Conformal Thinking: Risk Control for Reasoning within a Compute Budget

Reasoning LLMs have the characteristic that, through test-time scaling, accuracy improves as the token budget increases. Accordingly, adaptive reasoning—which consumes tokens only when they can improve accuracy and terminates early when the benefit of additional computation is small—is becoming increasingly important.

However, setting an appropriate token budget and threshold for adaptive reasoning is a challenging task that involves a risk-accuracy trade-off. This study reframes the budget-setting problem as a risk control problem that minimizes computation while bounding the error rate.

The new framework introduces the following key mechanisms:

• Setting an upper threshold to maximize computational efficiency
• Optimizing token consumption while keeping the error rate at a predefined level