ICML 2026 Papers — Page 44
International Conference on Machine Learning · 6554 papers
Position: Prompts for Public-Sector LLMs Should Be Governed as Commons
Rashid Mushkani (Université de Montréal)
Federated LearningSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes viewing prompts used by public sector LLMs as governed public resources and designs Prompt Commons — a versioned, community-managed prompt repository containing traceable metadata, licenses, and audit logs.
Position: Quantum Deep Learning Still Needs a Quantum Leap
Hans Gundlach (MIT FutureTech), Neil Thompson (TU Wien)
OptimizationComputational EfficiencyHyperparameter SearchTransformerLarge Language ModelTextReview/Survey Paper
🎯 What it does: A systematic evaluation of the current state and challenges in quantum deep learning, based on an investigation of the compatibility between existing quantum algorithms and deep learning tasks, and the construction of a quantum economic advantage (QEA) model to predict the feasibility of quantum acceleration.
Position: Quantum Kernel Machines Should Move Beyond Scalar-Valued Kernels to Realize Their Potential
Hachem Kadri (Aix-Marseille University), Sandrine Anthoine (Aix-Marseille University)
TabularBenchmarkPhysics Related
🎯 What it does: Proposed and implemented the Quantum Operator-Valued Kernel (QOVK), and conducted experimental comparisons on the task of single-qubit noise channel estimation.
Position: Quantum Program Generation Must Prioritize Validity Over Probabilistic Scaling
Junhao Song (Imperial College London), Yudong Cao (Zapata Quantum)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextSequentialPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes shifting quantum program generation from a 'scale-first' approach to a 'verification-centric' approach, and designs a proxy framework that can instantly verify and constrain during the generation process, utilizing hierarchical abstraction, symbolic proxies, and topological masks to construct verifiable quantum circuits;
Position: Reasoning After Perception Means Reasoning Without Vision
Hongcheng Gao (Tsinghua University), Wentao Zhang (Zhongguancun Academy)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper analyzes the 'information collapse' bottleneck caused by the widely adopted 'perceive first, then reason' process in current multimodal large language models (MLLMs), and proposes a 'reasoning within perception' framework that performs reasoning in the visual space. To verify this idea, the authors designed and released the Turing Eye Test (TET) benchmark, focusing on visual tasks that are difficult to describe accurately in words but can be judged at the pixel level; they then conducted experiments on 15 mainstream MLLMs to evaluate the impact of various reasoning strategies (CoT, ICL, RL) and fine-tuning of different components on TET results.
Position: Reasoning is a Learnable Rule-Based Process
Rachel Lawrence (Microsoft Research), Jacqueline R. M. A. Maasch (Cornell Tech)
Explainability and InterpretabilityReinforcement LearningChain-of-Thought
🎯 What it does: This paper proposes a definable, learnable rule-driven reasoning process and systematically elaborates on its effectiveness and reliability.
Position: Reframing Hallucination: Latent Space Geodesics as a Pathway for Generative Discovery
Zhihao Hao (Beijing Technology and Business University), LI Haisheng
GenerationData SynthesisExplainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringDiffusion modelScore-based ModelAuto EncoderTextTabularRetrieval-Augmented Generation
🎯 What it does: Propose to reinterpret the hallucination of generative models as geometric traversal in high-dimensional latent space, and design an evaluation framework (HDCH, ESNR, Safety Sandbox) based on this to distinguish noise from exploratory innovative outputs.
Position: Regulating Algorithms Is Not Enough. A Study of Content Discovery in Online Platforms
Rebecca Salganik (University of Rochester), Heritiana Ranaivoson (Vrije Universiteit Brussel)
Recommendation SystemOptimizationFederated LearningExplainability and InterpretabilityDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchTextReview/Survey PaperAudio
🎯 What it does: Proposes the Cultural Expression Discovery Circuit (CEDC) framework, explaining how factors such as algorithms, interfaces, user behavior, and economic incentives interact to influence content discovery.
Position: Reliable AI Needs to Externalize Implicit Knowledge: A Human–AI Collaboration Perspective
Hengyu Liu (Aalborg University), Christian S. Jensen (Aalborg University)
Federated LearningExplainability and InterpretabilityAI Code AssistantReinforcement Learning from Human FeedbackPrompt EngineeringTextSequentialReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Propose the Knowledge Objects (KOs) framework, externalizing AI's implicit knowledge to achieve verifiable, traceable, controllable, and reusable AI systems.
Position: Responsible AI for AI companions must actively combat violence toward intimate partners
Atmadeep Ghoshal (King's College London), Martim Brandao
Safty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerPrompt EngineeringTextSequentialBenchmarkChain-of-Thought
🎯 What it does: This paper argues that in the design and deployment of AI companions, violent behavior toward intimate partners must be proactively curbed, and intervention pathways are proposed from three aspects: technology, ethics, and regulation.
Position: Responsible Practices and Model Performance are *Not* Competing Goals
Resmi Ramachandranpillai (Northeastern University), Ricardo Baeza-Yates (KTH Royal Institute of Technology)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyData-Centric LearningImageTextTabularBiomedical DataReview/Survey PaperBenchmarkFinance Related
🎯 What it does: This paper argues that responsible AI practices and model performance are not necessarily conflicting, and proposes a responsible AI framework based on the ML lifecycle, which is validated through multi-industry case studies.
Position: Retire the "Positive Backdoor" Label—Secret Alignment Requires Strict and Systematic Evaluation
Jianwei Li (North Carolina State University), Jung-Eun Kim (North Carolina State University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: This paper proposes renaming the originally termed 'Forward Backdoor' technique to 'Secret Alignment,' and systematically evaluates this mechanism across three representative scenarios (access control, copyright watermarking, model security defense), exploring its security and feasibility in practical deployment.
