Get Started
Home
Topics
Search
Library
Research questionHow do LLMs integrate external evidence into partially formed answers, and why can the same evidence help or harm?Evidence from tools, retrieval systems, other agents, or users can alter an LLM’s existing answer rather than simply add information. The effect depends on the receiving model and can persist even when it has internally verified the supplied candidate as invalid.
AI
Mechanistic Interpretability
Natural Language Processing
Reasoning
Retrieval-Augmented Generation
Latest papersRecent research connected to this question, newest first.Evidence Integration in Large Language ModelsThis concerns LLM decisions involving externally supplied evidence from tools, retrieval-augmented generation, other agents, or users. The evidence covers more than ten million trials across twelve LLMs from four families and eight domains, including quantum mechanics, physics, genetics, and molecular biology; mechanistic claims are based on causal interventions and representation analyses of the tested models.research paper · Sep 3, 2026
Related questions
How can long-term LLM agents reconcile evolving textual evidence across interactions?How can we evaluate LLM reasoning quality beyond final-answer accuracy across deployment contexts?How can LLMs preserve evidence-based financial judgments despite personalized user context?How can we evaluate LLM knowledge updates over time without contamination or inconsistent facts?