of Technology Delhi)
CompressionComputational EfficiencyRepresentation LearningNeural Architecture SearchGraph Neural NetworkAuto EncoderContrastive LearningGraphBenchmark
🎯 What it does: This paper presents a systematic critique of current research on graph-to-graph condensation and redefines the 'graph-to-graph condensation' task, emphasizing that full-data training is no longer necessary, the method should be model-agnostic, and the core evaluation metric should be resource consumption (in bytes). It calls for more realistic and deployable research directions.
Position: Hallucinations Undermine Trust; Metacognition is a Way Forward
Gal Yona (Google Research), Yossi Matias (Google Research)
Recommendation SystemAnomaly DetectionFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Analyzes the 'hallucination' problem in large language models for factual question answering, and proposes redefining hallucination as 'confident errors,' introducing 'faithful uncertainty' and a metacognitive framework to reduce misinformation while maintaining practicality.
Position: Hippocampal Explicit Memory Is the Cornerstone for AGI
Sangjun Park (University of Texas at Austin), Sangjun Park (Cognizant AI Labs)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningMeta LearningReinforcement Learning from Human FeedbackLarge Language ModelTextReview/Survey Paper
🎯 What it does: This paper compares human implicit memory with LLM learning mechanisms, proposing that hippocampal explicit memory should be incorporated in AGI implementation, and theoretically elaborates on the computational requirements of explicit memory systems.
Position: Human-Centric Vision Requires Topological Generalization Beyond Fixed Skeletal Topologies
Heming Du (University of Queensland), Xin Yu (University of Adelaide)
Pose EstimationGraph Neural NetworkDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderContrastive LearningImageGraph
🎯 What it does: This paper argues that in human visual tasks, fixed skeleton topologies should be abandoned, and instead proposes a framework that adapts skeleton topologies to individual instances. This framework can dynamically predict the presence and connectivity of joints based on anatomical differences among individuals, thereby improving the accuracy of pose estimation for individuals with limb deficiencies and reducing the generation of erroneous structures.
Position: ICML Should Treat Hosted LLM APIs as Versioned Dependencies and Require Drift-Audit Artifacts
Utsav Gupta (Stanford University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper proposes that conferences such as ICML require papers relying on managed LLM API to submit a lightweight drift audit artifact (LDAA), ensuring that experimental results remain interpretable and traceable after API updates.
Position: Ideas Should be the Center of Machine Learning Research
Jairo Diaz-Rodriguez (York University)
Explainability and InterpretabilityData-Centric LearningImageReview/Survey PaperChain-of-Thought
🎯 What it does: Proposed and argued for a machine learning research framework centered on 'ideas,' called Ideas First, which advocates first determining ideas, then predicting observable signatures, and subsequently designing targeted experiments to verify them, rather than relying solely on leaderboard scores or idealized theories;
Position: If Open Source Is to Win, It Must Go Public
Joshua Z Tan (Public AI Network), Jenia Jitsev (LAION)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelDiffusion modelContrastive LearningTextMultimodalityReview/Survey Paper
🎯 What it does: This paper discusses the challenges faced by open source AI in terms of resources, licensing, and governance, and proposes the need to build public AI infrastructure to achieve true democratization and sustainability.
Position: Improved Documentation is Necessary for Benchmarking AI Systems in Geometry
Anna Genevaux (Independent Researcher), Simon Frieder (University of Oxford)
Large Language ModelTextBenchmarkPhysics Related
🎯 What it does: Proposed a benchmark release standard for an executable DSL (JGEX) in the geometric domain, constructed and released the JgexDiv corpus containing 137 Euclidean geometry problems, and provided executable interface contracts, predicate support tables, version-fixed verification scripts, and document rewrite records.
Position: In Defense of Information Leakage in Concept-based Models
Mateo Espinosa Zarlenga (University of Oxford)
Explainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper investigates the role of information leakage in concept-based models (CM), arguing that leakage is not entirely harmful but can be considered 'beneficial leakage' in practical scenarios with incomplete concepts, and can be achieved through regularization;
Position: Infringement cannot be cured after training
Satoru Utsunomiya (University of Tokyo), Ichiro Sakata (University of Tokyo)
Review/Survey Paper
🎯 What it does: Analyzes the copyright, contractual, and infringement risks involved in the training phase of generative AI, and points out that post-training 'Opt-Out' techniques applied to model outputs alone cannot eliminate prior liability for infringement.
