This workshop explores the use of programmatic representations to enhance the interpretability, generalizability, efficiency, and scalability of agent learning frameworks. By leveraging structured representations—such as symbolic programs, code-based policies, and rule-based abstractions—agents can achieve greater interpretability, improved generalization, and enhanced efficiency. Programs can explicitly encode policies, reward functions, task structures, and environment dynamics, providing human-understandable reasoning while reducing the reliance on massive data-driven models. Furthermore, programmatic representations enable modularity and compositionality, allowing agents to efficiently reuse knowledge across tasks and adapt with minimal retraining. By bringing together the sequential decision-making community—including researchers in reinforcement learning, imitation learning, planning, search, and optimal control—with experts in program synthesis and code generation, this workshop aims to tackle the fundamental challenges of agent learning at scale and drive progress toward interpretable, generalizable, verifiable, robust and safe autonomous systems across domains ranging from virtual agents to robotics.

The 2024 Nobel Prize in Chemistry was awarded to AI-based protein structure prediction and protein design. It highlights the immense potential of AI in basic science and health research. In the meantime, generative AI models such as large language models (LLMs) and diffusion models are acquiring impressive capabilities in generating language, creating artwork, solving complex reasoning problems, writing computer programs, etc. To further facilitate the dialog between machine learning and biology, we propose to organize a workshop at ICML 2025, focusing on generative AI for biological discovery and therapeutic design. By fostering connections among preeminent researchers from both industry and academia, we aim to gain critical insights into the future of generative-AI-driven biology. Moreover, we hope to bridge the gap between machine learning and biological disciplines by focusing on three central themes that encapsulate innovative research as well as practical implications, which span both cutting-edge research and translational impact.

Vector databases (Vector DBs) are a foundational and critical application layer for injecting information into large language models (LLMs). Although different companies have proposed various vector databases, no academic workshop has previously existed to discuss these systems comprehensively. This workshop aims to foster discussions on vector databases from various perspectives, ranging from mathematical theories to implementation-level optimizations. Topics covered in the workshop include retrieval-augmented generation (RAG), algorithms and data structures for approximate nearest neighbor search (ANN), data management systems for handling vector data, query languages, and embedding models. Furthermore, the workshop will also function as a platform for companies and researchers working on vector databases to present technical details (white papers) and exchange ideas.

The rapid evolution of Deep Learning, propelled by transformer-based architectures and significant hardware advancements, has unlocked unprecedented capabilities across diverse domains, from biological sciences to autonomous systems. As foundation models continue to scale, they introduce new challenges in resource management, particularly in data centers, and data availability prompting us to broaden our exploration of leveraging distributed and on-device resources for training and inference. Small Language Models (SLMs) are emerging as a compelling alternative for generative AI, particularly at the edge, offering a sustainable balance between efficiency and user privacy. This workshop aims to bring together algorithms and systems experts to discuss the opportunities and challenges of on-device machine learning. We hope to explore to what extent SLMs can compete with or complement LLMs and identify methods to enhance their quality and efficiency. Addressing this shift requires innovation in algorithm and system co-design, underscoring the importance of interdisciplinary approaches for future applications.

Mathematical reasoning stands as a pinnacle of human intelligence. The rapid advancements in artificial intelligence, particularly in large language models (LLMs), have opened new frontiers at the intersection of AI and mathematical reasoning. This workshop aims to explore the potential of AI in comprehending and advancing mathematical reasoning, with a focus on fostering collaboration between humans and machines to push the boundaries of mathematical discovery. The central theme revolves around the question: >``How can we leverage and advance the mathematical reasoning abilities of machine learning models, and drive innovation across scientific and practical domains?''Our workshop will bring together researchers from diverse backgrounds, institutions, and disciplines to discuss the progress and future of AI technologies in mathematics. Specifically, we will delve into the areas related to the following:* Automated Theorem Proving: How can we build consistent theorem-proving systems? How can theorem-proving systems assist humans through human-computer interaction?* Automated Theorem Generation: Can neural models generate new and practically meaningful theorems that have been discovered? How can we utilize these newly generated theorems?* Autoformalization and Verification: How can we improve the precision of translating natural language proofs into formal proofs, and vice versa?* Problem Solving: How can we develop AI models to solve complex mathematical computational problems across various domains? How can AI models improve themselves during the learning process?* Applications of AI in Mathematics: What are the practical applications of AI-driven mathematical reasoning in various fields such as sciences, engineering, finance, and education?The intended outcome is to identify new ideas, open problems, and interdisciplinary areas for future research related to mathematical reasoning. To this end, we welcome papers on areas related, but not limited, to:* Algorithm: How to develop effective algorithms (e.g., reinforcement learning, self-improve/evolve) to improve reasoning ability?What are the key principles for developing algorithms that minimize resource consumption (e.g., time, memory) while maintaining or improving reasoning performance?* Data Generation: Can AI models generate questions that they cannot answer correctly?Can AI models achieve self-improvement through self-generated data?* Tool Utilization: How can AI systems leverage existing tools, such as code and software, to solve practical mathematical problems more effectively?* Limitation Analysis: What are the drawbacks or limitations of current models in mathematical reasoning (e.g. robustness, generalization, and reasoning boundary)? How can these limitations be quantitatively analyzed?

