In this workshop, we aim to address the critically under-discussed issue of AI's dual-use nature, focusing on how machine learning technologies are being adapted for military purposes, potentially without the researchers' knowledge or consent. While attending to the heightened risks associated with particular areas and systems of research, we will also be collectively thinking through what it looks like to engage productively in research and development activities that considers ethics and international law at its core.

Recursive self-improvement (RSI) is moving from thought experiments to deployed AI systems. LLM agents now rewrite their own codebases or prompts, scientific discovery pipelines schedule continual fine-tuning, and robotics stacks patch controllers from streaming telemetry, even improving product-level code. The ICLR 2026 Workshop on AI with Recursive Self-Improvement brings together researchers to discuss a simple question with big consequences: how do we build the algorithmic foundations for powerful and reliable self-improving AI systems? As loops that update weights, rewrite prompts, or adapt controllers move from labs into production, we will surface the methods that work — how to design, evaluate, and govern these loops without hand-waving. This workshop examines algorithms for self-improvement across experience learning, synthetic data pipelines, multimodal agentic systems, weak-to-strong generalization, and inference-time scaling, and will discuss and refine methods for recursive self-improvement. In short, we care about loops that actually get better — and can show it. To give the workshop a clear spine, we organize contributions around five lenses: change targets inside the system, temporal regime of adaptation, mechanisms and drivers, operating contexts, and evidence of improvement. This framing synthesizes recent perspectives on self-evolving agents while grounding them in practical, auditable deployment settings. We are paradigm-agnostic: we welcome work on foundation models, agent frameworks, robots, learning algorithms and optimizers, control and program synthesis, as well as data and infrastructure systems and evaluation tooling that enable recursive self-improvement.

Artificial intelligence has reached a pivotal stage: while current agentic systems excel in static benchmarks, they struggle to adapt to dynamic, real-world environments. This workshop introduces the concept of lifelong agents, AI systems that continuously learn, align, and evolve across their operational lifespan. Such agents must integrate continual learning, long-term alignment with human values, and self-improvement under resource constraints to remain robust, trustworthy, and sustainable. By uniting research from reinforcement learning, large language models, alignment, embodied AI and more, the workshop seeks to establish shared principles, frameworks, and evaluation methods for creating AI that grows intelligently and responsibly over time.

AI agents are rapidly being deployed in critical real-world applications, yet their unique safety and security challenges remain underexplored. Unlike standard safety or security settings, agents act autonomously and make irreversible real-world decisions. This creates novel vulnerabilities and fundamental safety challenges for agents in real-world deployments. Our workshop provides the first dedicated venue for addressing the safety, security, and trustworthiness of agents in the wild. We bring together interdisciplinary researchers and practitioners to establish foundational theories and methods for safe agent deployment, identify critical open problems, and chart research directions for trustworthy agentic AI systems.

Artificial General Intelligence (AGI) has long seemed distant, but rapid advances in large-scale learning, autonomous reasoning, and open-ended discovery make its emergence increasingly plausible. The Post-AGI Science and Society Workshop asks what comes next. If AGI becomes ubiquitous, reliable, and affordable, how will it reshape scientific inquiry, the economy of knowledge, and human society? Will humans remain central to discovery or become curators and interpreters of machine-generated insights? The workshop brings together researchers from machine learning, philosophy of science, and policy to explore human-AI scientific coexistence. Topics include automated hypothesis generation, causal reasoning in AGI, collaborative discovery, epistemic alignment between humans and machines, and socio-economic shifts driven by pervasive intelligence. Through keynotes, talks, and a panel, we will examine how science and our understanding of knowledge might evolve in a post-AGI world.

Foundation models have transformed multimodal intelligence, enabling open-ended reasoning, dialogue, and generation across vision, language, and audio. A growing body of work now frames this progress under the unifying paradigm of next-X prediction, where X may denote tokens, frames, or scales across discrete or continuous spaces. Discrete autoregressive models, such as Chameleon, extend next-token prediction beyond text, while continuous formulations like VAR, MAR, TransFusion, BAGEL, and Fluid capture next-frame or next-scale dynamics in latent space. Meanwhile, predictive encoders—exemplified by V-JEPA 2—eschew token emission to forecast future representations, focusing on salient, structured aspects of perception and behavior. Complementary to both, discrete diffusion models such as Diffusion-LM, LLaDA, and LaViDa redefine generation as iterative denoising, offering parallelism and improved global consistency. This workshop provides a timely venue to connect these emerging paradigms—next-token generation, predictive encoding, and diffusion-based modeling—and to explore how they can be integrated into unified multimodal systems. Key questions include: Which learning paradigm scales most effectively? How do they differ in representation quality, efficiency, and controllability? And can hybrid models combine their strengths? By bringing together researchers from these diverse communities, the workshop aims to chart a coherent roadmap for the next generation of multimodal foundation models—beyond token prediction alone.

