Announcing RSI Bench

We are entering the emergence of AI systems entering the loop that builds AI. The open question is whether models can achieve recursive self-improvement: autonomously building the next generation of models without human intervention. A prerequisite of self-improvement is the ability to master skills involved in all stages of frontier AI development.

What is RSI Bench?

RSI Bench is an ongoing effort to evaluate whether AI agents can develop the capabilities required to advance AI R&D. Our goal is to collect tasks curated by experts in frontier AI development that evaluate autonomous research capabilities for open-ended, long-horizon workflows requiring agents to independently experiment, iteratively improve, and complete complex multi-step research processes.

Each task provides a starting environment, a compute budget, and a verifier suite that measures performance against a baseline. Beyond final outcomes, RSI Bench evaluates broader agent behaviors, including:

• Reliability: whether agents produce real improvements and distinguish meaningful progress from noise,
• Efficiency: how effectively they find solutions under limited resources,
• Generality: whether their solutions extend beyond narrow task-specific fixes.
• Idea quality: whether they can extract insights, synthesize knowledge, and develop novel approaches.

How We’re Building RSI-Bench

We are excited to invite the community to contribute tasks in their domain of expertise to RSI-Bench. See Call for Contributions for incentives and the process.

Sample Tasks