Truncated Reasoning Under a Budget: Evaluating System-1 Models as Decision-Makers
When a reasoning model is cut off mid-thought, a System-1 decision model can read the partial trace and answer. At a 1,500-token budget it beats forced answering by 4.1 pp, and a learned stopping policy saves 71.3% of the full-run cost for a 0.58 pp accuracy loss.