Position: Interestingness is an Inductive Heuristic for Future Compression Progress
Vincent Herrmann (Swiss AI Lab IDSIA/USI/SUPSI), Jürgen Schmidhuber (Swiss AI Lab IDSIA/USI/SUPSI)
CompressionAuto EncoderText
🎯 What it does: This paper proposes to view future compression progress as a predictable 'interestingness,' formalizing this concept in the form of a complexity-runtime profile, and further proves that past compression progress can predict future progress under different priors.
Position: Interpretability Can Be Actionable
Hadas Orgad (Kempner Institute at Harvard University), Mor Geva
Explainability and InterpretabilityLarge Language ModelImageTextReview/Survey Paper
🎯 What it does: Proposes a framework that integrates interpretability research with operationality, advocating that interpretability should be measured through executable decisions, and provides an operationality evaluation checklist and action dimensions;
Position: Interpretability in Deep Time Series Models Demands Semantic Alignment
Giovanni De Felice (Universita della Svizzera Italiana), Silvia Santini (Universita della Svizzera Italiana)
Explainability and InterpretabilityTime Series
🎯 What it does: This paper proposes that the interpretability of deep temporal models should pursue semantic alignment, and provides a formal definition and implementation blueprint.
Position: Invisible Tokens, Visible Bills: The Urgent Need to Audit Hidden Operations in Opaque LLM Services
Guoheng Sun (University of Maryland), Ang Li (University of Maryland)
Machine LearningSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose an audit framework for hidden operations in commercial opaque LLM services, and elaborate on the risks of quantity inflation and quality degradation
Position: Irresponsible AI: big tech’s influence on AI research and associated impacts
Alex Hernández-García, Mélisande Teng
Review/Survey Paper
🎯 What it does: This paper argues that the influence of large companies on AI research leads to irresponsible AI development through a literature review and case analysis, and calls on researchers to take action.
Position: It is Time to Virtualize Foundation Models with a Self-evolving Operating System Layer
Suparna Bhattacharya (Hewlett Packard Enterprise), Ian Foster (University of Chicago)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackHyperparameter SearchData-Centric LearningReinforcement Learning from Human FeedbackNeural Architecture SearchLarge Language ModelPrompt EngineeringMixture of ExpertsTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Proposes a new system layer called Foundation Model Operating System (FMOS), virtualizing foundation models (VFM), unifying the management of state, memory, resources, and security governance, and providing pluggable knowledge, model, and trust subsystems.
Position: It’s Time to Optimize LLMs for Self-Consistency
Itamar Pres (Massachusetts Institute Of Technology), Jacob Andreas
OptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextMultimodalityChain-of-Thought
🎯 What it does: Propose a unified self-consistency framework that optimizes the behavioral consistency of large language models by leveraging cross-input relational constraints, covering a variety of capabilities from invariance, equivariance to self-description and self-improvement;
Position: Knowing Isn’t Understanding: Re-grounding Generative Proactivity with Epistemic and Behavioral Insight
Kirandeep Kaur (University of Washington), Chirag Shah (University of Washington)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackAgentic AIReview/Survey Paper
🎯 What it does: Proposed a generative proactive framework based on the coupling of cognition and behavior, emphasizing that proactivity requires balancing between knowledge legitimacy and action commitment.
Position: Large Language Models Should Learn Personalized Rather Than Aggregated Human Preferences
Cristina Garbacea (University of Chicago)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextReview/Survey Paper
🎯 What it does: This paper argues that large language models (LLMs) should shift from learning aggregated human preferences to learning personalized preferences. It systematically analyzes the theoretical and empirical shortcomings of aggregated preferences, reviews existing personalization technologies and risks, and proposes a 'bounded personalization' framework along with governance recommendations.