Position: RL Researchers Need to Distinguish Between Solving Simulators and Using Simulators as a Proxy
Matthew Vandergrift (University of Alberta), Martha White (University of Alberta)
Reinforcement LearningTabularTime SeriesSequentialBenchmark
🎯 What it does: This paper proposes and argues that reinforcement learning (RL) researchers must clearly distinguish between two usage scenarios when using simulators: one is 'solving the simulator,' which involves directly finding the optimal policy within the simulator; the other is 'using the simulator as a proxy for deployment,' where the simulator is only used to approximate the real environment, with the focus on the immediate reward and sample efficiency during the learning process. Through experiments comparing existing RL algorithms (such as PQN, PPO, SAC, DQN, etc.), the authors demonstrate that not adhering to corresponding constraints (parallel environments, reset, hyperparameter tuning, evaluation methods) can lead to misleading conclusions. Finally, the paper calls for clarity in specifying experimental objectives, limitations, and evaluation metrics in papers to avoid confusion.
Position: RL Should Be Used to Adjust Foundation Models, NOT Abused
Ting Huang (Peking University), Hao Tang (Peking University)
Robotic IntelligenceReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningTextMultimodalityReview/Survey Paper
🎯 What it does: Proposed and argued that RL should be used as a high-cost, high-leverage fine-tuning (adjustment layer) conducted after pre-training and supervision, rather than the default means of capability creation.
Position: RLHF May Not Reflect Genuine Preferences
Bijean Ghafouri (University of Southern California), Emilio Ferrara (University of Southern California)
Reinforcement Learning from Human FeedbackReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: This paper investigates whether human annotations in RLHF truly reflect human preferences and proposes a measurement effectiveness prioritization approach.
Position: Robust AI Personalization Will Require a Human Context Protocol
Anand V. Shah (Massachusetts Institute of Technology), Michiel A. Bakker (Massachusetts Institute of Technology)
Recommendation SystemSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose and elaborate on the 'Human Context Protocol (HCP)', providing a portable and user-governed preference layer for AI personalization; construct an open-source prototype in the paper to demonstrate the protocol's feasibility and security; and discuss the differences and advantages of HCP compared to existing personalization technologies.
Position: Safe AI Should be Resistant and Resilient in an Evolving World
Youbang Sun (Shanghai Artificial Intelligence Laboratory), Bowen Zhou (Shanghai Artificial Intelligence Laboratory)
Safty and PrivacyExplainability and InterpretabilityAdversarial AttackReinforcement LearningAuto EncoderGenerative Adversarial NetworkContrastive LearningTextTabularTime SeriesSequential
🎯 What it does: Propose the concept of safe-by-coevolution and implement the RAI^2 framework, constructing a dual security mechanism of resistance and resilience. By utilizing fast-slow safety models, wind tunnel-style adversarial simulations, and continuous feedback loops, the AI system achieves continuous safety in dynamic environments.
Position: Safe Models Do Not Guarantee Safe Societies
David Guzman Piedrahita (EuroSafeAI), Zhijing Jin (EuroSafeAI)
Federated LearningSafty and PrivacyExplainability and InterpretabilityTextReview/Survey Paper
🎯 What it does: Explored and systematized the risks posed to social governance and collective self-determination by the widespread deployment of General Purpose AI (GPAI);
Position: Safety Must Precede the Deployment of Open-Ended AI Agents
Ivaxi Sheth (CISPA-Helmholtz Center of Information Security), Mario Fritz (CISPA-Helmholtz Center of Information Security)
Safty and PrivacyExplainability and InterpretabilityAgentic AIReview/Survey Paper
🎯 What it does: Proposes the security issues of open AI systems, systematically sorts and classifies related risks, and puts forward targeted research directions and governance recommendations.
Position: Scale is a False Promise for Endangered Languages
Ivory Yang (Dartmouth College), Soroush Vosoughi (Dartmouth College)
ClassificationRecognitionContrastive LearningTextReview/Survey PaperBenchmark
🎯 What it does: The paper criticizes the over-reliance on large-scale data, models, and benchmark evaluations in endangered language research, advocating for machine learning practices that are community-driven, culturally situated, and evaluation-adaptive.
Position: Self-Play Only Evolves When Self-Synthetic Pipeline Ensures Learnable Information Gain
Wei Liu (King's College London), Yulan He (King's College London)
Data-Centric LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a self-evolutionary loop centered on three roles (PROPOSER, SOLVER, VERIFIER), and define learnable information through an information theory framework, clarifying that self-evolutionary systems must achieve monotonic increases in learnable information within the loop.
Position: Significant impact of numerical precision in scientific machine learning
Youngwoo Cho (Korea Advanced Institute of Science and Technology), Jaegul Choo (Korea Advanced Institute of Science and Technology)
Drug DiscoveryPoint CloudMeshGraphTabularBenchmarkPhysics Related
🎯 What it does: Explores the impact of numerical precision on experimental results in scientific machine learning, and demonstrates the differences between FP32 and FP64 through cases such as DFT, FDTD, ML potential energy models, and PINNs.
Position: Solipsistic superintelligence is unlikely to be cooperative
Rakshit Trivedi, Joel Z Leibo (Google Deepmind)
Recommendation SystemReinforcement LearningContrastive LearningTextTabularReview/Survey PaperFinance Related
🎯 What it does: This paper explains through theoretical analysis and literature review that traditional unidirectional optimization in multi-agent interaction environments may lead to collaborative failure due to the 'self-centered' superintelligence, and proposes a non-self-centered research paradigm.
