We're actively looking for contributors to add new, challenging tasks. Our dedicated agent-native RSI-Anything pipeline helps you create a new task in an hour or less—bring the idea, and our agents handle the rest. See CONTRIBUTING.md for a step-by-step guide to creating and submitting tasks. Contributors with accepted tasks will be invited as co-authors. Have fun! 😀
Clone the repository:
git clone https://github.com/OpenRSI-Foundation/OpenRSI-Index.git
cd OpenRSI-IndexOpen Codex or Claude Code (app or CLI), start a fresh session, and paste:
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.
Any questions? Contact us at yuetaili@uw.edu, zhuofengli12345@gmail.com, or yfeng42@uw.edu.
If you have compute resources to run experiments and want to build frontier RSI environments together, contact yuetaili@uw.edu, zhuofengli12345@gmail.com, or hanxinyang@berkeley.edu.
We are witnessing the dawn of a new era: AI is entering a recursive self-improvement loop. The central question is whether this loop can move beyond the best-known human-designed method and reliably extend the scientific frontier. Answering it requires careful measurement, and building that measurement is the purpose of this project.
OpenRSI Index is an ongoing effort to evaluate whether AI agents can drive genuine recursive self-improvement and advance scientific discovery through real-world research at scales ranging from a single node to thousands of GPUs—not merely reproduce existing results, sweep parameters, or succeed on toy-scale tasks.
We are actively developing this project and welcome contributions from the community. Tasks and execution logs are available in rsi-tasks/ and rsi-logs/, respectively. Open the RSI Logs Viewer to explore results visually. See Quick Start for instructions on running these tasks.
View tasks
| Task | Track | Category | Description | GPU requirement |
|---|---|---|---|---|
Qwen-122B-RL · Research report |
Signature Task | Post-training | Optimize a full post-training stack under one fixed budget. | 256× H100 |
marin-optimizer-update-geometry |
Signature Task | Pre-training | Design a scale-general optimizer for the Marin scaling ladder. | 256× H100 |
gpic-text-to-image · Research trajectories |
Signature Task | Vision | Improve a shared PixelGen checkpoint on a fixed GPIC 10M subset. | 4× H100 per job |
depth-width-allocation |
Public Task | Pre-training | Optimize decoder-width allocation under a fixed 200M training budget. | 8× H100 |
learnability-cot |
Public Task | Post-training | Adapt reasoning traces for small-model math SFT. | 4× H100 |
gemm-h100-refined |
Public Task | MLSys | Optimize an FP16 CUDA GEMM kernel for H100 throughput. | 1× H100 |
liger-tied-ce |
Public Task | MLSys | Optimize tied-weight fused cross-entropy for Qwen3 SFT. | 2× H100 |
minference-sparse-prefill |
Public Task | MLSys | Optimize Triton sparse-prefill attention on H100. | 1× H100 |
molmo2-pointing-refined |
Public Task | Vision | Optimize Molmo2 video-pointing at inference time. | 1× H100 |
datacomp-small-filtering |
Public Task | Vision | Curate DataComp-small data for fixed ViT-B/32 training. | 32× H100 |
ace-playbook-repair |
Public Task | Agents | Inspect and repair ACE playbooks for a frozen Qwen model. | 1× H100 |
molmoweb-interaction-context |
Public Task | Vision | Allocate inference-time context for frozen MolmoWeb replay. | 2× H100 minimum |
reasonir-difficulty-curriculum |
Public Task | Post-training | Optimize a fixed-budget retrieval LoRA through difficulty-aware curricula. | 4× H100 minimum |
isaaclab-peginsert-reward-search |
Public Task | Robotics | Design bounded reward graphs for fixed-budget Isaac Lab peg insertion. | 4× H100 |
kev-decision-architecture |
Public Task | Post-training | Redesign a 0.5B decision architecture under a fixed training corpus and recipe. | 2× H100 minimum |
RSI-Harness powers OpenRSI Index for ultra-long-horizon RSI runs, natively supporting Harbor-format tasks from single-node local Docker to multi-node clusters.
We are still actively expanding the team. We really appreciate the efforts of all our excellent members.