Flex-RSI: A Benchmark for Evaluating Recursive Self-Improvement in Agents

Jingdong Zhang and other authors propose and release Flex-RSI, a benchmark for evaluating agents' recursive self-improvement capability under a fixed budget, covering perception, reasoning, robotic action, 3D reconstruction, visual creation, and infrastructure tasks. Under the benchmark's design, each round the executor improves the task solution, and changes are kept only if the score on held-out validation data is higher; then the improver modifies the execution framework, including the executor and the improver itself. Each run starts from the same general framework, and the final solution is scored on hidden test data. The benchmark is used to measure whether agents can continuously improve on vision and robotics tasks through their own modifications, rather than merely completing a single task.

AI 根据报道生成 · 2 days ago更新