How to benchmark RSI

How much of realistic LLM training do today's benchmark environments cover?

Recursive self-improvement (RSI) could have major consequences for AI development, yet its definition, its possible forms and its measurement remain open questions. Recent reports from Anthropic, OpenAI and Z.ai describe AI systems taking on parts of model research and infrastructure engineering, which makes these questions increasingly practical. Whatever definition settles, the ability to train and improve AI models is a core capability worth evaluating. For RSI that runs through model training, this capability is necessary but not sufficient. One way to evaluate progress toward RSI is to ask an AI system to develop and train a model end to end and deliver it ready for real-world use. At production scale, such an evaluation would be expensive in both compute and time. We therefore examine whether existing coding-agent benchmarks provide practical tests of the capabilities this process requires, and which capabilities their environments leave untested.

Meanwhile, coding benchmarks such as SWE-Bench Pro, Terminal-Bench, DeepSWE and ProgramBench assess progress in coding agents, and executable task environments of the same kind also support agent training. How well these environments represent the work of developing production LLMs is therefore an important question. This post audits selected tasks and documented environments from ten benchmark suites against eighteen documented model-development pipelines; the task inventory and the method and sources are given below. We identify relatively few tasks involving LLM infrastructure in the general-purpose benchmarks examined. The specialized benchmarks provide substantial training and inference coverage, but the audited tasks largely test individual components or bounded experiments, leaving gaps in coordinated, end-to-end workflows, particularly distributed agentic RL and recovery from failures.