RealFeedApp Large language models systematically overuse epanorthosis
The study finds that large language models systematically employ epanorthosis, a rhetorical figure of self-correction, due to training data rich in promotional prose and preference tuning via RLHF. The authors propose a mitigation program based on this finding.
- Field
- training language models
- What they did
- Researchers discovered that large language models systematically overuse the rhetorical figure of epanorthosis (self-correction like 'This is not a course, it is a journey of transformation'), and found that the cause lies in training data rich in promotional text and settings that reward confident phrasing.
- Why it matters
- This knowledge will help developers adjust training settings to reduce the unnatural style of model responses, making them sound more naturally and less like advertisements.
#llm#epanorthosis#rlhf#rhetoric#language model
