IC-1133LLMs are highly receptive to coherent counter-memory as sole evidence, contradicting prior findings of stubbornness with entity-substitution counter-memory

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 counter-memory is the only external evidence presented, LLMs adopt the counter-answer at high rates if the counter-memory is generated coherently by an LLM, in contrast to the low adoption rates observed with incoherent entity-substitution counter-memory. The authors manually inspected 50 stubborn cases and found most were due to hard-to-override commonsense or lack of strong direct conflict, confirming that the high receptiveness is genuine. This contradicts the prior conclusion from Longpre et al. (2021) that LLMs are stubborn and cling to parametric memory.

Evidence
correlational
Key metric
Manually checked 50 stubborn (mem-ans.) cases; most attributed to hard-to-override commonsense or lack of strong direct conflicts
Caveat
The counter-memory is generated by ChatGPT and may not represent all forms of coherent disinformation; the finding is limited to the PopQA and StrategyQA question formats.
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-1134, IC-1135, IC-1136
Extraction
automatic-extraction