Jeeves. Reasoning improves Jev-like decision models (github.com)

A new AI model named Jeeves has been introduced, successfully enhancing decision-making in Jev-like classifiers. This model employs a reasoning framework combined with a diffusion drafter, utilizing techniques from Stochastic Fine-Tuning (SFT) and Contrastive In Situ Policy Optimization (CISPO). The Jeeves model, built on the Qwen3.5-9B architecture, demonstrates a significant improvement in accuracy, achieving a score of 0.889 on out-of-domain test data—surpassing the performance of existing models Kev-9B (0.822) and Jev (0.857). Moreover, it excels in JevBench benchmarks, scoring 0.935 compared to Jev's 0.866. The significance of Jeeves lies in its ability to reason before making decisions, which enhances performance on complex, real-world tasks. It supports multiple types of queries such as yes/no, multiple choice, and ratings within the same request through a Jev-compatible API. Notably, the model operates efficiently on a single H100 GPU, processing requests in approximately 0.3 seconds without reasoning, and averaging 3.3 seconds with reasoning activated. Its architecture combines advanced techniques like LoRA and the pointer head approach, calibrated for better decision probability outputs, marking a substantial step forward for AI/ML researchers focusing on decision-making frameworks.