LLMs can detect own hallucinations using confidence signals without labels

Researchers have developed a novel method for large language models to identify and abstain from answering questions they are uncertain about, without requiring labeled datasets. This label-free approach leverages the model's internal confidence signals, which tend to decrease when the model generates incorrect information. By fine-tuning models to abstain when their confidence is low, the method performs comparably to traditional supervised abstention-tuning techniques across various open-weights models. This technique offers a nearly free substitute for costly labeled datasets in teaching models when to admit uncertainty. AI

IMPACT This research offers a cost-effective way to improve LLM reliability by enabling models to self-identify and abstain from answering uncertain queries.

RANK_REASON The item is an academic paper detailing a new method for LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• judge model
• large-language models
• LoRA+
• open-weights models
• ScienceCast

• paper
• model release

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