RSIHub: A Composable Framework for Agent-Evolution Methods
Zimu Wang, Xiaobo Wang, Siyu Ye, Jiaqi Li, Silin Chen, Yuling Shi, Ruobing Wang, Jingjing Zhang and 6 others
University of California, Berkeley · University of Science and Technology of China · Douyin Group · Shanghai Jiao Tong University · Tongji University · Beijing Institute for General Artificial Intelligence
Published October 23, 2026 · on the feed October 26, 2026 · ▲ 49 · arXiv 2610.32253
The abstract
Agent-evolution methods use execution feedback to improve prompts, skills, tools, and orchestration, but their implementations often rebuild the same experimental machinery. We present RSIHub, an open-source framework for building, comparing, and extending these methods on shared infrastructure. Recipes compose reusable operators and declare mutable resources; RSIHub supplies execution, evaluation, recovery, and lineage. Researchers can focus on evolution policy while evaluation stays outside the mutable surface and results remain tied to candidates and evaluation conditions. We instantiate AHE, Hyperagents, A-Evolve, and GEPA without modifying the framework core, keeping method-specific logic in recipes and operators. We evaluate them with MiniSWE and Codex on Terminal Bench 2 and τ^3 Banking under fixed model and evaluator configurations. The largest observed overall pass@1 gain is 13.5 percentage points, while method rankings vary across settings and optimization gains do not always transfer to held-out tasks. RSIHub provides a common protocol for comparing evolution policies, inspectable candidate histories, and reusable components for developing new methods.
The Brief
2:19A short podcast summary of the paper. Written and voiced by AI.
Transcript
RSIHub. A Composable Framework for Agent-Evolution Methods. By Zimu Wang and thirteen others, from the University of California, Berkeley, the University of Science and Technology of China, and five other institutions. Published on the twenty-third of October. Its code repository is worked out: about three hundred and fifty source files, with a readme and a licence.
Agent-evolution methods use execution feedback to improve an agent's prompts, skills, tools and orchestration. Their implementations, though, often rebuild the same experimental machinery from scratch. Each new method pays again for execution, evaluation and bookkeeping that have nothing to do with its own idea. When each method arrives with its own machinery, comparing methods on equal terms is hard, and so is building on them.
RSIHub is an open-source framework for building, comparing and extending these methods on shared infrastructure. What it lets researchers focus on is the evolution policy itself. Methods are evaluated under fixed model and evaluator configurations, and results remain tied to the candidates and the evaluation conditions that produced them.
A method is written as a recipe. Recipes compose reusable operators and declare which resources are mutable. RSIHub supplies execution, evaluation, recovery and lineage. Evaluation stays outside the mutable surface: whatever evolves, the measurement does not. To show the design holds, the authors instantiate four existing methods, AHE, Hyperagents, A-Evolve and GEPA, without modifying the framework core, keeping each method's specific logic in its recipes and operators.
The four are run with MiniSWE and with Codex, on Terminal Bench 2 and on tau cubed Banking. The largest observed overall pass at one gain is thirteen point five percentage points. Two other results temper that. Method rankings vary across settings, and optimization gains do not always transfer to held-out tasks. Put another way, which method looks best depends on where it is tested, and improvement on the tasks a method evolves against is no guarantee of improvement elsewhere.
Thirteen point five percentage points is the largest gain any method posts, yet the ranking of methods behind that number shifts with the setting. Seeing that takes the methods running side by side under the same conditions, which is the job RSIHub was built for. What it leaves behind is a common protocol for comparing evolution policies, inspectable candidate histories, and reusable components for whoever develops the next method.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every new idea for self-improving agents starts by rebuilding the same plumbing.
You have an idea for a better way to let an agent improve its own prompts and tools from feedback. Before testing it, you must build workspaces, runners, evaluation, crash recovery and record keeping, again.
The road ahead
- 1 · The problem ~7 min Every new agent-evolution idea forces its authors to rebuild the same testing machinery, and that makes methods hard to compare and gains hard to trust.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Fix the shared lifecycle slots, let each method fill them and declare what may change, and keep evaluation and records under framework control.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Write a method as a recipe of operators with a declared surface, run and score candidates on shared infrastructure, and tie every result to its conditions.
- 6 · The experiment ~6 min You’ll guess: Across four agent-evolution methods run in RSIHub, how big was the largest overall pass@1 gain?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.