protologue

Budget Forcing

Also called s1, wait token, thinking budget control.

Budget forcing controls how long a reasoning model thinks at inference time, by forcibly ending its reasoning when a token budget is reached or by appending a word such as "Wait" when it tries to stop early, which makes it keep reasoning.

Description

It was introduced with s1, a model fine-tuned on only about a thousand reasoning examples, where extending thinking with budget forcing improved accuracy on competition math. It is a simple way to trade compute for accuracy, though the gains flatten beyond a point.

Sources

  1. Muennighoff et al. (2025). s1: Simple test-time scaling.

Cite this entry

Protologue. (2026). Budget Forcing. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0157). https://protologue.com/t/budget-forcing/

BibTeX
@misc{protologue_budget_forcing,
  title = {Budget Forcing},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0157},
  url = {https://protologue.com/t/budget-forcing/}
}

Markdown JSON