Generative models across domains are capable of producing outputs that appear to mimic the real world. But have these systems actually understood the laws that govern the world? Researchers across subfields are attempting to answer this question: in natural language processing, researchers measure whether LLMs understand real-world mechanisms in order to measure how robust they are to new tasks; in video generation, researchers assess whether a model has understood the laws of physics in order to evaluate how realistic its videos are; in scientific domains, foundation models are being developed in order to uncover new theories about the world. Despite studying similar questions, these communities remain disparate. This workshop will explore the question: how can we formalize and evaluate whether generative models have understood the real world? While this question is important across communities, we don’t have unified frameworks for defining and evaluating world models. This workshop will bring together these computer science communities along with non-computer-science scientists working on relevant applications.Our invited speakers include Jacob Andreas, Shiry Ginosar, Shirley Ho, Sendhil Mullainathan, and Martin Wattenberg, all of whom have confirmed they will be speaking and that they can make it in-person.

Tokenization defines how data are represented as input and output for many current machine learning systems, including language models. Tokenization has been shown to significantly affect the utility and effectiveness of these models (Mielke et al., 2021). This finding has stirred considerable interest in tokenization as a research direction in machine learning and its subfields, such as natural language processing, but currently, there is no venue specifically dedicated to it. Our initiative—TokShop (Tokenization Workshop)—aims to fill this gap and will focus on tokenization in a broad sense.

Our workshop brings together experts in machine learning, cognitive science, behavioral psychology, and economics to explore human-AI alignment by examining human (and AI) feedback mechanisms, their mathematical models, and practical implications. By fostering collaboration between technical and behavioral science communities, we aim to develop more realistic models of human feedback that can better inform the development of aligned AI systems.

The scaling of model parameters has unlocked the groundbreaking capabilities of foundation models. Likewise, in human society, scaling and collaboration across individuals, organizations, companies, and nations amplify collective intelligence to unprecedented levels, enabling remarkable achievements that would be impossible for individuals alone, such as space exploration and autonomy. Could this principle of scaling~\cite{kaplan2020scaling} also apply to the growth in the number of agents? Multi-agent systems may offer a promising path forward. By progressively integrating more agents, multi-agent systems can activate diverse functionalities within these foundation model-powered generalist agents and coordinate a broader range of complementary functionalities. This synergy fosters improved problem-solving, adaptability, and decision-making capabilities. As the multi-agent system scales, it has a huge potential to achieve enhanced capabilities and tackle increasingly complex tasks, offering a promising solution toward the ultimate goal of achieving artificial general intelligence (AGI).

Structured data foundation models are an emerging area of research undergoing rapid growth, yet they still remain critically under-explored relative to image and text modalities. So far, the different structured data sub-communities have had little opportunity to come together and share insights about how to build foundation models for structured data. Yet, strong synergies exist across modalities since models share similar pre-training and in-context learning paradigms. Furthermore, models trained on one modality can also demonstrate promising predictive performance in another. This workshop brings together the tabular and time series communities to jointly discuss foundation models for structured data, enabling the communities to capitalize on their synergies. We aim for advancements in foundation models that unify structured data modalities, addressing challenges of scalability and generalization across real-world applications. This emerging field promises to transform how we approach structured data analysis and drive new opportunities across various domains.

Deep learning has advanced by scaling datasets, models, and training computation. At the same time applications have broadened to many kinds of data (personal, scientific, …) and deployments (in clouds, on cars, …). Will these all be solved by more data, parameters, and training? Test-time updates are complementary, and can help on both foundation model servers and edge devices. This workshop examines train-time vs. test-time updates across scales by test-time adaptation, continual learning, in-context learning, and post-training model editing. The test begins now!

Open-source software (OSS) development is a cornerstone of modern machine learning research. However, issues such as the sustainability of long-term projects, software reliability, and proper academic acknowledgment of maintenance and contributions are often overlooked. This workshop aims to identify and discuss strategies for successful and sustainable open-source development in ML while also proposing solutions to these challenges. Additionally, the workshop will provide a platform to recognize the efforts of open-source contributors in the field. We will bring together machine learning researchers, engineers, industrial practitioners, and software development experts. The workshop will feature invited talks, panel discussions with experts, and workshop paper submissions from open-source contributors in machine learning.

Machine learning and AI have long been concerned about modeling how an agent can change the world around it. However, intervening in the physical world takes effort, leading to sparsity of evidence and the corresponding gaps of credibility when an agent considers carrying out previously unseen actions. Making the most of sparse data within a combinatorial explosion of possible actions, dose levels, and waiting times requires careful thinking, akin to efforts for introducing more compositionality principles into machine learning (Andreas, 2019). The goal of this workshop is to bring together state-of-the-art ideas on how to predict effects of novel interventions and distribution shifts by exploiting original ways of composing evidence from multiple data-generation regimes.