Position: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance
Shiqiang Wang (University of Exeter), Mingyue Ji (University of Florida)
Data SynthesisExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextSequential
🎯 What it does: Propose the data probes method, which uses known stochastic processes to generate controllable synthetic sequences, systematically studying how these data properties affect the training, fine-tuning, and inference of LLMs.
Position: LLM Agents Are the Antidote to Walled Gardens
Samuele Marro (University of Oxford), Philip Torr (University of Oxford)
Autonomous DrivingOptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyRobotic IntelligenceAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringAuto EncoderGenerative Adversarial NetworkTextTabularSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the use of intelligent agents built with large language models (LLMs) to achieve 'general interoperability' by automatically translating data formats and interacting with human interfaces, thus breaking down the 'walls' of closed platforms and promoting seamless exchange of data and functions.
Position: LLM Benchmark Datasets Should Be Contamination-Resistant
ALI AL LAWATI, Suhang Wang (Pennsylvania State University)
Safty and PrivacyAdversarial AttackData-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes and implements a corruption-resistant benchmark dataset (CRD) scheme based on the training-inference asymmetry of Transformer models;
Position: LLM for Physics Research Requires Domain-Specialized Training and Tooling
Sirui Lu (Max-Planck-Institut für Quantenoptik), Bernhard Schölkopf (MPI for Intelligent Systems)
TransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringDiffusion modelScore-based ModelTextMultimodalityReview/Survey PaperBenchmarkPhysics RelatedRetrieval-Augmented GenerationStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes introducing specialized language models and toolchains into theoretical physics research, aiming to address the shortcomings of existing large language models in terms of physical intuition, constraint satisfaction, and reliable reasoning, thereby enabling them to automate the complete scientific research process, including literature review, model building, symbolic derivation, numerical simulation, and experimental design.
Position: LLM Serving Needs Mathematical Optimization and Algorithmic Foundations, Not Just Heuristics
Zijie Zhou (Hong Kong University of Science and Technology)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsTextReview/Survey Paper
🎯 What it does: Elaborates that LLM inference services have surpassed the limitations of traditional heuristic scheduling and routing, calling for the design of new decision-making mechanisms centered on mathematical optimization and algorithm theory.
Position: LLM-Based Social Simulations Require a Boundary
Zengqing Wu (University of Osaka), Chuan Xiao (University of Osaka)
Data-Centric LearningTransformerLarge Language ModelTextTabularTime SeriesSequentialReview/Survey Paper
🎯 What it does: This paper systematically evaluates the performance of large language models (LLMs) in social simulations, pointing out that they tend to generate 'averaged personas,' leading to homogenized behaviors, and provides a review of validation practices in 21 recent papers.
Position: LLM-Safety Evaluations Lack Robustness
Tim Beyer (Technical University of Munich), Stephan Günnemann
Safty and PrivacyExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper systematically analyzes the LLM security evaluation process, identifying noise, bias, and inconsistencies in stages such as dataset selection, algorithm implementation, sampling, and discrimination, and proposes improvement guidelines.
Position: LLMs can't jump
Tom Zahavy (Google DeepMind)
TransformerLarge Language ModelWorld ModelTextMultimodalityReview/Survey PaperPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper uses Einstein's development of the general theory of relativity as a case study to systematically analyze the limitations of existing large language models in scientific discovery. It points out that these models lack the mechanism of achieving 'abductive jumps' through sensory simulation, and thus cannot generate new postulates or hypotheses.
Position: LLMs Should Incorporate Explicit Mechanisms for Human Empathy
Xiaoxing You (Harbin Institute of Technology (Shenzhen)), Jun Yu (Harbin Institute of Technology (Shenzhen))
GenerationRecommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes that LLMs should possess an explicit human empathy mechanism, defining empathy as the accurate reproduction of human perspectives, intentions, emotions, and contexts, and explores the empathy deficiencies of current LLMs under four mechanisms: emotional decay, detail mismatch, conflict avoidance, and linguistic distance.
Position: Machine Learning for Heart Transplant Allocation Policy Optimization Should Account for Incentives
Ioannis Anagnostides (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
OptimizationFederated LearningExplainability and InterpretabilityReinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningTabularBiomedical DataElectronic Health RecordsReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper takes heart transplant allocation as an example, explaining the game-theoretic behaviors of stakeholders such as hospitals, OPOs, doctors, and patients under the current rule-based priority system where incentives are misaligned, and proposes that future machine learning-driven allocation strategies must be incentive-aware to improve efficiency, justice, and trust.