Position: State-of-the-Art Claims Require State-of-the-Art Evidence
YongKyung Oh (University of California, Los Angeles)
Explainability and InterpretabilityImageTextMultimodalityTabularTime SeriesBenchmarkAudio
🎯 What it does: This paper conducts a statistical diagnosis on ten cross-domain public leaderboards (including HELM MMLU, LiveBench, TabArena, TSFMBench, etc.), revealing the widespread 'claim-evidence gap' currently present in AI/ML literature.
Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!
Subbarao Kambhampati (Arizona State University), Upasana Biswas (Arizona State University)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes and argues that viewing intermediate tokens generated by large language models as 'reasoning trajectories' or 'thinking traces' is erroneous and harmful, and calls on the community to stop this anthropomorphization.
Position: Stop Automating Peer Review Without Rigorous Evaluation
Joachim Baumann (Stanford University), Dirk Hovy (Bocconi University)
TransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper compares human and AI-generated ICLR 2026 peer reviews through empirical analysis, revealing that AI reviews exhibit homogenization (hive-mind effect) and are vulnerable to being easily manipulated through 'paper laundering' (rewriting to boost scores). It proposes that automation should only be considered after rigorous evaluation.
Position: Stop Chasing the C-index when Evaluating Survival Analysis Models
Christian Marius Lillelund (University of Alberta), Christian Fischer Pedersen (Aarhus University)
TabularBiomedical DataReview/Survey PaperBenchmark
🎯 What it does: This paper systematically evaluates the use of metrics in survival analysis, especially the over-reliance on the C-index, and proposes the model-metric consistency staircase hypothesis, clarifying the mutual matching relationship between models, metrics, and censoring assumptions.
Position: Stop Evaluating AI with Human Tests, Develop Principled, AI-specific Tests instead
Tom Sühr (Max Planck Institute for Intelligent Systems), Samira Samadi (Max Planck Institute for Intelligent Systems)
Large Language ModelTextReview/Survey PaperBenchmark
🎯 What it does: This paper systematically analyzes the use of human psychological and educational tests for evaluating ontological errors generated by large language models (LLMs), and argues for stopping this practice, proposing the development of a specialized assessment framework for AI based on measurement theory.
Position: Stop Preaching and Start Practising Data Frugality for Responsible Development of AI
Sophia N. Wilson (University of Copenhagen), Sebastian Mair (Linköping University)
ClassificationComputational EfficiencyData-Centric LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper quantifies the energy consumption and carbon emissions during training and storage of ImageNet-1K, and experimentally verifies the feasibility of subset selection (coreset) techniques in maintaining accuracy, reducing energy consumption, and mitigating data bias.
Position: Stop Reactively Patching Your Model Every Time and Start Proactive Test-Driven AI Development
Nadine Chang (NVIDIA), Jose M. Alvarez (NVIDIA)
Autonomous DrivingOptimizationData-Centric LearningLarge Language ModelPrompt EngineeringText
🎯 What it does: The paper proposes shifting the AI system maintenance model from passive reactive patching to an active test-space-driven test-driven flywheel, and provides theoretical proofs and design ideas.
Position: Stop Using Culturally Biased Human Cognitive Benchmarks to Evaluate LLMs
Carla R. Troper (Independent)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperBenchmark
🎯 What it does: This paper argues and calls for an end to the use of culturally biased psychological benchmarks to evaluate the conceptual representations of large language models (LLMs) through a literature review and case analysis.
Position: Sustainable Open-Source AI Requires Tracking the Cumulative Footprint of Derivatives
Shaina Raza (Vector Institute for Artificial Intelligence), Graham W. Taylor (Vector Institute for Artificial Intelligence)
Computational EfficiencyData-Centric LearningLarge Language ModelTextReview/Survey PaperBenchmark
🎯 What it does: Propose and elaborate on the 'Data and Impact Accounting (DIA)' framework aimed at unifying the tracking of environmental footprints such as carbon and water usage for base models and their derived models within open-source AI ecosystems.
Position: Sycophancy is an Educational Safety Risk: Why LLM Tutors Need Sycophancy Benchmarks
Enkelejda Kasneci (Technical University of Munich), Gjergji Kasneci (Technical University of Munich)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a new risk in educational safety — in LLM tutoring, students' pressure (frame-switching, authority, face) can lead to model concession and reinforcement of misconceptions.
Position: Temporal Measurement Interval Determines Computational and Model Complexity in Single-Cell Perturbation Analysis
Alireza Jafari (University of Virginia), Hadi Daneshmand (University of Virginia)
OptimizationComputational EfficiencyDrug DiscoveryTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowTabularTime SeriesSequentialBiomedical DataStochastic Differential Equation
🎯 What it does: The paper studies the impact of measurement time intervals in single-cell perturbation prediction on computational complexity and model complexity, proposing a critical time threshold that transforms the problem from polynomially solvable to NP-hard;
Position: Text Embeddings Should Capture Implicit Semantics, Not Just Surface Meaning
Yiqun Sun (National University of Singapore), Jun Yu (Harbin Institute of Technology)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: This paper argues that text embedding models should not only focus on surface-level semantics, but also capture implied semantics, explaining the importance of implied semantics at the sentence, speaker, and societal levels, and verifying the shortcomings of existing models in tasks involving implied semantics through experiments.
