Today’s LLMs are extremely capable, but they still often misread unfamiliar situations, simply because of how they’re trained.

Belletrist shows that fine-tuning on reasoning reconstructed from humanistic texts can significantly improve model judgment while preserving general capability.