IC-1134LLMs show strong confirmation bias in multi-source settings, preferring evidence consistent with parametric memory, with stronger bias for popular entities

Jian Xie, Kai Zhang, Jiangjie Chen, Renze Lou, Yu Su

SourceAdaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts

When both parametric memory and counter-memory are presented as evidence, LLMs systematically prefer the evidence consistent with their parametric memory. GPT-4 shows an 80% memorization ratio for the most popular PopQA questions, indicating stronger confirmation bias for well-known facts. The memorization ratio increases monotonically with the proportion of parametric memory evidence, from near 0% at 0/2 to near 100% at 2/2, but even at 1/2 (50/50) the ratio exceeds 50% for most models, confirming a genuine bias toward parametric memory.

Evidence
correlational
Key metric
GPT-4 memorization ratio 80% for most popular questions; at 1/2 proportion: ChatGPT 43.0, GPT-4 65.4, PaLM2 56.8, Qwen-7B 52.3, LLaMA2-7B 58.4, LLaMA2-70B 58.4, Vicuna-7B 45.9, Vicuna-33B 51.7
Caveat
The counter-memory is LLM-generated and may not represent all forms of conflicting evidence; experiments use zero-shot setting.
Model
ChatGPT, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, PaLM 2, Qwen Qwen-7B, Llama 2 / Llama 2 base Llama 2 7B, Llama 2 70B, Vicuna Vicuna-7B, Vicuna-33B
Datasets
PopQA [eval], StrategyQA [eval]
Related findings
IC-1133, IC-1135, IC-1136
Extraction
automatic-extraction