Large language models (LLMs) have achieved remarkable breakthroughs in natural language understanding and generation, but their logical reasoning capabilities remain a significant bottleneck. Logical reasoning is crucial for tasks requiring precise deduction, induction, or abduction, such as medical diagnosis, legal reasoning, and scientific hypothesis verification. However, LLMs often fail to handle complex logical problems with multiple premises and constraints, and they frequently produce self-contradictory responses across different questions. These limitations not only restrict the reliability of LLMs in complex problem-solving but also hinder their real-world applications. In response to these emerging needs, we propose the workshop on Logical Reasoning of LLMs. This workshop will explore the challenges and opportunities for improving deduction, induction, and abduction capabilities of LLMs, implementing symbolic representation and reasoning via LLMs, avoiding logical contradictions across responses to multiple related questions, enhancing LLM reasoning by leveraging external logical solvers, and benchmarking LLM logical reasoning and consistencies. As LLMs continue to expand their role in AI research and applications, this workshop will serve as a platform to discuss and refine the methods for advancing logical reasoning within LLMs.

The workshop will explore how explicit models of human knowledge, cognitive capabilities, and mental states can be integrated into AI reasoning processes. We will examine approaches that combine neural and symbolic methods inspired by human cognition, incorporate human causal reasoning patterns, and leverage human teaching signals to create more interpretable and aligned AI systems. More details at https://bit.ly/hcair26

While deep learning continues to achieve impressive results on an ever-growing range of tasks, our understanding of the principles underlying these successes remains largely limited. This problem is usually tackled from a mathematical point of view, aiming to prove rigorous theorems about optimization or generalization errors of standard algorithms, but so far they have been limited to overly-simplified settings. The main goal of this workshop is to promote a complementary approach that is centered on the use of the scientific method, which forms hypotheses and designs controlled experiments to test them. More specifically, it focuses on empirical analyses of deep networks that can validate or falsify existing theories and assumptions, or answer questions about the success or failure of these models. This approach has been largely underexplored, but has great potential to further our understanding of deep learning and to lead to significant progress in both theory and practice. The secondary goal of this workshop is to build a community of researchers, currently scattered in several subfields, around the common goal of understanding deep learning through a scientific lens.

Summary: This workshop will delve into aspects of time series prediction and analysis in the age of large models. This workshop builds upon our successful track record of fostering community engagement around large models for time series. Our inaugural NeurIPS 2024 workshop demonstrated strong community interest, attracting 99 submissions and over 500 participants (~1000 registered interest via Whova). Submissions spanned the full spectrum of the field—from building time series foundation models and leveraging pre-trained models from other modalities, to real-world applications and deployment experiences. The rich discussions at NeurIPS 2024 revealed both significant opportunities and fundamental limitations in current approaches, directly informing the research questions we aim to address in this iteration. Building on this momentum, we also organized the successful ICML 2025 Workshop on Foundation Models for Structured Data, which broadened our perspective by connecting time series researchers with the tabular data community. Focus and Innovation: For ICLR 2026, we are strategically refocusing to dive deeper into outstanding research questions that emerged from our previous workshops - particularly around agents, interpretability, and context-informed predictions. This iteration features an evolved organizing team and fresh speaker lineup, reflecting the field's rapid development. The nascent nature of large time series models makes this workshop particularly timely for ICLR 2026, as the community continues to establish foundational principles and explore novel applications in this emerging domain. Organizer Expertise: The organizers bring extensive research experience and proven leadership in the time series foundation models domain, with diverse backgrounds from industry and academia. Collectively, we have led advances on 3 key dimensions: foundational model development– creating some of the first time series foundation models including Lag-Llama, Chronos, Moment, Moirai, and TimesFM, advanced applications– establishing initial frameworks for reasoning and agents in time series through MLZero and TimeSeriesGym, and rigorous evaluation and benchmarking using tools such as Context-is-Key, GIFT-Eval, TimeSeriesExam, and fev-bench. Beyond research contributions, our team has demonstrated success in organizing impactful workshops at premier venues, including the NeurIPS 2024 workshop on Time Series in the Age of Large Models, AAAI’24 Spring Symposium on Clinical Foundation Models, ICAIF’24 Foundation Models for Time Series: Exploring New Frontiers, and ICML’25 Workshop on Foundation Models for Structured Data. This combination of deep technical expertise and proven workshop leadership positions us to facilitate meaningful discussions and foster collaboration in this rapidly evolving field.

