When to Think Deeply: Inhibitory Deliberation for LLM Reasoning

In three linesIDPR is a framework that dynamically decides when to invoke slow reasoning in LLMs. An inhibition controller analyzes the fast response (confidence, logit margin, cost) and suppresses or validates before slow reasoning. On 5000 math examples, IDPR uses slow reasoning on 8.20% of cases and improves accuracy from 47.90% to 48.92%.

Summary generated by Claude — human-verified