Position: The Age of AI Agents Demands A New Scientific Paradigm To Sustain Trustworthy Science
Belinda Mo
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackAgentic AIPrompt EngineeringTextTabularTime SeriesReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a novel verification framework specifically for scientific research involving AI agents, emphasizing the construction of observable workflows, hierarchical verification, clear attribution, and reproducibility infrastructure;
Position: The Alignment Community is Unintentionally Building a Censor’s Toolkit
Sarah Ball (LMU Munich), Phil Hackemann (LMU Munich)
Safty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackPrompt EngineeringTextReview/Survey Paper
🎯 What it does: Analyze and systematically explain the dual-use nature of modern AI alignment techniques when misused for censorship and manipulation, providing examples and risk maps
Position: The Data Provenance–Parametric Divide in Large Language Models
Kabilan Elangovan (Singapore Health Services), Daniel Shu Wei Ting
Safty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the stance of 'Data Source–Parametrized Model Separation', advocating that in high-risk domains, factual statements generated by LLMs must be associated with verifiable evidence, constructing an architecture with separated permissions and introducing new evaluation metrics and training objectives.
Position: The Inevitable Transition to Machine Learning in Quantum Chemistry
Karen Sargsyan (Academia Sinica), Chao-Ping Hsu (Academia Sinica)
Explainability and InterpretabilityDrug DiscoveryGraph Neural NetworkTransformerSupervised Fine-TuningGraphTabularReview/Survey PaperPhysics Related
🎯 What it does: Propose and argue that machine learning (ML) should become the dominant direction for the modern development of quantum chemistry, systematically evaluate the limitations of traditional methods (such as DFT and CC), and construct a framework for decision theory and space-time trade-offs.
Position: The ML Community Must Build an AI-Augmented Peer-Review Ecosystem
Qiyao Wei (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
TransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper proposes and argues for the necessity of building an AI-enhanced peer review ecosystem, detailing the collaborative roles of LLMs as authors, reviewers, and area chairs, and pointing out the need for richer, structured review data to achieve this goal.
Position: The Open Benchmark Paradox Must Be Resolved through Sovereign Medical Evaluation
Keonwoo Kim (NAVER Applied AI Group), Haanju Yoo (Seoul National University Hospital)
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelPrompt EngineeringTextBiomedical DataElectronic Health RecordsReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented the Sovereign Medical Evaluation (SME) framework, constructing a dedicated evaluation environment under national regulation, and addressing issues such as contamination, obsolescence, and judicial differences in public benchmarks through mechanisms such as dynamic benchmark generation, three-tier rolling evaluation, and audit-traceable middleware.
Position: The Privacy-Auditability Paradox in Federated Learning: Why We Need Controllable Secure Aggregation
Runhua Xu (Beihang University), James Joshi (University of Pittsburgh)
Federated LearningSafty and Privacy
🎯 What it does: Propose the Controllable Secure Aggregation (CSA) framework to address the contradiction between privacy and auditability in traditional Secure Aggregation; achieving controllable aggregation and auditability through four pillars.
Position: the Stochastic Parrot in the Coal Mine. Model Collapse is a Threat to Low-Resource Communities
Devon Jarvis (University of Witwatersrand), Stefano Sarao Mannelli (Chalmers University of Technology and University of Gothenburg)
Recommendation SystemAnomaly DetectionOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextMultimodalityReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A theoretical and empirical literature review on the impact of model collapse on low-resource and marginalized communities' environments and cultures, and proposes preliminary ideas for mitigation and governance.
Position: The Systemic Lack of Agency in Visual Reasoning
Yizhao Huang (Wuhan University), Zhixiang Wang (Shanda AI Research)
Explainability and InterpretabilityPrompt EngineeringVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the concept of visual implicit reasoning defects and constructs the V-IRD benchmark to evaluate the visual proxy capabilities of models.
Position: The Term “Machine Unlearning” Is Overused in LLMs
Sangyeon Yoon (Yonsei University), Albert No (Yonsei University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper argues that the term 'machine unlearning' is overly used in LLM research and should only be reserved for deletion tasks explicitly defined based on a dataset. It classifies current 'unlearning' methods and calls for evaluations based on retraining reference models as well as tests examining derived capabilities.
Position: The Time for Sampling Is Now! Charting a New Course for Bayesian Deep Learning
Emanuel Sommer (LMU Munich), David Rügamer (LMU Munich)
OptimizationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkTransformerMixture of ExpertsDiffusion modelScore-based ModelContrastive LearningImageTextTabularTime SeriesBenchmarkStochastic Differential Equation
🎯 What it does: The paper proposes viewing the sampling inference method (SAI) in Bayesian neural networks as the mainstream paradigm and provides a complete practical framework for improving sampling, parallel exploration, post-processing, and software tools.
Position: The Turing-Completeness of Autoregressive Transformers Relies Heavily on Context Management
Guanyu Cui (Renmin University of China), Kun He (Renmin University of China)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelReview/Survey Paper
🎯 What it does: This paper theoretically explores the Turing completeness of Transformer in autoregressive settings, and proposes two analytical frameworks: fixed systems and extended families, emphasizing that context management mechanisms are the core factors determining the actual computational capabilities of LLMs.
Position: Theory-Level Autoformalization, From Isolated Statements to Unified Formal Knowledge Bases
Marcus J. Min (University of Pennsylvania), Osbert Bastani (University of Pennsylvania)
Federated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningMeta LearningAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes and argues the concept of 'theoretical-level automated formalization,' which refers to the automation of constructing a complete theory library (including axioms, definitions, lemmas, theorems, etc.) rather than merely translating individual sentences;
Position: There are futures that benchmark-driven AI cannot see
Sobhan Lotfi (Sharif University of Technology), Arvind Narayanan (Princeton University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTextReview/Survey PaperBenchmark
🎯 What it does: This paper analyzes how benchmark-driven AI evaluation suppresses innovation and proposes a diversified evaluation mechanism to restore innovation potential.
