New strategy cuts LLM hallucinations to under 5% using confidence guidance

Researchers have developed ConfQA, a fine-tuning strategy designed to significantly reduce hallucinations in large language models (LLMs). By training models to respond with "I am unsure" when they lack confidence, hallucination rates have been lowered from 20-40% to below 5% across various factuality benchmarks. This approach uses a dampening prompt and training data derived from factual statements to improve model confidence calibration. Building on ConfQA, ConfRAG is introduced as a retrieval-augmented generation strategy that only triggers external retrievals when the model indicates uncertainty, achieving over 95% accuracy while cutting unnecessary retrievals by more than 30%. AI

IMPACT This research could lead to more reliable and efficient LLM applications by reducing hallucinations and optimizing retrieval processes.

RANK_REASON The cluster contains an academic paper detailing a new method for improving LLM factuality and retrieval efficiency. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• ConfQA
• ConfRAG
• large-language models
• retrieval-augmented generation
• Yin Huang

• paper
• model release

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