Flex-RSI benchmark tests whether AI agents can improve their own harness
The arXiv paper has an improver agent revise the executor's harness on vision and robotics tasks and finds binary feedback holds back open-weight models.
Why it matters
Recursive self-improvement, where AI systems upgrade the tools and scaffolding they run in, is one of the paths forecasters watch most closely, and few benchmarks measure it directly. Flex-RSI sets an improver agent to revise an executor agent's harness on vision and robotics tasks and scores the gains. The paper finds that binary success-or-failure feedback holds back open-weight models. It is a preprint and has not been peer reviewed. Watch for results from frontier closed models and independent replications.
Key facts
- Improver agent revises an executor agent's harness
- Tasks: vision and robotics
- Binary feedback holds back open-weight models
- arXiv preprint 2611.08946, not peer reviewed