Long reasoning is not deep reasoning
A reasoning chain that fills 8,000 tokens with restatements, hedging, and shallow exploration is not deeper than a 200-token chain that names the constraints, tries one approach, verifies. Length looks like effort but doesn't always correlate with accuracy.
Symptoms of theatre
- Reasoning that paraphrases the question for paragraphs before doing anything.
- Repeated "let me reconsider" with no actual reconsideration.
- Lists of considerations that don't influence the final answer.
- Branch counts that grow with budget but don't change conclusions.
What to do
- Lower the thinking budget. If accuracy holds, the budget was bloat.
- Prompt the reasoning structure: "name the constraints, then propose one approach, then verify."
- Switch to a non-reasoning model. Compare accuracy.