xu2024-multi-turn-belief-alteration-range
IN premise — summaries/2026/08/24/xu-2024-knowledge-conflicts-survey-sR-references-chunk-2.md
Created 2026-08-25T02:59:04+00:00
Multi-turn belief alteration in LLMs ranges from 20.7% to 78.2% cumulative shift as counter-memory rounds increase (Xu et al. 2023, as reported in the survey).
Summary
Repeatedly feeding an LLM counter-evidence over multiple turns can shift its answers by anywhere from roughly one-fifth to more than three-quarters, meaning its apparent knowledge is far more malleable in conversation than it appears in a single query. This matters because any system that treats LLM responses as stable, settled facts is exposed to significant drift whenever a user or an attacker keeps pushing back on what the model previously stated.