Budget forcing intervenes in the token stream instead of the weights. To cap thinking, the decoder appends the end-of-thinking delimiter once a token budget is reached, forcing the model to answer from whatever reasoning it has produced. To extend thinking, the decoder suppresses that delimiter when the model tries to stop and appends a continuation cue such as “Wait”, which reliably triggers another round of self-checking.
The s1 paper introduced the technique and showed accuracy scaling smoothly with the enforced budget, no retraining required. It is the blunt, guaranteed counterpart to learned reasoning-effort dials, and some models (Nemotron, Qwen3) are trained or found to be robust to being cut off mid-thought so the forced answer stays good.