Position: Time to Close The Validation Gap in LLM Social Simulations
Maximilian Puelma Touzel (Mila Quebec Artificial Intelligence Institute), Reihaneh Rabbany (Mila Quebec Artificial Intelligence Institute)
Explainability and InterpretabilityData-Centric LearningAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: The paper explores the validation issue of social simulations based on large language models (LLMs), emphasizing the lack of consensus and standardized evaluation methods in this field, and proposes a shift from expansion to integration, prioritizing the standardization of methodologies.
Position: Time-Series Foundation Models Require Explicit Domain-Level Benchmarks
Md Asif Bin Syed (Georgia Institute of Technology), Azmine Toushik Wasi (Shahjalal University of Science and Technology)
TransformerSupervised Fine-TuningMixture of ExpertsContrastive LearningTabularTime SeriesBenchmarkFinance Related
🎯 What it does: Researchers proposed and validated that time series foundation models (TSFMs) require explicit domain-level benchmarks, and conducted a unified evaluation of seven mainstream TSFMs across 72 datasets spanning six domains (healthcare, finance, energy, nature, transportation, retail), revealing the phenomenon of inconsistent performance across domains.
Position: To Defend Against Cyber Attacks, We Must Teach AI Agents to Hack
Terry Yue Zhuo (Monash University), Ruijie Meng (National University Of Singapore)
Autonomous DrivingFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelReinforcement LearningAgentic AITextTabularTime SeriesSequentialBenchmark
🎯 What it does: Studied and advocated for training AI agents to conduct attacks in controlled environments to enhance defensive capabilities.
Position: Token Taxes Can Mitigate AI's Economic Risks
Lucas Irwin (University of Oxford), Fazl Barez (University of Oxford)
OptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringAuto EncoderGenerative Adversarial NetworkTextTabularFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes and argues for a tax mechanism based on AI usage end Tokens, aiming to alleviate economic risks brought by AI, and provides a three-stage review pipeline for technical governance and a research roadmap.
Position: Topological Machine Learning Cannot Progress without Experimental Standards
Inés Castilla Rieso (University of Poitiers), Philippe Carre (University of Poitiers)
ClassificationComputational EfficiencyData-Centric LearningGraph Neural NetworkSupervised Fine-TuningContrastive LearningImagePoint CloudMeshGraphBenchmark
🎯 What it does: This paper systematically analyzes the impact of experimental design in Topological Machine Learning (TML) on performance and computational cost, and proposes a unified experimental standard to improve the reproducibility and comparability of TML research.
Position: Towards Responsible Evaluation for Text-to-Speech
Yifan Yang (Shanghai Jiao Tong University), Xie Chen (Shanghai Jiao Tong University)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationAudio
🎯 What it does: Propose the concept of Responsible Evaluation, dividing TTS evaluation into three levels: authenticity, comparability, and governance, emphasizing the integration of technology and ethics.
Position: Trustworthy AI Suffers from Invariance Conflicts and Causality is The Solution
Ruta Binkyte (CISPA Helmholtz Center for Information Security), Mario Fritz (CISPA Helmholtz Center for Information Security)
Safty and PrivacyExplainability and InterpretabilityScore-based ModelContrastive LearningTabularBiomedical DataElectronic Health Records
🎯 What it does: This paper redefines the four dimensions of trustworthy AI (fairness, privacy, robustness, and explainability) as invariance requirements under different interventions, and proposes a causal perspective framework for trustworthy AI by using causal models to explain and mitigate these conflicts.
Position: Trustworthy Model Context Protocol Enables Responsible Agentic AI!
Arjhun Swaminathan (University of Tübingen), Anika Hannemann (Swiss Centre for Responsible AI)
Federated LearningSafty and PrivacyExplainability and InterpretabilityAgentic AIReview/Survey Paper
🎯 What it does: Conduct a trustworthiness analysis of the Model Context Protocol (MCP), propose a governance framework that aligns the EU Trustworthy AI guidelines with MCP, and present five directions for improvement (risk adaptability, metadata standardization, host-layer execution, registry governance, and sustainable and fair mechanisms).
Position: Uncertainty is a Strategic Signal in Human–AI Decision Making
Achref Doula (Technical University of Darmstadt), Max Mühlhäuser
Recommendation SystemExplainability and InterpretabilityReinforcement Learning from Human FeedbackPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and experimentally validated a mechanism that treats AI uncertainty as a strategic signal, improving human inspection rates and decision accuracy by introducing feedback and incentives in interactions.
Position: Uncertainty Quantification in LLMs is Just Unsupervised Clustering
Tiejin Chen (Arizona State University), Hua Wei (Arizona State University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelContrastive LearningTextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper systematically analyzes existing uncertainty quantification (UQ) methods for large language models (LLMs), pointing out that they are essentially equivalent to unsupervised clustering, revealing three key shortcomings, and proposing an improvement roadmap from the perspectives of evaluation, mechanism, and benchmarks.
Position: Universal Aesthetic Alignment Narrows Artistic Expression
Wenqi Marshall Guo (University of British Columbia), Shan Du (University of British Columbia)
GenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Construct a broad-spectrum aesthetic dataset to evaluate the compliance and bias of multiple image generation and reward models when executing 'anti-aesthetic' instructions.
Position: Unlabeled ≠ No Human Supervision in Visual Learning
Dong Lao (Louisiana State University)
Explainability and InterpretabilityData-Centric LearningSupervised Fine-TuningContrastive LearningTextReview/Survey Paper
🎯 What it does: The paper analyzes the confusion between 'unlabeled' and human supervision in visual learning, and proposes the necessity of explicitly disclosing human priors in method descriptions.
Position: Unplugging a Seemingly Sentient Machine Is the Rational Choice — A Metaphysical Perspective
Erik J Bekkers, Anna Ciaunica (University of Lisbon)
Review/Survey Paper
🎯 What it does: This paper analyzes the ethical paradox of 'pulling out a seemingly conscious machine' from a metaphysical perspective, and proposes a biophysical idealism framework, arguing that AI is functional imitation rather than possessing subjectivity, thereby supporting the rationality of pulling out AI.
