1 paper
Pierre-Luc St-Charles, Alessandro Palmas, Damiano Fornasiere +5
Reasoning models can remain capable of solving a task while still defaulting to cheaper but misleading shortcuts. This creates a central oversight problem: when a model gives an an…