A new paper on arXiv investigates the relationship between planning and error recognition in Transformer models. The authors note that being able to plan does not automatically mean a model can spot its own mistakes.
However, they find that error recognition can improve planning capacity. The research aims to understand the internal mechanisms behind this phenomenon, following a mechanistic approach.
The abstract is brief, so specific results and methods are not detailed here. The study appears to focus on how Transformers handle planning tasks and whether error detection plays a role in improving performance.