CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?
Jiaqi Su, Guibin Zhang, Jie-Jing Shao, Cong Pang, Tiankuo Yao, Ivor Tsang, Haiyan Yin, Shuicheng Yan
National University of Singapore · Agency for Science, Technology and Research (A*STAR) · SenseTime Group Inc.
Published October 18, 2026 · on the feed October 21, 2026 · ▲ 120 · arXiv 2610.25078
The abstract
Model--harness co-evolution has emerged as a prominent paradigm for recursive self-improving (RSI) agents, jointly evolving their execution scaffolds and intrinsic capabilities. We study the persistent yet often overlooked brain--body mismatch problem in this paradigm, where an evolving harness (body) demands behaviors its model (brain) has not learned or restricts capabilities it already possesses. We introduce CoEvo-RSI, a RSI framework that advances model--harness co-evolution through agent-driven, bidirectional historical inheritance. Drawing on a shared evolutionary archive and execution feedback, a meta-agent coordinates harness development from historical designs with model checkpoint selection or merging and subsequent adaptation under the selected harness, aligning inherited capabilities with evolving execution requirements to reduce brain--body mismatch. Using only supervised fine-tuning (SFT) on its own rejection-sampled trajectories, CoEvo-RSI makes Qwen3.5-9B competitive with larger models across diverse vertical applications, including multi-agent failure attribution, role-aware text-to-SQL, and smart-home instruction following. It surpasses DeepSeek-V4-Pro on TraceElephant (+14.16 agent accuracy) and GLM-5.3 on HomeBench VM (+4.90 success rate), underscoring the need for brain--body alignment and the potential of bidirectional historical inheritance to advance agent co-evolution.
The Brief
2:16A short podcast summary of the paper. Written and voiced by AI.
Transcript
CoEvo-RSI. When Harnesses Evolve, Which Models Should Follow?. By Jiaqi Su and seven others, from the National University of Singapore, the Agency for Science, Technology and Research, and one other institution. Published on the eighteenth of October. Its code repository is worked out: about two hundred source files, with a readme.
Model and harness co-evolution has become a prominent paradigm for recursive self-improving agents. Two things evolve together: the execution scaffold, or harness, and the intrinsic capabilities of the model running inside it. The authors study a problem in this paradigm that is persistent yet often overlooked, which they call brain-body mismatch. The evolving harness is the body. The model is the brain. And the body can change in ways the brain is not ready for. It may demand behaviors the model has never learned, or it may restrict capabilities the model already possesses.
CoEvo-RSI is a recursive self-improvement framework that advances co-evolution through agent-driven, bidirectional historical inheritance. Inheritance runs both ways: harness designs are drawn from history, and so are model checkpoints. That lets a new harness and a suitable model be settled together, rather than one being forced onto the other.
The work is coordinated by a meta-agent, which draws on a shared evolutionary archive and on execution feedback. It develops the harness from historical designs. Alongside that, it selects or merges model checkpoints, and then adapts the result under the harness it chose. The aim is to align inherited capabilities with the evolving execution requirements, which is exactly the gap the mismatch describes. Training uses only supervised fine-tuning, on trajectories the system rejection-samples from its own runs.
With that alone, Qwen3.5-9B becomes competitive with larger models across diverse vertical applications: multi-agent failure attribution, role-aware text-to-SQL, and smart-home instruction following. It surpasses DeepSeek-V4-Pro on TraceElephant, by fourteen point one six in agent accuracy. And it surpasses GLM-5.3 on HomeBench VM, by four point nine in success rate.
Upgrading a harness, on this account, is not free for the model inside it. Keeping brain and body matched is presented as its own requirement, and inheritance running in both directions is the tool offered for meeting it.
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
Brain and body evolve together. Here is what happens when they stop fitting.
You run a self-improving agent: each round you rewrite the harness around the model, its tools, formats and workflow, and then train the model to fit.
The road ahead
- 1 · The problem ~7 min When a self-improving agent keeps rewriting its harness and retraining its model, the two can drift apart until neither fits the other.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Keep every past harness and model checkpoint in an archive, and let each round draw on any of them, pairing each new harness with a model that suits it.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Each round, edit harnesses from archived designs, pick or merge checkpoints for them, fine-tune the best pair, then add everything to the archive.
- 6 · The experiment ~6 min You’ll guess: After co-evolution with CoEvo-RSI, how did Qwen3.5-9B compare with DeepSeek-V4-Pro on TraceElephant agent accuracy?
- 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.