New research links fine-tuning to LLM hallucinations, proposes solutions

A new research paper explores how supervised fine-tuning (SFT) of large language models can inadvertently increase hallucinations, which are factually incorrect statements. The study proposes that this issue stems from knowledge degradation and interference among semantic representations. To address this, the researchers suggest using continual learning techniques, such as self-distillation, to improve factual learning while minimizing hallucinations. They also found that freezing certain model parameters can reduce hallucinations when new knowledge acquisition is not required. AI

IMPACT This research could lead to more reliable LLMs by reducing factual inaccuracies introduced during fine-tuning.

RANK_REASON Research paper published on arXiv detailing a novel method to address LLM hallucinations. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Guy Kaplan
• hallucinations
• Hugging Face
• Influence Flower
• Large language models
• ScienceCast
• supervised fine-tuning

• paper
• safety

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