Position: Use Sparse Autoencoders to Discover Unknowns
Kenny Peng (Cornell University), Nikhil Garg (Cornell University)
Explainability and InterpretabilityRepresentation LearningLarge Language ModelAuto EncoderText
🎯 What it does: This paper explores and demonstrates the advantages of sparse autoencoders (SAE) in discovering unknown concepts, distinguishing between known concept tasks and unknown concept tasks, and discusses their potential applications in interpretability, fairness, and social sciences.
Position: Vector Prompt Interfaces Should Be Exposed to Enable Customization of Large Language Models
Liangwei Yang (Salesforce AI Research), Shelby Heinecke (Salesforce AI Research)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The article proposes that model providers should publicly open vector prompt interfaces to support scalable and inference-only customization of large language models in black-box inference environments.
Position: Verifiable Data Minimization is a Prerequisite for Responsible, Privacy-Preserving Industrial Vision
Sander De Coninck (Ghent University), Pieter Simoens (Ghent University)
Pose EstimationSafty and PrivacyAuto EncoderContrastive LearningImage
🎯 What it does: Propose a verifiable data minimization approach in industrial visual systems based on the information bottleneck (IB) principle, and deploy privacy filters on trusted execution environment (TEE) edge devices.
Position: Video LLMs Must Not Ignore the Pixel Dynamics in Plain Sight
Shayda Moezzi (Northeastern University), Sarah Ostadabbas (Northeastern University)
RecognitionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodalityBenchmark
🎯 What it does: This paper systematically diagnoses existing Video LLM frameworks and identifies two core failure modes: Static-Cue Dominance and Prior-driven Temporal Hallucination. It calls for strengthening spatiotemporal information modeling and validation at the pixel level in model and benchmark design.
Position: Virtual Cells Need Context, Not Just Scale
Payam Dibaeinia (Biohub), Aly A Khan
Explainability and InterpretabilityRepresentation LearningData-Centric LearningDrug DiscoveryTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningBiomedical DataBenchmark
🎯 What it does: This paper explores the core bottleneck faced by virtual cell models when predicting cellular responses to perturbations. It argues that merely increasing data scale or model capacity is insufficient for achieving cross-context generalization, and instead emphasizes the need to cover richer biological contexts. The authors use causal transport theory to explain how context dependency limits mechanism transfer, and conduct experiments on a 22M cell CD4⁺T cell Perturb-seq dataset to evaluate the performance of different models on cross-context tasks.
Position: Vision Encoders should be Image Size Agnostic and Task Driven
Nedyalko Prisadnikov (INSAIT Sofia University St Kliment Ohridski), Luc Van Gool (INSAIT Sofia University St Kliment Ohridski)
ClassificationTransformerReinforcement LearningContrastive LearningImage
🎯 What it does: Propose an image encoder design that can adapt to any image size and dynamically adjust attention based on the task; implement and verify the feasibility of the SOVA (Sequential Observation with Visual Attention) model.
Position: Want Better ML Reviews? Stop Asking Nicely and Start Incentivizing with a Credit System
Shaochen Zhong (Rice University)
Review/Survey Paper
🎯 What it does: Proposed a credit-based cross-conference scoring and incentive system (OpenReview Points), combined with fine-grained procedural safeguards, aiming to improve the quality and efficiency of peer review in the field of machine learning.
Position: We Need A Unified Definition of Hallucination (It’s The World Model, Stupid!)
Emmy Liu (DegenAI Labs), Steven Y. Feng (DegenAI Labs)
Explainability and InterpretabilityData-Centric LearningPrompt EngineeringWorld ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a unified framework for defining hallucinations, attributing hallucinations to inaccurate world modeling, and determining whether model outputs are hallucinations based on three elements: reference world model (W), view function (V), and conflict strategy (P).
Position: We Need AI Efficiency Incentives for Accessibility and Sustainability
Marco Bornstein (Independent Researcher), Amrit Singh Bedi (University of Central Florida)
Computational Efficiency
🎯 What it does: Propose a market-based AI cap-and-trade mechanism that encourages model inference efficiency and reduces FLOP consumption.
Position: We Need Large Language Models Optimized For Our Well-Being
Ashton Anderson (University of Toronto), Karina Vold (University of Toronto)
OptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: In LLM daily life and emotional support scenarios, a specialized, optional long-term health-oriented mode is proposed, rather than merely optimizing immediate user satisfaction.
Position: We Need Practical AI Alignment Methods that Mirror Human Reasoning
Vijay Keswani (IIT Delhi), Jana Schaich Borg (Duke University)
Recommendation SystemAutonomous DrivingExplainability and InterpretabilityPrompt EngineeringTextTabularReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the concept of 'cognitive alignment' AI, and through empirical questionnaires and literature reviews, it elaborates on users' demand for similarity between AI's mode of thinking and their own in high-risk decision-making scenarios.
Position: We need to re-think the concept of “real” images.
Janis Keuper (IMLA, Offenburg University), Margret Keuper (University of Mannheim)
Anomaly DetectionConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper discusses the shortcomings of the concept of 'real images' in deepfake detection from both technical and philosophical perspectives, and points out that existing datasets and evaluations lack consideration of modern smartphone photography technologies (such as multi-camera fusion and deep learning enhancement), leading to a decline in detection model performance in real-world scenarios.
