LLMs systematically overuse rhetorical self-correction, new study finds

A new research paper identifies a rhetorical figure called epanorthosis, characterized by self-correction, as being systematically overused by large language models. The authors attribute this overreliance to training data rich in promotional content and reinforcement learning techniques that favor emphatic phrasing. The study proposes an "Epanorthosis Index" to measure this figure against human baselines across different genres, finding that models over-index in oratory and under-index in informal Q&A, while aligning with human rates in argument, journalism, and encyclopedic prose. Mitigation strategies, including lightweight adapters and instruction tuning, are presented to calibrate model output closer to human rhetorical styles. AI

IMPACT Identifies a specific linguistic artifact in LLMs, potentially impacting the naturalness and trustworthiness of AI-generated text.

RANK_REASON Academic paper detailing a specific linguistic phenomenon in LLMs and proposing mitigation strategies. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Cicero
• epanorthosis
• Fontanier
• large language models
• LoRA
• Quintilian
• RLHF

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →