RSIFlow: Adaptive Self-Improvement Workflows with an Evolving Meta-Agent

RSIFlow. Adaptive Self-Improvement Workflows with an Evolving Meta-Agent. By Huan Zhang and six others, from Beihang University, Tsinghua University, and one other institution. Published on the fifteenth of October. Its code repository is worked out: about one thousand source files, with a readme and a licence.

Large language model agents now handle complex tasks such as code generation, tool use and multi-hop question answering, and a family of self-improvement methods makes them better. Some optimize the model, some evolve the harness around it, some refine executable artifacts. Recursive self-improvement goes a level further and optimizes the improvement mechanisms themselves. The open difficulty, as the authors frame it, has two parts: adapting the self-improvement workflow as the agent's performance bottlenecks shift, and improving how those workflows are orchestrated using feedback from execution.

RSIFlow is a recursive self-improvement framework that combines adaptive workflow orchestration with evolution of a meta-agent's own capability. Rather than running one fixed improvement routine, it decides which routine to run based on where the agent is currently falling short.

The diagnosis draws on two sources: the trajectories of the task agent, and feedback from the environment. From these, RSIFlow identifies performance bottlenecks and dynamically selects improvement workflows specific to the component at fault. After a workflow runs, the meta-agent distills its outcome into two kinds of knowledge: skills specific to a component, and principles that cut across components. With those it evolves its own harness, so the next round of orchestration is better informed than the last.

Experiments span six benchmarks and two backbone models, and RSIFlow reaches the state of the art in average performance on all of the benchmarks. Ablation studies attribute contributions to both pieces: the adaptive workflow orchestration, and the evolution of the meta-harness. Code is publicly available.

There are two loops here, one inside the other. The inner loop improves the task agent, choosing a workflow to fit the current bottleneck. The outer loop improves the meta-agent doing the choosing, by turning each workflow's outcome into skills and principles it carries forward. The ablations are what tie the result to both loops rather than to either one alone.