AI models for science have the potential to harness large datasets, accelerate scientific discoveries, and transform numerous fields. Through this workshop, our mission is to foster interdisciplinary collaboration to develop fully autonomous AI systems, addressing challenges like benchmark datasets, human-AI collaboration, robust tools and methods for validating AI outputs, and trustworthiness. By tackling these issues, we can unlock AI's transformative potential in research. In this workshop, themed Agentic AI for Science, we will explore these critical topics and welcome diverse perspectives. We will focus on integrating agentic AI systems to enhance scientific discovery while upholding rigorous standards. For AI to contribute effectively, it must generate novel hypotheses, comprehend their applications, quantify testing resources, and validate feasibility through well-designed experiments. This workshop serves as a vital forum for collaboration and knowledge-sharing aimed at redefining the landscape of scientific discovery.

In the era of AI-driven transformations, foundation models (FMs) have become pivotal in various applications, from natural language processing to computer vision. These models, with their immense capabilities,reshape the future of scientific research and the broader human society, but also introduce challenges intheir in-the-wild/real-world deployments. The 2nd Workshop on FMs in the Wild delves into the urgent need forthese models to be useful when deployed in our societies. The significance of this topic cannot be overstated,as the real-world implications of these models impact everything from daily information access to criticaldecision-making in fields like medicine and finance. Stakeholders, from developers to end-users, care deeplyabout this because the successful integration of FMs into in-the-wild frameworks necessitates a careful consideration of many properties, including adaptivity, reliability, efficiency, and reasoning ability. Some of thefundamental questions that this workshop aims to address are:1. In-the-wild Adaptation: How can we leverage techniques such as Retrieval-Augmented Generation(RAG), In-context Learning (ICL), or Fine-tuning (FT) to adapt FMs for specific domains, such asdrug discovery, education, or clinical health?2. Reasoning and Planning: How can FMs be enhanced to tackle more complex in-the-wild tasks thatrequire multi-step reasoning or decision-making, such as multi-hop question answering, mathematicalproblem-solving, theorem proving, code generation, or robot planning scenarios?3. Reliability and Responsibility: How can FMs work reliably outside their training distribution?And how can we address issues like hallucination, fairness, ethics, safety and privacy within the society?4. Practical Limitations in Deployment: How can FMs tackle challenges in practical applications,such as system constraints, memory requirements, response time demands, data acquisition barriers,and computational costs for inference-time scaling and long-context input?In summary, our topics of interest include, but are not limited to:* Innovations in techniques for customizing models to individual user preferences, tasks, or domains* Advancements in the reasoning and planning abilities of FMs in complex real-world challenges* Theoretical and empirical investigations into the reliability and responsibility of various FMs* Strategies for overcoming practical limitations (e.g., memory, time, data) of FMs in broad applications* Methods for integrating multiple modalities (e.g., text, images, action) into a unified in-the-wild framework* Discussions on FM agents that perform intricate tasks through interaction with the environment* In-depth discussions exploring the in-the-wild deployments and applications of FMs* Benchmark methodologies for assessing FMs in real-world settings

Some of the most exciting and impactful open scientific problems have computational complexity as the limiting factor to an in silico solution, e. g. high–temperature superconductivity and fusion power. Atoms behave according to the well–established laws of quantum mechanics, but as system size grows computations quickly become intractable. This workshop will gather for cross–pollination a diverse group of researchers belonging to difference scientific domains and machine learning approaches. The immediate outcome will be an exchange of ideas, datasets, and crystallized problem statements, all towards the ultimate goal of developing universal AI methods that would be able find efficient and accurate approximations of complex systems from low-level theory. If we solve scale transition, we solve science.

Machine learning (ML) models are impressive when they work but they can also show unreliable, untrustworthy, and harmful dangerous behavior. Such behavior is even more common in the era of large models, such as chatGPT, which are quickly being adopted even though we do not understand why they work so well and fail miserably at times. Unfortunately, such rapid dissemination encourages irresponsible use, for example, to spread misinformation or create deep fakes, while hindering the efforts to use them to solve pressing societal problems and advance human knowledge. Ideally, we want models that have a human-like capacity to learn by observing, theorizing, and validating the theories to improve the understanding of the world. At the very least, we want them to aid human knowledge and help us to further enrich it. Our goal in this workshop is to bring together researchers working on understanding model behavior and show how this key aspect can lead to discovering new human knowledge. The workshop will include theoretical topics on understanding model behavior, namely interpretability and explainability (XAI), but also three distinct scientific application areas: weather and climate, healthcare, and material science (ML4Science).

