Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agents
Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agents
This research introduces Flex-RSI, a benchmark for evaluating recursive self-improvement in frontier agents performing vision and robotics tasks within a fixed budget. The study demonstrates that editing the agent's harness contributes to performance gains, while recording only success or failure hinders open-weight models. Results indicate that open-weight models struggle most when outputs require strict checks of physical stability or exact correctness.
The paper introduces Flex-RSI, a benchmark evaluating recursive self-improvement in frontier agents for vision and robotics tasks under fixed budgets, finding that harness editing contributes to gains while strict physical stability checks widen the gap between open-weight and closed-source models.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
Source：arXiv Robotics — research abstracts · arxiv.org