A post-00s master's student unveils an AI inference "black technology": with just two tokens, the base model matches the reinforcement learning version.

What happened: An MIT-led study found that fixing the first two tokens of a base model's answer can bring its math reasoning close to its reinforcement learning version, without changing parameters or adding examples. Key numbers: Olmo-3-7B accuracy on MATH-500 rose from 42% to 78% with the ".\n\nOkay" prefix, above its RL version's 75%. Why it matters: The paper suggests part of RL's effect may be making models pick effective answer openings, and that openings also affect refusal and safety behavior. The authors note results are limited to the tested models.

Two tokens let a base model match reinforcement learning, signaling major AI inference efficiency gains.

Two tokens let a base model match reinforcement learning, signaling major AI inference efficiency gains. — impact context not separately stored; see What changed.