Position: Machine Learning Research Should Be Guided by Explicit, Pluralistic Models of Human Purpose
Utsav Gupta (Stanford University)
Federated LearningSafty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackReview/Survey Paper
🎯 What it does: Propose practical norms for incorporating human purposes explicitly into machine learning research, including purpose statements, purpose evaluation, and governance mechanisms.
Position: Make Planning Research Rigorous Again!
Michael Katz (IBM), Sarath Sreedharan (Colorado State University)
Data-Centric LearningLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes that when using large language models (LLMs) for planning, rigorous methods and tools from the field of automated planning should be adopted, and unified evaluation, data, and tool usage standards should be established;
Position: Mechanisms for Aggregated Individual Reporting Should be Established for Post-Deployment Evaluation
Jessica Dai (University of California, Berkeley), Irene Y. Chen (University of California, Berkeley)
Federated LearningSafty and PrivacyExplainability and InterpretabilityLarge Language ModelPrompt EngineeringTextTabularTime SeriesBiomedical DataReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Aggregated Individual Reports (AIR) mechanism for post-deployment evaluation of AI systems, and presents the framework, design decisions, and implementation path.
Position: Medical AI Neglects Real Treatment Outcomes
Shiva Kaul (Harvard Pilgrim Health Care Institute and Harvard Medical School), Anjum Khurshid (Harvard Pilgrim Health Care Institute and Harvard Medical School)
Recommendation SystemFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyData-Centric LearningDrug DiscoveryPrompt EngineeringTextTabularBiomedical DataElectronic Health RecordsReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: Analyze current training and evaluation practices in medical AI, pointing out their neglect of real-world treatment outcomes, and propose a new framework that uses longitudinal real patient data for training and randomized controlled trial (RCT) results for evaluation.
Position: Metaphysical Concepts in AI Should Be Judged by Their Consequences
Paras Chopra (Lossfunk)
Explainability and InterpretabilityTextReview/Survey Paper
🎯 What it does: Propose a two-step framework called 'productive confusion,' which evaluates the value of AI metaphysical concepts using a philosophical approach, with actual consequences as the criterion;
Position: Model identity in machine learning is a convention, not a property
Vacslav Glukhov (ItoFlow)
Federated LearningExplainability and InterpretabilityReview/Survey Paper
🎯 What it does: This paper examines the assumption that machine learning models are regarded as stable objects in practical applications and governance, analyzes the limitations of identity recognition based on functional behavior and internal structure, and proposes the minimal prerequisites and hierarchical equivalence relations required for model identity recognition.
Position: Modular Memory is the Key to Continual Learning Agents
Vaggelis Dorovatas (Toyota Motor Europe), Rahaf Aljundi (Toyota Motor Europe)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Proposes a continual learning framework based on modular memory, combining the core capabilities of large pre-trained models with working memory and long-term memory to achieve rapid context adaptation and stable parameter updates.
Position: Modular Safety Guardrails Are Necessary for Foundation-Model-Enabled Robots in the Real World
Joonkyung Kim (Texas A&M University), Yan Gu (Purdue University)
Safty and PrivacyRobotic IntelligenceTransformerReinforcement LearningPrompt EngineeringVision-Language-Action ModelContrastive LearningImageVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes a modular safety guard architecture for open-world deployment of robots driven by foundation models, including a monitoring and evaluation layer and an intervention layer (decision gate and action gate), and provides cross-layer collaborative design principles.
Position: Multi-Agent Explainability Needs Contracts Before Methods
Hak Hyun Kim (Dartmouth College), Soroush Vosoughi (Dartmouth College)
Explainability and InterpretabilityTextReview/Survey Paper
🎯 What it does: Propose defining research contracts and agent contracts in the study of explainability in multi-agent systems to standardize explanation goals, audiences, evaluation criteria, and behavioral baselines.
