Artificial Epanorthosis: Why large language models overuse a classical rhetorical figure, and how to mitigate it

En 30 secondes

Ce qui s'est passé

LLMs systematically overuse epanorthosis, a classical self-correction rhetorical figure, driven by training data and preference tuning rather than generation mechanics.

Pourquoi ça compte

Matters for engineers building or fine-tuning language models who want output matching human rhetorical patterns across different writing genres and registers.

Vigilance

The paper measures only one model family; mitigation techniques like LoRA adapters and instruction tuning show promise but require genre-specific calibration rather than outright elimination.