chameleon-gpt4-vicuna-resist-counter
IN premise — summaries/2026/08/24/xie-2024-chameleon-sloth-s0-abstract-chunk-2.md
Created 2026-08-25T02:58:58+00:00
GPT-4 and Vicuna-33B resist counter-memory most strongly, still predominantly following their parametric memory even at a 1:3 evidence ratio favoring the counter-answer.
Summary
GPT-4 and Vicuna-33B are stubborn in a specific way: even when you feed them three times as much evidence pointing to a different answer, they mostly stick with whatever they absorbed during training. This means their baked-in knowledge acts as a heavy default that's hard to override through prompting alone, so a system relying on them to update in response to new data should expect pushback rather than clean revision.