Rohan Paul (@rohanpaul_ai)T2·78 pts
- Drex Max emits no text, only one probability per supplied option in a single forward pass
- RLAF training rewards verified answers plus a reasoning judge's probabilities, with a KL penalty limiting drift from supervised training
- $0.06 and 183 ms per 1K decisions, but only 21% on sampled HLE without chain-of-thought