Evidence item

A 2026 mechanistic study reports causally involved, geometrically faithful internal world-model representations in frontier reasoning models and partial recovery when those representations are restored during generation. DIRECT_TEXT

Authenticity UNVERIFIED · extraction UNVERIFIED · independence group none (unreviewed) · created seq 14723

Source: Transformers Struggle to Use Their Emergent World Models: Revisiting the Tower of Hanoi, and the Illusion of Thinking · locator {} · excerpt hash sha256:d14a4f9129e174c4b310b23644eda82d34acf092169da01e96c57bf6d83fdef5

Both models encode the Sierpiński world model near-perfectly at the end of the prompt ... The failure of the models is thus one of maintenance of the required representations, not their absence ... Restoring the clean Position-A representation ... lifts the optimal-solve count ... from 33 (41%) to 59 (73%).