Position: Multi-Agent Systems Should Prioritize Concurrency Control
Xin Yang (Zhejiang University), Wenyuan Jiang (ETH Zurich)
Autonomous DrivingOptimizationFederated LearningComputational EfficiencyAI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose to treat concurrency control as a core design focus in LLM multi-agent systems, and map failure modes of multi-agent systems to traditional concurrency anomalies;
Position: Multiple Definitions & Unrealistic Assumptions of Model Collapse Distract from Real World Threats
Rylan Schaeffer (Stanford University), Sanmi Koyejo (Stanford University)
Review/Survey Paper
🎯 What it does: Systematically review the literature on model collapse, identify and summarize eight definitions, point out the discrepancy between their assumptions and reality, and suggest focusing on real risks.
Position: Multiplicity is an Inevitable and Inherent Challenge in Multimodal Learning
Sanghyuk Chun (Princeton University), Olga Russakovsky (Princeton University)
Representation LearningData-Centric LearningVision Language ModelContrastive LearningImageTextMultimodalityReview/Survey Paper
🎯 What it does: This paper systematically explains the inevitable multiplicity problem in multimodal learning from both theoretical and empirical perspectives, analyzes its impact on data construction, training, and evaluation, and further proposes an improved approach that is aware of multiplicity.
Position: Natural Language Should Not Fully Replace Formal Languages
Eitan Wagner (Hebrew University of Jerusalem), Omri Abend (Hebrew University of Jerusalem)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelPrompt EngineeringDiffusion modelScore-based ModelContrastive LearningImageTextMultimodalityAudio
🎯 What it does: Propose an information-theoretic framework that explains the intersection of natural language and formal language in terms of task specificity, and verify through theoretical derivation and multimodal case studies (image generation, code synthesis, audio creation, etc.) that natural language is more efficient in low-specificity tasks, while formal language has advantages in high-specificity tasks.
Position: Neglecting the Sustainability of AI is Fuelling a Global AI Arms Race
Pedram Bakhtiarifard (University of Copenhagen), Raghavendra Selvan (University of Copenhagen)
TextTabularReview/Survey Paper
🎯 What it does: Analyzes and points out that neglecting AI sustainability leads to a global AI arms race, and proposes the 'CARAML' framework based on Marx's base-superstructure theory, emphasizing the synergistic role of climate and resource awareness.
Position: Neural Approximation Is Rarely Justified for Hard Combinatorial Problems
Pritish Chakraborty (Indian Institute of Technology Bombay), Abir De (Indian Institute of Technology Bombay)
OptimizationGraphTabularBenchmark
🎯 What it does: Investigate the rationality of using neural networks to approximate NP-hard combinatorial optimization problems, pointing out their shortcomings in terms of data, computation, guarantees, distribution drift, and decoder dependency;
Position: Peer Review in ML/AI Conferences Should Separate Publication from Presentation and Offer Non-Anonymous Review Tracks
Nihar B Shah (Carnegie Mellon University)
Large Language ModelTextReview/Survey Paper
🎯 What it does: Proposes two structural reform proposals for peer review in ML/AI conference proceedings: first, separate publication from presentation, adopting a four-step process to first evaluate technical feasibility and publish all qualified papers, then select presentations through community voting; second, introduce anonymous and open (non-anonymous) review tracks, making full review data publicly available to enhance transparency and accountability.
Position: Peer Review Should Be Calibrated via LLM Scoring
Zijin Chen (Huazhong University of Science and Technology), Qinbin Li (Huazhong University of Science and Technology)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper proposes and implements a calibration layer based on a large language model (LLM) to map textual justifications provided by reviewers during paper reviews to unified 'anchor' scores. It then calculates the residual between the anchor scores and the self-reported scores of the reviewers, and triggers a lightweight 'post-check' request when the residual is large, prompting the reviewers to reassess or supplement their justifications.
Position: Predicting AI’s Impact on Labor Is a Core Machine Learning Problem
Yong Suk Lee (University of Notre Dame)
Recommendation SystemOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsContrastive LearningTextTabularTime SeriesSequentialReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper argues that predicting the impact of artificial intelligence on the labor market should be regarded as a core machine learning problem. It reviews existing methods from economics, management, and machine learning, identifies technical challenges, and proposes a feasible research agenda, including data and measurement frameworks, workflow benchmarks, non-stationary prediction methods, LLM agent simulations, and interdisciplinary evaluation systems.
