RSI Data

RSI data is the source’s label for long-running trajectories in which one model, agent, or training loop helps improve another model or its own future behavior. In 从蒸馏到合成数据到 RSI，模型竞争的下一个焦点是什么？｜对谈 Evolvent AI 联创孟繁青, [[MengFanqing|孟繁青]] describes this as a likely next demand after ordinary agent data: a model post-trains a smaller model, checks whether it improves, and leaves a complete trace of data, training, evaluation, and revision.

This page treats RSI data as a bridge between Synthetic Agent Data and Recursive Self-Improvement. It is not just a good task answer; it is evidence from an improvement loop. That makes it expensive because the trajectory may require hours of environment execution, training knowledge, evaluation design, and anti-cheating checks.

Key Claims

• RSI data records improvement attempts, not only task completion.
• Valuable trajectories may include environment setup, generated data, training runs, evaluation results, error analysis, and revised strategy.
• The source expects model labs to want this kind of data as ordinary agent-data demand becomes more concentrated around fewer high-value benchmarks.
• RSI data needs hands-on model-training expertise; it cannot be produced reliably by generic crowdsourcing alone.
• The concept sits between current post-training data markets and stronger Auto Research or Recursive Self-Improvement systems.

Connections

• Recursive Self-Improvement, Auto Research, and AI For AI — broader self-improvement and research automation frame.
• Synthetic Agent Data, Environment-Based Agent Benchmarks, and RSIbench-data — data, benchmark, and project layer.
• Agent Post-Training, Model Post-Training Bottleneck, and AI Verification — training and validation requirements.
• Data Pricing In AI and AI Data Infrastructure — why these long trajectories can carry high value.
• Evolvent AI and Meng Fanqing / 孟繁青 — source company and speaker.