LLMs' source attribution causally linked to cached state, study finds

Researchers have investigated how large language models attribute their answers to either provided context or their internal knowledge. Experiments involving interventions on the model's key-value (KV) cache revealed that modifying the document's cache had no impact on source reporting. However, transplanting the question-and-answer (Q+A) cache between runs significantly influenced the reported source, with transplanted Q+A caches dictating the reported origin of the answer. This suggests that the cached state produced during answer generation causally determines source reporting in these models. AI

IMPACT This research clarifies how LLMs determine answer sources, potentially impacting how their outputs are trusted and audited.

RANK_REASON Academic paper detailing novel research findings on LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• KV cache
• LLMs
• Q+A cache

• paper
• model release

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