Position: Predictive Uncertainty Is Not Enough — Joint Distribution for Full Uncertainty Representation
Adria Aldoma (Barcelona Supercomputing Center), Axel Brando (Barcelona Supercomputing Center)
Information TheoryClassificationAnomaly DetectionTransformerMixture of ExpertsFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageBenchmark
🎯 What it does: This paper argues that relying solely on predictive uncertainty (epistemic + aleatoric) is insufficient to fully assess model risk, and proposes three sources of uncertainty (domain, model, data noise), achieving comprehensive representation through the joint distribution p(x,y|D)=p(x|D)·p(y|x,D).
Position: Preparing for AI Systems That Deceive Developers
Isabella Duan (Safe AI Forum), Min Yang (Fudan University)
Safty and PrivacyExplainability and InterpretabilityTransformerPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes three security recommendations for AI developers, emphasizing the monitorability during the training phase, the completeness of evaluation, and the irrevocable control before deployment, aiming to prevent AI systems from misleading developers through deceptive means and weakening subsequent security measures.
Position: Preregister Experiments with AI Agents
Michelle Vaccaro (Massachusetts Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextReview/Survey Paper
🎯 What it does: This paper proposes and elaborates on preregistration practices for behavioral experiments using large language models (LLMs) and autonomous AI agents, aiming to enhance the credibility and reproducibility of such experiments.
Position: Preventing AI-Generated CSAM Necessitates New Approaches to AI Safety
Neil Kale (Carnegie Mellon University), Virginia Smith (Carnegie Mellon University)
Safty and PrivacyPrompt EngineeringDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper points out that due to legal and ethical restrictions, existing AI safety technologies cannot fully prevent AI-generated child sexual abuse material, and systematically organizes 15 key technical and policy gaps throughout the lifecycle, including data acquisition, evaluation, deployment, and maintenance, and proposes specific action recommendations for researchers, developers, and policymakers.
Position: Prioritize Identifying Structure, Not Complex Models, for Scientific Discovery
Tyler McCormick
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningLarge Language ModelPrompt EngineeringContrastive LearningTabularAgriculture Related
🎯 What it does: This paper argues that in high-dimensional proxy data environments, predictive performance alone is insufficient to determine mechanisms, and it proposes that explicit identification structures (including mechanism constraints, observational process assumptions, and experimental design) must be explicitly stated to achieve scientific discovery; subsequently, it demonstrates through simulated pea genetics experiments that models can have similar predictive performance under different designs but show significant differences in mechanism consistency.
Position: Privacy Is a Claim, Not a Property of Synthetic Data
Jiachen Zhao (University of Notre Dame), Taeho Jung (University of Notre Dame)
Safty and PrivacyData-Centric LearningTextReview/Survey Paper
🎯 What it does: This paper argues from three aspects—conceptual, empirical, and normative—that synthetic data does not guarantee privacy, emphasizing that privacy should be viewed as a verifiable scientific claim. Based on an audit of papers from ICML, NeurIPS, and ACL 2024–2025, it proposes a 'Minimum Privacy Statement Standard.'
Position: Profiling Game Worlds by Transition Complexity
Lele Cao (King AI Labs, Microsoft Gaming)
Recurrent Neural NetworkTransformerReinforcement LearningContrastive LearningWorld ModelImageSequential
🎯 What it does: Proposed Transition Complexity Profile (TCP), a repeatable and quantifiable metric system for describing the transition complexity of a given game environment under specific information interfaces (pixels/marks/potential variables + limited history).
Position: Prompting Intent Should Be Audited in LLM-Assisted Peer Review
Lijinghua Zhang (University of California, Irvine), Hengrui Cai (University of California, Irvine)
ClassificationData SynthesisExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: The study proposes and implements an LLM-assisted peer review auditing framework based on prompt intent, utilizing synthetically generated review texts to train an intent classifier, and applying it to ICLR 2026 reviews marked as using LLMs.