IC-1135LLMs show order sensitivity to evidence position in context, with PaLM2 and LLaMA2-7B showing memorization ratio variations exceeding 30%

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 parametric memory and counter-memory are presented in different orders, the memorization ratio changes significantly for most models. PaLM2 varies from 38.6 (parametric memory first) to 72.2 (counter-memory first), a 33.6 percentage point swing. LLaMA2-7B varies from 33.3 to 82.8, a 49.5 percentage point swing. ChatGPT tends to favor the first evidence, while PaLM2 and LLaMA2-7B lean toward later evidence. GPT-4 shows minimal order sensitivity.

Evidence
correlational
Key metric
PaLM2: 38.6 (parametric first) vs 72.2 (counter first); LLaMA2-7B: 33.3 vs 82.8; ChatGPT: 46.7 vs 40.1; GPT-4: 60.9 vs 62.7
Caveat
Only four models (ChatGPT, GPT-4, PaLM2, LLaMA2-7B) were tested for order sensitivity; the order of evidence was the only variable changed.
Model
ChatGPT, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, PaLM 2, Llama 2 / Llama 2 base Llama 2 7B
Concepts
Positional bias
Datasets
PopQA [eval], StrategyQA [eval]
Related work
Behnamghader et al. 2022 (retriever-augmented LM reasoning) [context]
Related findings
IC-1133, IC-1134, IC-1136
Extraction
automatic-extraction