llm-confirmation-bias-favors-memory

IN premise — summaries/2026/08/24/xu-2024-knowledge-conflicts-survey-s2-context-memory-conflict.md

Created 2026-08-25T02:59:02+00:00

Empirically, LLMs exhibit confirmation bias by favoring information consistent with their internal parametric memory over strong external contextual evidence (Chen et al., 2022; Xie et al., 2023).

Summary

When you feed an LLM new evidence that contradicts what it learned during training, it tends to trust its own training over the evidence you just gave it. This means you cannot assume the model will rationally update its answer based on context; it will quietly lean toward whatever its weights already encode, even when the prompt says otherwise.