CoEvo-RSI aligns evolving agent harnesses with model checkpoints
Original title: CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?
Authors say SFT on self-sampled trajectories makes Qwen3.5-9B beat larger models on two agent benchmarks.
CoEvo-RSI is a recursive self-improvement framework that jointly evolves agent harnesses and model checkpoints to reduce brain-body mismatch. A meta-agent uses a shared evolutionary archive and execution feedback to inherit harness designs and select or merge checkpoints, then adapts with supervised fine-tuning on rejection-sampled trajectories. The authors report Qwen3.5-9B becoming competitive on failure attribution, role-aware text-to-SQL, and smart-home instruction following.
Unconfirmed: Benchmark gains are the authors' reported results and are not independently confirmed.
- Training uses only supervised fine-tuning on the agent's own trajectories.
- Reported TraceElephant gain is +14.16 agent accuracy over DeepSeek-V4-Pro.
- Reported HomeBench VM gain is +4.90 success rate over GLM-5.3.
- Tasks include multi-agent failure attribution and role-aware text-to-SQL.
- Products
- CoEvo-RSI
- AI models
- Qwen3.5-9BDeepSeek-V4-ProGLM-5.3
From the original article
1 min read at huggingface.co
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.
By Jiaqi Su, Guibin Zhang, Jie-Jing Shao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.