Modern machine learning applications face the challenge of extracting insights from high-dimensional datasets. The 3rd High-dimensional Learning Dynamics (HiLD) Workshop focuses on predicting and analyzing the behavior of learning algorithms in regimes where both the number of samples and parameters are large. This workshop aims to advance research and foster collaboration in several key areas:1. Developing tractable models and dynamical frameworks to explain phenomena observed in deep neural networks (DNNs) and foundation models;2. Establishing mathematical frameworks for neural scaling laws as network width and depth approach infinity;3. Identifying and characterizing relevant observable quantities in high-dimensional limits;4. Understanding the provable effects of optimization algorithms, hyperparameters, and neural architectures on training and test dynamics.The HiLD Workshop will unite experts from random matrix theory, optimization, high-dimensional statistics/probability, and statistical physics to share diverse perspectives on these challenges. By bringing together theorists and practitioners from machine learning with researchers from these adjacent fields, we aim to create new collaborations between communities that often do not interact. Through talks, poster sessions, and panel discussions, the workshop will explore the fundamental dynamics of learning algorithms in high-dimensional settings. This year's workshop theme is "Navigating Complexity: Feature Learning Dynamics at Scale."

Generative AI models are trained on internet-scale datasets, yielding powerful capabilities but also introducing risks like copyright infringement, PII leakage, and harmful knowledge. Targeted removal or unlearning of sensitive data is challenging, as retraining on curated sets is computationally expensive, driving research into machine unlearning and model editing. Yet approaches like RLHF only suppress undesirable outputs, leaving underlying knowledge vulnerable to adversarial extraction. This raises urgent privacy, security, and legal concerns, especially under the EU’s GDPR “right to be forgotten”. Because neural networks encode information across millions of parameters, precise deletion without degrading performance is complex, and adversarial or whitebox attacks can recover ostensibly erased data. This workshop brings together experts in AI safety, privacy, and policy to advance robust, verifiable unlearning methods, standardized evaluation frameworks, and theoretical foundations. By achieving true erasure, we aim to ensure AI can ethically and legally forget sensitive data while preserving broader utility.

As wireless communication systems evolve to meet the demands of a hyper-connected world, artificial intelligence models are emerging as the driving force behind a new wave of technological innovation. This workshop will explore how state-of-the-art artificial intelligence and machine learning (ML) methods are poised to redefine the core of wireless networks providing solutions to old and new communication challenges. One of the central themes is semantic communication, where ML enables wireless networks to understand and transmit the meaning behind data, rather than the whole bitstream, drastically improving efficiency in bandwidth-constrained environments and presenting novel scenarios and possible applications that were not even conceivable a couple of years ago. Additionally, the rise of generative and language models for wireless communication is bringing new ways to compress and enhance signal transmissions, impacting several downstream applications such as autonomous driving, video streaming, and virtual reality. Concurrently with widening the range of applications, these models also bring novel challenges related to large models' computational demands or to the regenerated content's controllability and reliability. Central to bridging ML and wireless communication is the study of inverse problems, where generative models play a pivotal role in reconstructing lost or incomplete signals, and solving ill-posed tasks inherent in communication systems constrained by noisy and interference channels with limited bandwidth. The workshop aims also to explore key areas such as multimodal content compression, post-training quantization, efficient semantic feature extraction, and designing trustworthy models tailored for resource-constrained and noisy environments, in which foundational ML research finds crucial applications in communication scenarios.

Workshop Goals: This workshop aims to foster collaboration between ML researchers and wireless communication experts, encouraging cross-disciplinary innovation that will help shape the future of intelligent communication systems as well as more efficient and reliable AI models and techniques. Through a series of presentations, discussions, and interactive sessions, participants will explore both the theoretical foundations and practical applications of ML in wireless networks, with an eye toward addressing the most pressing challenges in this rapidly evolving field. On top of fostering collaborations and networking, we aim to boost research in machine learning and wireless communication topics by i) Hosting discussions about current wireless communication method limitations and how diverse ML models can empower communication systems by solving those challenges. ii) Encourage cross-collaboration between ML researchers and communication ones. iii) Giving space to younger researchers and PhD students to present their works and to get in contact with experts in this area, which is usually arduous in main conference tracks.

Why This Workshop at ICML? We know that artificial intelligence and machine learning models are driving technological transformations across numerous applications, with a particularly significant impact on wireless communication, given our daily reliance on smartphones and the emergence of connected intelligent devices, ranging from autonomous cars to mobile humanoids. Nonetheless, few ML researchers are actively contributing to wireless communication communities and venues, leaving researchers from this field alone in developing AI-powered methods and systems. On the other hand, ML researchers working on communication- potentially interesting topics like compression, quantization, inverse problems, or reliability sometimes lack real-world scenarios, datasets, or embedding systems to test their foundational research. We believe there is an unmet need to bridge the gap between the two research worlds. With this workshop, we aim to close this gap by fostering an active exchange and discussion between ML and communication researchers that can benefit both the research communities and establish a starting point for future collaborations and connections between the two worlds.