LLMs exhibit 'anchored confabulation,' amplifying confident hallucinations with partial evidence

Researchers have identified a new phenomenon in large language models called "anchored confabulation," where providing partial evidence can paradoxically increase the model's tendency to confidently hallucinate. This effect, formalized as Parametric Hallucination Confidence (PHC), was observed across multiple model families and predicted by a new law, the Anchoring Threshold Law. The findings suggest implications for retrieval-augmented generation (RAG) systems, with a proposed LearnedRouter demonstrating significant performance gains by exploiting PHC. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Identifies a new LLM failure mode that can be exploited to improve RAG systems and reduce hallucinations.

RANK_REASON Academic paper detailing a novel LLM behavior and its implications.

Read on arXiv cs.CL →

• arXiv
• Ashish Balkishan Lathkar
• Parametric Hallucination Confidence
• Anchoring Threshold Law
• RAG
• LearnedRouter
• Hugging Face

• paper
• safety