AI4Mat-ICLR-2026 explores the automated discovery of advanced materials through three interconnected pillars: 1. AI-Guided Design; 2. Automated Synthesis; 3. Automated Characterization. By bringing together leading researchers at the intersection of machine learning and materials science, the workshop fosters discussion of cutting-edge advances while building a cohesive, multidisciplinary community tackling some of the field's most pressing challenges. To that end AI4Mat-ICLR-2026's program highlights two leading topics to foster scientific dialogue in relevant subject areas, each featuring carefully curated invited speakers: 1. Reinforcement Learning & Beyond: The Role of Feedback in AI for Materials Science; 2. Cross-Modal, Unified Materials Representations – From Structure to Properties to Performance. In addition to invited talks and technical discussions, AI4Mat-ICLR-2026 continues its commitment to community development through established initiatives, including a Tiny Papers track for early-stage work, travel grants to support broad and inclusive researcher participation, and a dedicated journal venue for high-quality submissions.

Partial Differential Equations (PDEs) are foundational to modeling complex phenomena across the natural sciences and engineering, from fluid dynamics and quantum systems to climate modeling and materials science. Despite their ubiquity, solving PDEs remains computationally intensive, especially in high-dimensional, multi-physics, and uncertain regimes. Recent advances in machine learning—such as neural operators, physics-informed networks, and foundation models—offer transformative potential to accelerate and generalize PDE solutions. However, realizing this promise requires addressing critical challenges in representation, stability, generalization, and benchmarking. The AI\&PDE-ICLR-2026 workshop will convene researchers from machine learning, applied mathematics, physics, and engineering to explore the future of AI-driven PDE modeling. We aim to (1) define the roadmap toward foundation models for PDEs, (2) investigate next-generation representations and architectures, and (3) foster a globally inclusive community. The program will feature invited talks, contributed papers, and themed tracks, including a full papers track for mature research and a tiny papers track for emerging ideas. By bridging disciplines and promoting open benchmarks and datasets, AI&PDE-ICLR-2026 will catalyze progress toward scalable, general-purpose AI solvers for PDEs.

Real-world data often originates from physical systems that are governed by geometric and physical laws. Yet, most machine learning methods treat this data as abstract vectors, ignoring the underlying structure that could improve both performance and interpretability. Geometry provides powerful guiding principles, from group equivariance to non-Euclidean metrics, that can preserve the symmetries or the structure inherent in data. We believe those geometric tools are well-suited, and perhaps essential, for representation learning and generative modeling. We propose GRaM, a workshop centered on the principle of grounding in geometry, which we define as: An approach is geometrically grounded if it respects the geometric structure of the problem domain and supports geometric reasoning. This year, we aim to explore the relevance of geometric methods, particularly in the context of large models, focusing on the theme of scale and simplicity. We seek to understand when geometric grounding remains necessary, how to effectively scale geometric approaches, and when geometric constraints can be relaxed in favor of simpler alternatives.

AI has transitioned from predictive models to interactive, autonomous agents capable of reasoning, planning, and executing complex goals. As the systems increasingly influence social, economic, and scientific decisions, they determine whose interests are represented and whose opportunities are constrained. Ensuring fairness, therefore, is no longer an ethical preference but a practical imperative. As the fairness challenges are fundamentally transformed by advanced AI systems, traditional algorithmic fairness frameworks developed primarily for prediction and/or prediction-based decision-making no longer suffice. This workshop, Algorithmic Fairness Across Alignment Procedures and Agentic Systems (AFAA), emerges at this pivotal moment as a timely forum for rethinking fairness in AI alignment processes and agentic system development. By examining fairness across alignment procedures and agentic systems, this workshop creates a crucial platform for bridging the gap between rapid technical advances in model capabilities and the equally important advances needed in frameworks of algorithmic fairness to govern these powerful systems.

Several areas at the forefront of AI research are currently witnessing a convergence of interests around the problem of learning high-quality concepts from data. Concepts have become a central topic of study in neuro-symbolic integration (NeSy). NeSy approaches integrate perception – usually implemented by a neural backbone – and symbolic reasoning by employing concepts to glue together these two steps: the latter relies on the concepts detected by the former to produce suitable outputs [1–5]. Concepts are also used in Explainable AI (XAI) by recent post-hoc explainers [6–9] and self-explainable architectures [10–13] as a building block for constructing high-level justifications of model behavior. Compared to, e.g., saliency maps, these can portray a more abstract and understandable picture of the machine’s reasoning process, potentially improving understandability, interactivity, and trustworthiness [14–17], to the point that concepts have been called the lingua franca of human-AI interaction [18]. Both areas hinge on learned concepts being “high-quality”. Causal Representation Learning (CRL) aims to identify latent causal variables and causal relations from high-dimensional observations, e.g., images or text, with theoretical guarantees [25]. As such, CRL is a generalization of disentangled representation learning, when the latent variables are dependent on each other, e.g., due to causal relations. CRL has been increasingly popular, with a plethora of methods and theoretical results [26–36]. The potential of leveraging CRL to learn more robust and leak-proof concept is an emerging area of research with a growing number of approaches [24, 37–40], but many open questions remain. In particular, what properties high-quality concepts should satisfy is unclear, and – despite studying the same underlying object – research in these areas is proceeding on mostly independent tracks, with minimal knowledge transfer. Efforts at adapting ideas and techniques are limited at best, meaning that approaches in one area completely ignore insights from the others. As a result, the central issue of how to properly learn and evaluate concepts is largely unanswered. This workshop brings together researchers from NeSy, XAI and CRL and from both industry and academia, who are interested in learning robust, semantically meaningful concepts. By facilitating informal discussion between experts and newcomers alike, it aims to tie together these currently independent strands of research and promote cross-fertilization.

