Roadmap

Welcome to the Shared RSI Community! We focus on open-source research and engineering for Recursive Self-Improvement (RSI). Previously, we built GDPevo, an enterprise-grade evaluation benchmark for self-evolving agents, and authored a comprehensive survey on RSI research.

Currently, we are iterating along the following tracks:

• Data: Curating agent Benchmark datasets and Evaluation results to build the largest RSI Leaderboard.
• Survey: Surveying the latest RSI papers and reorganizing their taxonomy to further improve the Awesome RSI repository.
• Framework: Optimizing the PenguinHarness unified RSI framework to build the most comprehensive self-evolving harness.

Evolution Roadmap

If you are new to core concepts such as Agents, Self-Evolution (RSI), Harnesses, the distinction between Harness-Level and Model-Level Evolution, and Benchmarks, check out Basic Concepts first.

In Stage 1, we focus on Data, Survey, and Harness-Level Evolution. Because harness-level evolution does not require updating underlying model weights, it demands significantly less GPU compute and lower experimental cost, while producing human-interpretable prompts, skills, and tools that transfer across models—making RSI far more accessible and inclusive for the broader research and developer community.

In Stage 2, building upon the benchmarks, algorithm plugins, and high-quality evolution traces accumulated by the community, we will expand into Model-Level Evolution:

• Trajectory Synthesis & Distillation: Leveraging high-scoring execution traces and evolved skill assets produced by self-evolving harnesses across diverse benchmarks in Stage 1 to synthesize and curate training data, internalizing explicit external experience into model weights.
• Self-Evolving SFT & Agentic RL: Driven by verifiable reward signals from standardized community sandboxes and automated evaluators, conducting supervised fine-tuning (SFT) and agentic reinforcement learning (Agentic RL) on long-horizon tasks to directly strengthen the base model's planning, tool use, and self-correction capabilities.
• Harness × Model Co-Evolution: Closing the bi-level loop between outer-loop harness search and inner-loop parameter updates—where stronger harnesses synthesize higher-ceiling trajectories for model training, and stronger models in turn empower harnesses to tackle more complex task boundaries.

1. Stage 1: Data (RSI Leaderboard), Survey (Awesome RSI), Harness Evolution (PenguinHarness).
2. Stage 2 (Model-Level Evolution): Trajectory Synthesis (Experience Distillation), Model Training (SFT & Agentic RL), Co-Evolution (Harness × Model).

How to Participate

For detailed workflows, please refer to the Workflow pages:

• How to submit a Benchmark: https://club.penguin.ooo/workflow/benchmark/
• How to submit an Evaluation: https://club.penguin.ooo/workflow/evaluation/
• How to submit an Algorithm Plugin: https://club.penguin.ooo/workflow/algorithm/
• How to contribute to PenguinHarness: https://club.penguin.ooo/workflow/harness/
• How to submit papers to Awesome RSI: https://club.penguin.ooo/workflow/survey/

We hold Bi-weekly Sync community meetings for discussion, with meeting links posted in our community group the day before.

We warmly welcome everyone to join our open-source community, and we commit to the following for contributors:

• Active monthly contributors will receive token resource packages for frontier models such as Gemini.
• Open-source contributors with significant contributions will be credited as authors in our open-source repositories.
• Contributors who submit original methods or original experimental results will be invited as co-authors on our upcoming papers.