The Illusion of Thinking
Controllable puzzles show reasoning models collapse to zero accuracy past a complexity threshold and reduce thinking effort as problems get harder.
Uses Tower of Hanoi, River Crossing and similar puzzles with tunable complexity. Three regimes appear, with non-reasoning models winning on easy tasks, reasoning models on medium, and both failing on hard. Went viral and prompted immediate methodological rebuttals.
- Date
- Saturday, 7 June 2025
- Lab
- Apple
- Kind
- paper
- Access
- paper only
Lead author Parshin Shojaee; Apple ML Research post dated June 2025. The result is contested. Rebuttals say output-token limits and impossible River Crossing instances explain much of the 'collapse'.
Sources
This record was checked against its sources on 6 October 2026. How we check