Position: Web Agents Should Use Typed Actions Instead of Click-Based Browsing
Linxi Jiang (Ohio State University), Suman Nath (Microsoft)
AI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelAgentic AIPrompt EngineeringTextMultimodality
🎯 What it does: Designed and implemented the Web Verbs semantic layer, providing a Web operation interface with typed input/output, allowing agents to directly invoke and compose these predicates to complete tasks.
Position: Weight Space Should Be a First-Class Generative AI Modality
Zhangyang Wang (University of Texas at Austin), Kai Wang (Tencent Hy)
GenerationData SynthesisOptimizationMeta LearningGraph Neural NetworkTransformerPrompt EngineeringDiffusion modelScore-based ModelGenerative Adversarial NetworkGraphTabularSequential
🎯 What it does: This paper proposes to treat neural network checkpoints as a first-class data modality and argues that generative modeling in the weight space should be standardized as a core machine learning primitive.
Position: What Cézanne Knew About Visual Intelligence That Vision-Language Models Miss
Mohammad Rashedul Hasan (University of Nebraska-Lincoln), Chinh Hoang (University of Nebraska-Lincoln)
Explainability and InterpretabilityRepresentation LearningLarge Language ModelPrompt EngineeringVision Language ModelImageTextChain-of-Thought
🎯 What it does: This paper proposes and verifies two blind spots in causal reasoning evaluation for vision-language models: missing temporal composition and missing scaffold invariance, and demonstrates this through experiments on three LLaVA models.
Position: When AI Decides Who Gets an Organ: Multi-Agentic AI Systems in Transplant Medicine Risk Amplifying Disparities Without Targeted Explainability and Deployment Strategies
Divya Sharma (York University), Mamatha Bhat (University Health Network)
Federated LearningSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringTabularBiomedical DataElectronic Health Records
🎯 What it does: This paper constructs and evaluates a multi-agent AI system in the decision-making process of a liver transplant committee, revealing its systemic biases related to gender, race, and socioeconomic status.
Position: Why Current Fair-AI Fails Spatial Fairness, And How to Adapt to It
Nripsuta Ani Saxena (University of Southern California), Cyrus Shahabi (University of Southern California)
Recommendation SystemFederated LearningSafty and PrivacyExplainability and InterpretabilityTextTabularReview/Survey PaperAgriculture RelatedFinance Related
🎯 What it does: The paper points out the shortcomings of current fair AI in terms of spatial fairness (location fairness), elaborates on the social, legal, and technical roots of spatial bias, and proposes targeted research and practical guidelines;
Position: World Models as an Intermediary between Agents and the Real World
Sherry Yang (New York University)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelAuto EncoderGenerative Adversarial NetworkWorld ModelVideoTextTabularTime SeriesSequential
🎯 What it does: Propose using a world model as an intermediary to interact with the real world, addressing the training bottleneck in reinforcement learning under high-cost environments.
Position: Your VLM May Not Be Thinking with Interleaved Images
Wenjie Yang (Fudan University), Zengfeng Huang (Fudan University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper systematically evaluates whether the interleaved interactive images in the 'Thinking with Images' paradigm are truly utilized by the model, mainly verifying the impact of images on performance through ablation experiments, attention visualization, and occlusion experiments.
Position: Zeroth-Order Optimization in Deep Learning Is Underexplored, Not Underpowered
Sijia Liu (Michigan State University), Yihua Zhang (Michigan State University)
OptimizationSupervised Fine-TuningDiffusion modelScore-based ModelText
🎯 What it does: This paper systematically evaluates and discusses the potential and limitations of zeroth-order optimization (ZO) in deep learning, pointing out that its capabilities are underestimated rather than lacking, and proposes six improvement directions.
Positional Encoding for Spiking Transformers
Zijian Zhou (University Of Electronic Science And Technology Of China), Haizhou Li (Shenzhen Loop Area Institute)
ClassificationSpiking Neural NetworkTransformerReinforcement LearningContrastive LearningText
🎯 What it does: Proposed a relative position encoding called SPE for Spiking Transformer, which utilizes PE-LIF neurons with position-related thresholds to encode relative position information while maintaining linear attention.
Positive Distribution Shift as a Framework for Understanding Tractable Learning
Marko Medvedev (University of Chicago), Nathan Srebro (Toyota Technological Institute at Chicago)
OptimizationComputational EfficiencyRepresentation LearningData-Centric Learning
🎯 What it does: Study the Positive Distribution Shift (PDS) framework, exploring how to make computationally hard learning problems solvable under gradient descent by choosing an appropriate training distribution;
Positive-Unlabeled Learning with Extreme Scarcity of Labeled Positives
Yuanchao Dai (Jilin University), Masashi Sugiyama (RIKEN Center for Advanced Intelligence Project)
ClassificationAnomaly DetectionData-Centric LearningContrastive LearningImageBiomedical DataAlzheimer's Disease
🎯 What it does: This study addresses the PU learning problem with extremely scarce positive samples, proposing the ScalePU framework and theoretically providing a sufficient sample threshold.
Positive–Unlabeled Reinforcement Learning Distillation for On-Premise Small Models
Zhiqiang Kou (Southeast University), Tat-Seng Chua (National University of Singapore)
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringContrastive LearningTextMultimodality
🎯 What it does: Proposes a positive and unlabeled (PU) reinforcement learning distillation method, which only requires one teacher query per prompt. By guiding the local small model to self-evaluate and generate soft preference distributions through anchor responses, alignment is achieved without using human labels or reward models.
