The Cost of Compression: A Rate-Distortion Limit on Factual Hallucination
Not provided in the abstract
Abstract
This paper explores the dual sources of error in factual hallucination during closed-book question answering, focusing on the impact of finite memory and compression on factual recall.
Reality Card
The study establishes a rate-distortion limit on factual hallucination, demonstrating that both compression distortion and missing coverage contribute to errors in factual recall.
The derived bound shows that the expected error in factual recall can be quantified as a function of observed facts and memory constraints.
The results are theoretical and may not fully capture the complexities of real-world language models, limiting direct applicability.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.