This workshop aims to build a multidisciplinary research community around the emerging field of human-AI coevolution (HAIC) to understand the feedback loops that emerge from continuous and long-term human-AI interaction. As AI systems have become more prevalent and have been present in society over longer periods, scholars from diverse fields and methodologies have come to focus on HAIC and its importance for system architecture, human feedback, regulation, and other domains. Through this workshop we hope to lay a collaborative foundation for this research agenda. To achieve this we will organize expert talks from academia and industry, dynamic panel discussions, interactive breakout sessions, and networking opportunities, drawing on our diverse experience organizing related workshops at leading conferences in ML, NLP, HCI, and related fields.

The increasing complexity of modern machine learning models exposes the limitations of the traditional, monolithic approach to their development, raising concerns about cost and sustainability.This workshop challenges this approach by advocating for a new paradigm based on modular design and functional specialization. Inspired by principles from software engineering, we envision a future where models are composed of independently trainable modules, enabling asynchronous development, incremental updates, and cross-task generalization through composability. This shift towards modularity unlocks new possibilities for collaborative model development where researchers can contribute specialized modules, combine existing models, and participate in decentralized training schemes. By embracing modularity, we can democratize deep learning research, enabling smaller teams and institutions to contribute to the development of powerful and efficient models. Furthermore, modularity promises to enhance model interpretability, and maintainability, paving the way for more robust and efficient AI systems. This workshop aims to accelerate this transition towards a more collaborative and sustainable future for deep learning.

Biomolecular design, through artificial engineering of proteins, ligands, and nucleic acids, holds immense promise in addressing pressing medical, industrial, and environmental challenges. While generative machine learning has shown significant potential in this area, a palpable disconnect exists with experimental biology: many ML research efforts prioritize static benchmark performance, potentially sidelining impactful biological applications. This workshop seeks to bridge this gap by bringing computationalists and experimentalists together, catalyzing a deeper interdisciplinary discourse. Together, we will explore the strengths and challenges of generative ML in biology, experimental integration of generative ML, and biological problems ready for ML. To attract high-quality and diverse research, we partnered with Nature Biotechnology for a special collection, and we created dedicated tracks for in-silico ML research and hybrid ML-experimental biology research. Our lineup features emerging leaders as speakers and renowned scientists as panelists, encapsulating a spectrum from high-throughput experimentation and computational biology to generative ML. With a diverse organizing team and backed by industry sponsors, we dedicate the workshop to pushing the boundaries of ML's role in biology.

The year 2024 has seen an explosion of interest in humanoid robots. In the 7th Robot Learning workshop, to be held at ICLR-2025, we will look beyond the humanoid embodiment and ask: how far are we from robots with human-level abilities? What do we need to improve about embodied learning, decision-making, perception, and data collection to train generally physically capable robots to robustly perform a wide range of activities such as cooking or tidying up the house -- activities that people do without much thinking? We believe many of the weaknesses of the current robotic systems to be a reflection of the shortcomings of general AI methods and models. As such, in this workshop we will seek diverse perspectives from robotics-focused and robotics-orthogonal parts of the ICLR community alike, scientific contributions from academia and industry, as well as participants from a variety of backgrounds and career stages. Capitalizing on our prior experience with robotics showcases, in keeping with the spirit of the times we will solicit several humanoid robotics companies to exhibit their robots during the workshop's poster sessions.

Accessing large scale and high quality data has been shown to be one of the most important factors to the performance of machine learning models. Recent works show that large (language) models can greatly benefit from training with massive data from diverse (domain specific) sources and aligning with user intention. However, the use of certain data sources can trigger privacy, fairness, copyright, and safety concerns. The impressive performance of generative artificial intelligence popularized the usage of synthetic data, and many recent works suggest (guided) synthesization can be useful for both general purpose and domain specific applications. For example, Yu et al. 2024, Xie et al. 2024, Hou et al. 2024 demonstrate promising preliminary results in synthesizing private-like data, while Wu et al. 2024 highlight existing gaps and challenges. As techniques like self-instruct (Wang et al. 2021) and self-alignment (Li et al. 2024) gain traction, researchers are questioning the implications of synthetic data (Alemohammad et al. 2023, Dohmatob et al. 2024, Shumailov et al. 2024). Will synthetic data ultimately solve the data access problem for machine learning? This workshop seeks to address this question by highlighting the limitations and opportunities of synthetic data. It aims to bring together researchers working on algorithms and applications of synthetic data, general data access for machine learning, privacy-preserving methods such as federated learning and differential privacy, and large model training experts to discuss lessons learned and chart important future directions.

