People– no linked people
Updated 04/02/26Funding Details
Updated 04/02/26- Annual Budget
- -
- Current Runway
- -
- Funding Goal
- -
- Funding Raised to Date
- -
Org Details
Updated 04/02/26The Computational Rational Agents Laboratory (CORAL) is an AI alignment research group founded by Vanessa Kosoy that focuses on developing rigorous mathematical foundations for ensuring AI systems remain aligned with human values. CORAL spun off from ALTER (Association for Long Term Existence and Resilience) Israel as an independent organization in 2025, building on years of foundational research previously supported by MIRI and the Long-Term Future Fund. CORAL's core research program is the learning-theoretic agenda (LTA), which seeks to build a comprehensive mathematical theory of agents using tools from computational learning theory, control theory, algorithmic information theory, and categorical systems theory. The agenda's key insight is that future existential risks from AI will come from systems that are powerful agents -- masters at learning, planning, and adapting in pursuit of their objectives -- and that a rigorous mathematical theory of such agents would apply across different AI architectures, much like thermodynamics applies to any physical machine. The group's major theoretical contributions include infra-Bayesianism, a mathematical framework developed by Vanessa Kosoy and Alex Appel that extends Bayesian decision theory to handle non-realizability -- situations where the true environment is not contained in the agent's hypothesis space. This framework addresses critical problems in AI alignment including embedded agency, decision theory, and multi-agent interaction. More recent work focuses on compositional learning theory, ambidistributions, physicalist superimitation, and infra-Bayesian decision estimation theory. CORAL's team consists of Vanessa Kosoy (Founder and Research Lead) and Alex Appel (Researcher), with additional collaborators and mentees through programs like PIBBSS and MATS. Vanessa Kosoy is also a PhD student in Shay Moran's group at the Technion, where her doctoral research and CORAL's research agenda are closely intertwined. The group has published peer-reviewed work at venues including COLT 2025 and the Journal of Machine Learning Research, alongside extensive posts on the Alignment Forum and LessWrong. CORAL operates as a fiscally sponsored project of Ashgro, a US-based 501(c)(3) public charity that provides fiscal sponsorship to AI safety projects. ALTER Israel continues to provide some administrative support to CORAL.
Theory of Change
Updated 04/02/26CORAL believes that the core risk from advanced AI comes from powerful agents that learn, plan, and adapt in pursuit of their objectives. If such agents' objectives are misaligned with human values, the results could be catastrophic. CORAL's theory of change is that by developing a rigorous mathematical theory of computationally bounded agents -- drawing on learning theory, control theory, and algorithmic information theory -- researchers can produce formal proofs that guarantee alignment under well-defined assumptions. This architecture-independent theoretical framework would allow AI designers to understand agentic capabilities and failure modes, translate between AI and human ontologies, and verify that training procedures produce aligned systems. The work aims to provide the mathematical foundations that make provably safe AI possible, analogous to how thermodynamics provides guarantees about physical machines regardless of their specific design.
Grants Received
Updated 04/02/26Projects– no linked projects
Updated 04/02/26Discussion
Key risk: The main risk is that this highly formal, idiosyncratic agenda may remain siloed from frontier ML practice—with a two-person team and limited institutional ties—yielding elegant theory that sees little adoption or influence on real-world AI systems.
Case for funding: CORAL is uniquely advancing a rigorous, architecture-independent alignment framework—via infra-Bayesianism and the broader learning-theoretic agenda—with peer-reviewed results (e.g., COLT 2025, JMLR) that directly tackle non-realizability and embedded agency, offering a plausible path to provable guarantees that could reshape how future labs design safe agents.