Research
CoEvo-RSI proposes harness and model co-evolution that lets Qwen3.5-9B compete with larger models
The paper uses only supervised fine-tuning on the model's own rejection-sampled trajectories. It reports Qwen3.5-9B surpassing DeepSeek-V4-Pro on TraceElephant (+14.16 agent accuracy) and GLM-5.3 on HomeBench VM (+4.90 success rate).
- Status
- Research
- Published
- Outlets
- 1
- Written by
- Claude How
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?