RSIAgent: parameter-free self-improvement pushes open-source models past GPT-6 Astra on agent benchmarks
机器之心 · wechat · 2026-09-15
Causal-intelligence startup AetherAI released RSIAgent, proposing a new scaling axis: scale experience, not model parameters. The agent autonomously explores new software environments, distills stable action-condition-outcome causal rules into an evolving memory, and recursively self-improves.
Framework
- A three-agent loop: a curriculum agent picks what to explore, actor agents execute, verifier agents check against real environment feedback.
- Two-phase strategy: Broad exploration (parallel, pretraining-like coverage) then Deep exploration (sequential, adversarial attack on hard/hidden corner cases); memory is frozen and reused downstream.
Results
- With GLM-5.3 and Kimi-K3, OSWorld 2.0 jumps from 71.97% to 78.98%, beating GPT-6 Astra (72.60%) and Claude Opus 5; 84.82% on Agents' Last Exam (Near-term).
- Transfers to autonomous game development on GameCraft-Bench; some tasks starting at 0–40% success reach 100% after more RSI rounds.
- Ablations confirm both phases are necessary.
Paper, code and project page are public. The team frames this as a step toward general causal intelligence.
Related event: RSIAgent Achieves Recursive Self-Improvement Without Training(3 posts)→
More from coding & agent
- Cohere Labs agent team-construction method cuts multi-agent costs by 70% — Cohere_Labs · 2026-11-02
- DeepSeek design doc: a person can only follow 3-5 agent sessions at once, attention is the real bottleneck — heyshrutimishra · 2026-11-02
- Software fundamentals are key to getting good results from AI coding agents — bendee983 · 2026-11-02
- Open-source LLM regression gate digline hits 1.0: quantifying judge noise before calling a regression — alexpran · 2026-11-02
- PR adds patient retry logic for model capacity errors in OpenAI Codex — bytebot · 2026-11-02
- Perplexity open-sources WANDR: 500 research tasks, 170,495 evidence points, URL-rechecking grader — thisguyknowsai · 2026-11-02