Mechanism behind improvements from test-time cognitive structure guidance

Determine whether test-time cognitive structure guidance that scaffolds consensus reasoning patterns causes large language models to deploy latent reasoning capabilities or merely retrieve cached reasoning patterns from training data.

Background

The paper proposes test-time reasoning guidance that scaffolds cognitive structures associated with success and reports performance gains, especially on ill-structured problems.

Despite these improvements, the authors state that the underlying mechanism is unresolved: whether guidance elicits genuinely latent capabilities or simply triggers retrieval of memorized patterns remains unclear.