Jeeves: Improving Decision Models via Reasoning-Before-Deciding
September 29, 2026
Jeeves is a 9B Qwen3.5 model that uses a diffusion drafter and CISPO to reason before making decisions. It achieves 0.889 accuracy on unseen test data, outperforming standard Jev-like models while maintaining reasonable latency on H100 GPUs.
HOW THIS AFFECTS YOU
●
builderYou can integrate reasoning steps into decision-making models to improve accuracy on hard tasks.
●
researcherThis highlights the performance gains from combining diffusion drafting with reasoning-heavy training.