IC-1148ChatGPT and InstructGPT exhibit positional bias in their reliance on in-context examples, with ChatGPT showing decreasing attention by position and InstructGPT showing a U-shaped pattern

Linyi Yang, Shuibai Zhang, Zhuohao Yu, Guangsheng Bao, Yidong Wang, Jindong Wang, Ruochen Xu, Wei Ye, Xing Xie, Weizhu Chen, Yue Zhang

SourceSupervised Knowledge Makes Large Language Models Better In-context Learners

Using an interpreter prompt that asks the model to identify which in-context example most influenced its prediction, the paper counts how often each of 16 positions is cited across 8 NLU tasks. ChatGPT shows a monotonic decrease: the second example is cited over 35,000 times while the last receives around 5,000. InstructGPT shows a U-shaped curve, attending to the first and last examples while attention degrades significantly in the middle. The authors note this is consistent with Liu et al. (2023) on long-context degradation.

Evidence
correlational
Key metric
ChatGPT: 2nd in-context example cited over 35,000 times, last example around 5,000 times; InstructGPT: U-shaped occurrence with peak at positions 1 and 16, trough in the middle
Caveat
The measurement relies on the model's self-reported identification of the influential example via an interpreter prompt, which may not perfectly reflect actual internal attention.
Model
ChatGPT, InstructGPT
Concepts
Positional bias
Datasets
GLUE-X [eval]
Related work
Lost in the middle: how language models use long contexts [context]
Related findings
IC-1149, IC-1150
Extraction
automatic-extraction