The primary focus of this workshop is to strengthen the analytical foundations of associative memory while exploring its emerging role in the design of agentic AI systems. By bringing together researchers from optimization and deep learning, statistical physics, neuroscience and machine learning systems, the workshop aims to catalyze cross-disciplinary exchange, identify open problems, and foster collaboration toward advancing the theoretical and practical frontiers of associative memory. A central goal is to build a cohesive community at the intersection of these fields, one that unites rigorous mathematical foundations with scalable architectures and applications where associative memories serve as the core drivers of reasoning, adaptation, and intelligent behavior.

This workshop series explores the intersection of scale-driven generative artificial intelligence (AI) and the correctness-focused principles of verification. In its first rendition at ICLR 2025, it focused in particular on how generative AI can address the scaling challenges faced by formal analysis tools such as theorem provers, satisfiability solvers, and execution monitoring. The special theme of VerifAI@ICLR'25 was thus Large Language Models (LLMs) for Code Generation, an undeniably active area of research across both industry and academia, which has benefited greatly from (and improved) formal analysis tools such as static analyzers. Now, in light of the recent emphasis on large-scale post-training through reinforcement learning (RL), we are excited to continue uniting the interests of industry and academia with a new special theme: Building verifiable tasks and environments for RL.

The past year has witnessed remarkable advances in foundation models (FMs): new post-training paradigms such as reinforcement learning with verifiable rewards (RLVR) that strengthen reasoning, increasingly multimodal and agentic systems, and renewed attention to benchmark design and evaluation. Each of these advances depends on distinct data innovations: verifiable reward signals and reasoning traces for RLVR; aligned cross-modal corpora and interaction logs for multimodality and agency; and leak-resistant, representative test sets for evaluation. Taken together, these dependencies underscore the continuing centrality of data as a design variable at the forefront of FM research. Meanwhile, longstanding challenges in data collection, curation, and synthesis remain unresolved, while concerns surrounding copyright, privacy, and fairness have only intensified. Building on the success of the first two DATA-FM workshops at ICLR 2024 and 2025, the third edition will revisit these persistent issues while highlighting emerging ones at the frontiers of post-training, multimodality, and evaluation. By convening researchers and practitioners across diverse research communities, DATA-FM seeks to advance understanding of data’s evolving role in FMs and foster innovative solutions shaping the next generation of models and applications.

Machine learning systems are increasingly deployed in high-stakes domains such as healthcare, finance, robotics, and autonomous systems, where data distributions evolve continuously. Without robust monitoring and timely adaptation, even high-performing models can degrade silently, compromising reliability, safety, and fairness. Continuous monitoring is therefore an absolute necessity. While there has been rapid progress in drift detection, test-time and continual adaptation, and the deployment of ML systems at scale, these topics are often studied separately. The Catch, Adapt, and Operate workshop brings them together around three themes: sensing drift through statistical and representation-based monitoring, responding through adaptive and self-supervised updates, and operating at scale in production pipelines. By connecting theory, systems, and real-world practice, the workshop aims to build a shared foundation for reliable, fair, and continuously adaptive machine learning under real-world drift.

The rapid advancement of artificial intelligence, particularly in machine learning and foundation models, is creating a new synergy with the classical fields of Mechanism Design (MD) and Strategic Decision Making (SDM). This workshop aims to catalyze interdisciplinary research at this intersection, exploring how modern AI methods can redefine, extend, and automate core problems in MD and SDM. We will bring together researchers from machine learning, economics, and computer science to investigate this symbiotic relationship, focusing on topics such as novel AI applications for MD \& SDM and theoretical models for AI-driven methods. This workshop will highlight not only cutting-edge research but also impactful real-world applications and case studies from industry. Overall, our goal is to provide a premier platform for disseminating novel ideas, fostering collaboration across communities, and charting the future of intelligent economic systems.