First Comprehensive Playbook and Open-Source Models for SOTA Decision-Making AI
Nace AI Research Lab has published a 50-page technical report alongside open-source models detailing the complete architecture, datasets, and training methodologies for their Drex decision models. The release covers innovations like Reinforcement Learning from Adjusted Feedback (RLAF), confidence calibration, and option shuffling. This release addresses a major knowledge gap in the AI community by moving beyond benchmark results to openly document how state-of-the-art decision-making models are trained. It provides practical resources for developers experimenting with local agent routers, tool selection, and specialized decision frameworks. The Drex models range from 9B to 35B parameters, utilizing architectures like Dense Causal Transformers and Mixture of Experts, and achieve high accuracy alongside sub-100ms latency on long contexts up to 128K tokens. Ablation studies revealed that most performance gains stemmed from Supervised Fine-Tuning (SFT), while RLAF primarily improved confidence calibration.
## BACKGROUND
Decision models in AI are specialized systems designed to evaluate complex states, weigh alternatives, and select optimal actions or choices rather than just generating general text. Techniques like Supervised Fine-Tuning and Reinforcement Learning are standard approaches used to align model outputs with desired performance metrics and verified logic.