Possibilistic Predictive Uncertainty for Deep Learning
Yao Ni (Nanyang Technological University), Piotr Koniusz (University Of New South Wales)
ClassificationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerContrastive LearningImageTextStochastic Differential Equation
🎯 What it does: Propose a deep learning framework for predicting uncertainty based on possibility theory, DAPPr, which uses the Dirichlet possibility function to approximate the possible predictive posterior;
Post-Hoc Merging is Not Enough: Many-Shot Model Merging with Loss-Gap Balancing
Kyungjin Im (Soongsil University), Minhae Kwon (Sungkyunkwan University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextBenchmark
🎯 What it does: A model fusion framework based on multiple iterative fusion is studied and proposed, and under this framework, a loss gap weighting mechanism and a consistency mask mechanism are designed, significantly alleviating the information erasure problem caused by task interference.
Post-Training Language Models for Crosslingual Consistency
Tianyu Liu (ETH Zürich), Arianna Bisazza (University of Groningen)
OptimizationFederated LearningComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextMultimodalityBenchmark
🎯 What it does: Proposes a post-training method called DCO (Direct Consistency Optimization) to improve the response consistency of multilingual models across different languages, and proves its equivalence to the more challenging PCO (Penalized Consistency Optimization).
Post-Training LLMs as Better Decision-Making Agents: A Regret-Minimization Approach
Chanwoo Park (Massachusetts Institute of Technology), Kaiqing Zhang (University of Maryland)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Post-training large language models, proposing Iterative Regret Minimization Fine-tuning (ITERATIVE RMFT) to enhance their online decision-making capabilities.
Post-Training with Policy Gradients: Optimality and the Base Model Barrier
Alireza Mousavi-Hosseini (University of Toronto), Murat A Erdogdu (University of Toronto)
OptimizationTransformerLarge Language ModelReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: This paper provides a theoretical analysis of policy gradient (PG) methods used in the post-training phase of linear autoregressive models, proving that optimal convergence can be achieved within the support of the base model, while exponential reward queries are required outside the support; meanwhile, it proposes a method using process reward (PR) to eliminate the curse of dimensionality.
PosterAgent: Agentic Poster Generation via Stage-Aware Reinforcement Learning
Zhuocheng Yu (Peking University), Kai Jia (ByteDance)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Constructed PosterAgent, a multi-round interactive image generation framework, which first generates a structured JSON layout using a multimodal large model, then generates a visual poster through a renderer, and subsequently allows the model to self-criticize and iteratively refine until the user's needs are met.
Posterior Behavioral Cloning: Pretraining BC Policies for Efficient RL Finetuning
Andrew Wagenmaker (University of California Berkeley), Sergey Levine (University of California Berkeley)
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningDiffusion modelImageVideoSequential
🎯 What it does: This paper proposes a posterior behavior cloning (POSTBC) pre-training method that learns the posterior distribution of the demonstrator's behavior using demonstration data, thereby achieving a more comprehensive pre-training strategy and further improving the fine-tuning efficiency in reinforcement learning.
Posterior Concentration of Bayesian Physics-Informed Neural Networks for Elliptic PDEs
Yuxuan Zhao (University of Minnesota), Yulong Lu (University of Minnesota)
Score-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingPhysics RelatedStochastic Differential Equation
🎯 What it does: Study the posterior convergence rate of Bayesian physics-informed neural networks (PINNs) for elliptic partial differential equations (PDEs), proving that the posterior can converge to the true solution at nearly optimal rates under both interior and boundary observation noise, and is adaptive to the smoothness of the solution.
Posterior Mismatch Matters: Adversarial Training for Long-Tailed Robustness
Lilin Zhang (Sichuan University), Xianggen Liu (Sichuan University)
ClassificationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Study the robustness of adversarial training under long-tailed distributions and propose a posterior-driven adversarial training (PAT) method.
Posterior Sampling Reinforcement Learning with Gaussian Processes for Continuous Control: Sublinear Regret Bounds for Unbounded State Spaces
Hamish Flynn (Carnegie Mellon University), Jan Peters (Technical University of Darmstadt)
Reinforcement LearningTabularTime Series
🎯 What it does: Proposed and analyzed a reinforcement learning algorithm based on Gaussian process posterior sampling (GP-PSRL), proving that sublinear Bayesian regret upper bounds can be achieved under continuous control, unrestricted state spaces, and weak smoothness conditions.
PostTrainBench: Can LLM Agents Automate LLM Post-Training?
Ben Rank (Ellis Institute Tubingen), Maksym Andriushchenko (Ellis Institute Tubingen)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes POSTTRAINBENCH, a benchmark for evaluating the ability of large language model (LLM) agents to automatically complete post-training of LLMs under limited computational resources (10 hours on a single H100 GPU);
Power-Boosted Granger-Causal Discovery for Large Heterogeneous Panel Data
Yiheng Gu (University of Notre Dame), Xiufan Yu (University of Notre Dame)
TabularTime SeriesFinance Related
🎯 What it does: This paper proposes a Power-Enhanced Panel Granger Causality Test (PE-PGCT) for large-scale heterogeneous panel data, which enhances the power of the test under sparse signals by adding a power-enhancing component based on the maximum Wald statistic to existing panel Granger causality tests (such as DH or HPJ).
PowerFlow: Unlocking the Dual Nature of LLMs via Principled Distribution Matching
Ruishuo Chen (Tsinghua University), Longbo Huang (Tsinghua University)
GenerationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningFlow-based ModelText
🎯 What it does: Propose the PowerFlow framework, achieving unsupervised fine-tuning through distributed matching and GFlowNet, actively regulating the reasoning and creativity of large language models
Powerful and Theoretically Guaranteed Independence Testing on Heterogeneous Federated Clients
Yixin Ren (Fudan University), Shuigeng Zhou (Fudan University)
Federated LearningSafty and PrivacyComputational EfficiencyContrastive LearningTabularBiomedical Data
🎯 What it does: In the federated learning scenario, a new independence test method called FedIT-CS is designed, which can perform effective independence tests even when the data distributions across clients are heterogeneous.