MIT-IBM & Amazon Launch OpenRSI-Index: Open Platform for AI-Driven Research

MIT-IBM Watson AI Lab and Amazon's A-EVO Lab have jointly open-sourced OpenRSI-Index , an experimental infrastructure for Recursive Self-Improvement (RSI) that lets AI agents conduct research autonomously. The project aims to lower the barrier for researchers to explore AI-driven scientific discovery without building custom experimental pipelines from scratch.

Advisory Board

The initiative is advised by prominent researchers including Wenhu Chen (University of Waterloo), Yejin Choi (Stanford), Pang Wei Koh and Luke Zettlemoyer (University of Washington), and Dawn Song (UC Berkeley).

Three Core Open-Source Components

1. RSI-Anything: Turning Ideas into Runnable Tasks

RSI-Anything is the task contribution workflow. Researchers select an existing open-source project or a familiar domain, then use coding agents (e.g., Codex, Claude Code) guided by a proposal-agent skill to define:

Research question and baseline approach

Evaluation metrics and success criteria

Allowed modification scope (which files can be changed)

Compute budget (e.g., maximum GPU hours)

Example: To test whether a new training strategy improves model performance, the agent helps specify the model, dataset, original training recipe, editable files, evaluation metrics, and GPU budget. These details form a task proposal submitted via GitHub Discussions for review before environment provisioning.

2. OpenRSI-Index: A Real-World Benchmark for AI Research

OpenRSI-Index serves as the central evaluation benchmark. Each task includes a clear research goal, baseline implementation, resource budget, and verification method. Coverage extends beyond large language models to vision models, data processing, training systems, inference optimization, and robotics.

3. RSI-Harness: Secure Execution and Evaluation Framework

RSI-Harness executes tasks in isolated environments (single-node Docker to multi-node clusters). Key design principles:

Separation of agent workspace and evaluation: The agent cannot modify scoring code, preventing reward hacking.

Explicit modification boundaries: Tasks declare which files/variables are mutable and which are frozen; violations invalidate results.

Full trajectory logging: Snapshots, configs, and logs are preserved for reproducibility and audit.

When an agent submits a candidate, Harness saves a workspace snapshot and runs an independent evaluation against predefined rules. The agent then receives feedback for iterative improvement.

Getting Started Tutorial

Clone the repository:

git clone https://github.com/OpenRSI-Foundation/OpenRSI-Index.git
cd OpenRSI-Index

Launch a coding agent and invoke the proposal skill:

Use the proposal-agent skill at .agents/skills/proposal-agent/SKILL.md in my local OpenRSI-Index repository. Locate the repository if needed, then follow the skill to complete setup and guide me step by step through creating and submitting an RSI proposal.

The agent interactively helps define goals, baselines, evaluation criteria, and resource budgets, then generates a proposal.

Submit and run the task. After review, the agent submits the proposal to GitHub Discussions for inclusion in the benchmark.

Community Expansion

Beyond code and tasks, OpenRSI is recruiting Collaborators and Domain Leads worldwide to grow the RSI experimental benchmark across research areas.