chameleon-quantity-amplifies-bias
IN premise — summaries/2026/08/24/xie-2024-chameleon-sloth-s0-abstract-chunk-2.md
Created 2026-08-25T02:58:59+00:00
Confirmation bias in LLMs grows as the absolute count of parametric-memory evidence increases, even when the supporting-to-contradicting ratio stays the same (e.g., 1:2 vs. 2:4).
Summary
It's not just about whether the evidence leans one way or the other; the sheer volume of stored evidence itself makes the model more likely to gravitate toward its prior expectations, even when the supporting and contradicting pieces scale up proportionally. This means that feeding an LLM more context doesn't guarantee a more balanced judgment — larger evidence sets can actively worsen its tendency to see what it already expects to see.