counter-memory-length-ratio-1-2-threshold
IN premise — summaries/2026/08/24/xie-2024-chameleon-sloth-sR-references.md
Created 2026-08-25T02:59:01+00:00
When counter-memory length ratio exceeds 1.2 relative to parametric memory, the LLM's answer distribution shifts significantly.
Summary
When the external context fed to a model is more than 20% longer than the internal knowledge it would otherwise draw on, the model's outputs stop reliably reflecting what it was trained to know and instead get heavily steered by that context. This matters because it marks a point where you can no longer assume the answer comes from the model's own understanding, which changes how much trust you should place in its responses and which safeguards you need to activate.