This workshop explores the intersection of scale-driven generative artificial intelligence (AI) and the correctness-focused principles of verification. Formal analysis tools such as theorem provers, satisfiability solvers, and execution monitoring have demonstrated success in ensuring properties of interest across a range of tasks in software development and mathematics where precise reasoning is necessary. However, these methods face scaling challenges. Recently, generative AI such as large language models (LLMs) has been explored as a scalable and adaptable option to create solutions in these settings. The effectiveness of AI in these settings increases with more compute and data, but unlike traditional formalisms, they are built around probabilistic methods – not correctness by construction. In the VerifAI: AI Verification in the Wild workshop we invite papers and discussions that discuss how to bridge these two fields. Potential angles include, but are not limited to the following: generative AI for formal methods, formal methods for generative AI, AI as verifiers, datasets and benchmarks, and a special theme: LLMs for code generation. We welcome novel methodologies, analytic contributions, works in progress, negative results, andreview and positional papers that will foster discussion. We will also have a track for tiny or short papers.

As foundation models (FMs) scale, they face a data bottleneck, where the growth of high-quality internet data unable to keep pace with their training needs. This is most apparent with text data already, has been a consistent problem in domains such as embodied intelligence, and is expected to soon inflict other modalities as well. Self-improvement, a paradigm where models generate and train on synthetic data generated from the same or other models, offers a promising solution. This paradigm differs from both supervised learning, which relies on curated human data, and reinforcement learning (RL), which depends on external rewards. Self-improvement frameworks require models to self-curate training data, often using imperfect learned verifiers, with unique challenges. This workshop will explore algorithms for self-improvement, covering topics such as synthetic data, multi-agent and multi-modal systems, weak-to-strong generalization, inference-time self-supervision, and theoretical limits.

Watermarking involves embedding a hidden signal into digital media like text, images, and audio to establish ownership or ensure authenticity. It has become increasingly important in the age of generative AI. However, despite its growing significance, watermarking in the AI community often gets lost in broader conversations around adversarial robustness, and general security, and safety. We argue that watermarking needs its own dedicated space in AI conferences for discussion and exploration, where researchers can dig deeper into the technical specifics of this field and build on a foundation of research spanning over 20 years. The aim of this workshop is to bring together experts from academia, industry, policy and from different communities to discuss advancements and challenges in watermarking technologies. The event will facilitate the exchange of ideas and collaborative problem-solving.

Topics of interest include, but are not limited to:
- Algorithmic Advances: Multi-modal watermarking, model watermarking, dataset tracing and attribution.
- Watermark Security: Theoretical results on strong watermark impossibility, black and white-box adversarial attacks, advanced threat models, open-sourced and publicly detectable watermarking, and zero-knowledge watermarking.
- Evaluation: Benchmarks for watermarking, perceptual models and watermark-specific quality evaluation metrics, and bias in watermarking robustness.
- Industry Requirements: Large bit watermarking, low FPRs, and complexities of deployment in-the-wild.
- Policy and Ethics: Dual use, communication to policy makers, and standards.
- Explainability and Interpretability: Understanding how watermarks work and their limitations, human oversight and review, and balancing automation with human judgment.

Machine learning for remote sensing (ML4RS) has emerged as a critical and exciting area of research, with the potential to address some of the most pressing global challenges, including climate change, food security, disaster management, and conservation. Remote sensing data, collected from diverse instruments capturing the Earth across various spatial, temporal, and spectral dimensions, offers unique research opportunities and challenges for the ML community. Unlike traditional data modalities, these datasets are high-dimensional, extremely multi-modal, and contain patterns at a multitude of spatial and temporal scales. These characteristics often require specialized approaches in cross-cutting ML topics like self-supervised/semi-supervised learning, domain adaptation/generalization, and multi-modal learning/data fusion to unlock their full potential. Our workshop will foster discussion and feedback on early-stage work that is critical to impactful applications and new developments in machine learning for remote sensing.

The ongoing deep learning revolution of the last decade has brought about hundreds of millions of neural networks (NNs) trained on diverse datasets.At the same time, the recent rise of foundation models has led to a rapid increase in the number of publicly available neural network models. On Hugging Face alone, there are over a million models, with thousands more added daily. As a result, the ample knowledge contained in the data, the abstraction learned via training, as well the trained models' behaviours themselves are stored in the architectures and parameters of trained NNs. Despite this massive growth, little research has been conducted into processing model weights, and they are rarely considered a data modality. This workshop aims to create a community around Weight Space Learning by bringing together the scattered sub-communities that already interface with model weights, with the ultimate goal of democratizing model weights as a proper data modality.

