Budget Forcing: Appending "Wait" Tokens at Reasoning Model's End-of-Thought Boundary Extends Deliberation Without Fine-Tuning

Introl Blog·medium signal

Budget forcing, formalized in the s1 paper, controls test-time compute by intercepting the end-of-thinking token during inference and replacing it with a "Wait" token to force continued deliberation — or stripping it early to truncate reasoning for latency-sensitive calls. With 1,000 carefully curated training examples plus budget forcing at inference time, s1-32B matched or beat OpenAI o1-preview on math and science benchmarks. Practically: add a token-level hook in your inference pipeline that monitors for the model's internal reasoning stop sequence and injects continuations until a per-request token budget is consumed.