Drex Open Decision Models Achieve State of the Art
Original title: Towards Foundation Decision Models: Drex
Drex family of decision models including 8B to 35B parameters sets new benchmarks in accuracy and speed over autoregressive LLMs.
Autoregressive LLMs suffer from inconsistencies, parsing needs, and high latency. Decision models output probability distributions over user-supplied options without decoding. Drex trains models with supervised fine-tuning and reinforcement learning from adjusted feedback, plus vision tuning for the largest variant. Max reaches 67.52 on Decision Index 0.4.1 and answers requests in 183 ms.
- Drex models: 8B, 9B, 35B active, 27B parameters
- RLAF improves accuracy and robustness
- Max sets new SOTA on Decision Index
- 183 ms per request vs 2.0 s for GPT-6 Sol
- All models open source
- Organizations
- nace-ai
From the original article
1 min read at huggingface.co
Autoregressive approaches with large language models, such as long chain-of-thought reasoning, have become the default for new AI applications because they reach higher accuracy on a wide range of tasks. In production, however, they give inconsistent answers, require parsing of outputs, and add cost and latency.
By Nischay Dhankhar, Alisher Amantay, Zhanibek Datbayev · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.