Nace AI has released details on their Drex family of decision models, which are designed to process states and questions to output probability distributions without autoregressive decoding. The models, including Drex DLM, v1.5, v2.0, and Max, utilize various architectures like diffusion LMs, Gated DeltaNets, and sparse MoE. Nace AI employed techniques such as Reinforcement Learning from Adjusted Feedback (RLAF) and consistency training to enhance performance, achieving high scores on benchmarks like the Decision Index and demonstrating low calibration error. The Drex models are available for local use and offer competitive latency and context window sizes compared to other leading models. AI
IMPACT Provides a detailed playbook for training specialized decision models, potentially influencing agent development and tool selection.
RANK_REASON The cluster details a technical report and associated models from a research lab, including novel training techniques and evaluation results.
- Decision Models
- Drex
- Nace AI Research Lab
- Reinforcement Learning from Adjusted Feedback (RLAF)
- Sota
- ArtificialAnalysis
- Drex DLM
- Drex Max
- Drex v1.5
- Drex v2.0
- Nace AI
- OpenRouter
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
