IC-1154LLaMA-65B exhibits a selection bias where zero-shot accuracy varies substantially by answer choice, with 'a' at 71.58% and 'd' at 52.28%

Jianhao Yan, Jin Xu, Chiyu Song, Chenming Wu, Yafu Li, Yue Zhang

SourceUnderstanding In-Context Learning from Repetitions

On a balanced MMLU test set (equal numbers of questions with golden answers a, b, c, d), LLaMA-65B's zero-shot accuracy differs by answer choice: class 'a' achieves 71.58%, 'b' 59.65%, 'c' 58.95%, and 'd' 52.28%. With 5 demonstrations, the overall accuracy improves from 60.61% to 63.16%, but class 'a' accuracy drops to 54.74% while 'b', 'c', and 'd' all increase. This indicates a persistent preference for earlier answer choices in zero-shot that partially corrects with more demonstrations.

Evidence
correlational
Key metric
Zero-shot: a=71.58%, b=59.65%, c=58.95%, d=52.28%, avg=60.61%. With 5 demos: a=54.74%, b=70.76%, c=67.60%, d=59.53%, avg=63.16%.
Caveat
Measured on 1140 randomly sampled MMLU test cases with the test set randomly permuted to ensure balance across answer choices. The paper presents this as a confound to be controlled for in the spurious-connection experiments rather than as a primary contribution.
Model
LLaMA
Datasets
MMLU / MMLU-Math [eval]
Related findings
IC-1151, IC-1152, IC-1153
Extraction
automatic-extraction