RSIAgent: Training-Free Agent Self-Improvement Beats GPT-6 on OSWorld

alex_verem · x · 2026-09-21

RSIAgent, from Aether AI with UC San Diego and UIC researchers (arXiv:2609.15364), is a training-free multi-agent framework achieving recursive self-improvement via autonomous memory construction.

Mechanism: like a new hire learning software — a curriculum agent sets practice tasks, an actor tries them, and a verifier checks results (without seeing the actor's reasoning, judging outcomes only). After every task the agent writes notes; the underlying model never changes. Exploration is broad-then-deep: wide parallel self-exploration to map the environment, then focused deep dives on hard cases and hidden causal dependencies. The resulting memory is frozen and reused directly for downstream tasks.

Results: on 82 OSWorld 2.0 long tasks, open models GLM-5.3 and Kimi-K3 improved from 71.97 to 78.98 average after practice, beating GPT-6 Astra's reported 72.60; a video watermark-removal task went from 0.40 to a perfect 1.0.

Caveats the authors list:

The agent practices on the test tasks themselves before being scored
Other models were scored with their own setups, so comparisons aren't matched; on a second benchmark GPT-6 Astra still finished more tasks perfectly
The verifier sometimes approved wrong work, which entered notes as rules — e.g., learning "missing information means no"

Takeaway: self-teaching agents need better verifiers before their notes can be trusted.

Original post →

RSIAgent: Training-Free Agent Self-Improvement Beats GPT-6 on OSWorld

More from coding & agent