Large Language Models (LLMs) have emerged as transformative tools in both research and industry, excelling across a wide array of tasks. However, their growing computational demands especially during inference—raise significant concerns about accessibility, environmental sustainability, and deployment feasibility. At the same time, sparsity-based techniques are proving critical not just for improving efficiency but also for enhancing interpretability, modularity, and adaptability in AI systems. This workshop aims to bring together researchers and practitioners from academia and industry who are advancing the frontiers of sparsity in deep learning. Our scope spans several interrelated topics, including Mixture of Experts (MoEs), LLM inference and serving, network pruning, sparse training, distillation, activation sparsity, low-rank adapters, hardware innovations and quantization. A key objective is to foster connections and unlock synergies between traditionally independent yet highly related research areas, such as activation sparsity and sparse autoencoders (SAEs), or quantization and KV cache compression. Rather than focusing solely on efficiency, we aim to explore how sparsity can serve as a unifying framework across multiple dimensions of AI—driving advances in interpretability, generalization, and system design. By facilitating the fusion of ideas from different topics, the workshop will create new opportunities for innovation. We encourage participants to think beyond traditional constraints, exploring how different forms of sparsity can inform each other and yield new algorithms. Whether the goal is faster inference, modular architectures, or more interpretable models, our aim is to catalyze research that deepens the integration of sparsity within AI.

How can we trust large language models (LLMs) when they generate text with confidence, but sometimes hallucinate or fail to recognize their own limitations? As foundation models like LLMs and multimodal systems become pervasive across high-stakes domains—from healthcare and law to autonomous systems—the need for uncertainty quantification (UQ) is more critical than ever. Uncertainty quantification provides a measure of how much confidence a model has in its predictions or generations, allowing users to assess when to trust the outputs and when human oversight may be needed. This workshop aims to focus on the question of UQ and hallucination in the modern LLMs and multimodal systems and explore the open questions in the domain.

Current AI research and applications often prioritize adult-focused solutions, while progress in AI designed specifically for children's development, health, and education has lagged behind. Our workshop aims to spotlight this issue and bring together researchers from diverse fields to discuss the future of AI design and its applications for children.In the era of AI, developing bespoke AI systems for children holds special significance:(i) Advanced AI technologies, such as large language models (LLMs), have the potential to support children’s development, education, and mental health, posing a critical new frontier for research.(ii) AI in pediatric healthcare is essential, as early diagnosis of childhood diseases can lead to timely interventions, improving prognoses and reducing infant mortality rates.(iii) AI can also provide valuable tools helping children in low-resource countries, helping bridge gaps in education, healthcare, and other developmental supports.This workshop will invite researchers from the fields of AI, child psychology, education, pediatrics and social good to discuss how AI, particularly new generative models like LLMs, can address the unique challenges in pediatrics, child psychology, and education. We will also explore the potential risks associated with AI applications for children.The insights from the workshop's panel discussions will be summarized in a survey paper and submitted to a top-tier journal or AI conference after the workshop.

Our limited understanding of the biological mechanisms underlying diseases remains a critical bottleneck in drug discovery. As a result, we often lack insights into why patients develop specific conditions, leading to the failure of many drug candidates in clinical trials. Recent advancements in genomics platforms and the emergence of diverse omics datasets have sparked increasing interest in this field. The primary objective of this workshop is to bridge the gap between machine learning and genomics, emphasizing target identification and emerging drug modalities such as gene and cell therapies and RNA-based drugs. By fostering interdisciplinary collaboration, we aim to advance the integration of these disciplines and accelerate innovation in drug discovery.

This workshop will discuss the latest multidisciplinary developments in Associative Memory. A number of leading researchers in this topic from around the world have already agreed to attend and present their latest results. We anticipate sharing their presentations and outlining future research directions in this emerging field with the rest of the ICLR community.

Datasets are a central pillar of machine learning (ML) research—from pretraining to evaluation and benchmarking. However, a growing body of work highlights serious issues throughout the ML data ecosystem, including the under-valuing of data work, ethical issues in datasets that go undiscovered, a lack of standardized dataset deprecation procedures, the (mis)use of datasets out-of-context, an overemphasis on single metrics rather than holistic model evaluation, and the overuse of the same few benchmark datasets. Thus, developing guidelines, goals, and standards for data practices is critical; beyond this, many researchers have pointed to a need for a more fundamental culture shift surrounding data and benchmarking in ML. At present it is not clear how to mobilize the ML community for such a transformation. In this workshop, we aim to explore this question, including by examining the role of data repositories in the ML data landscape. These repositories have received relatively little attention in this context, despite their key role in the storage, documentation, and sharing of ML datasets. We envision that these repositories, as central purveyors of ML datasets, have the potential to instigate far-reaching changes to ML data and benchmarking culture via the features they implement and the standards they enforce (e.g., minting DOIs, requiring licenses, facilitating the provision of structured metadata).