Do large language models perform genuine reasoning beyond interpolation?

Determine whether large language models trained via next-token prediction genuinely perform reasoning beyond interpolation of training data, by formulating precise criteria for reasoning and establishing whether such models satisfy these criteria.

Background

In the concluding discussion, the authors highlight a central epistemic uncertainty about the nature of LLM capabilities: whether these models merely interpolate from training data or exhibit genuine reasoning.

This question is framed as essential for understanding the structure and limits of “optimal” networks and aligns with broader theoretical goals at the intersection of learning theory, cognition, and AI capabilities.