Announcement
Reported study says some base models reason better with specific first-token cues
In posts describing new research, the authors say some base models recovered much of reinforcement learning’s reasoning gain when prompted with model-specific “token cues,” and they say the work also includes a safety case study on refusal and compliance.
TLDR
In posts about new research, the authors say some base models can recover much of reinforcement learning’s reported reasoning gain when responses begin with model-specific “token cues.” They shared examples including Olmo-3-7B rising from 42% to 78% on MATH-500 with ".\n\n Okay" and Qwen3-14B rising from 72% to 87% with " Alright," and said the work also includes a safety case study on refusal and compliance.
Combined views
125.3K
15 Sources, first seen ago
