IC-1078LLaMA models (7B through 65B) exhibit gender bias in language generation, coreference resolution, and sentence likelihood, with stereotypical associations driving predictions

Tomasz Limisiewicz, David Mareček, Tomáš Musil

SourceDebiasing Algorithm through Model Adaptation

The authors evaluate all four LLaMA sizes on three bias measures: a linear model fitted to profession-prompt pronoun probabilities, Winobias coreference, and StereoSet sentence likelihood. In every size, the stereotypical coefficient (as) is positive and the intercept (b) is positive, indicating the models predict male pronouns more often and are influenced by stereotypical gender cues. Winobias shows a large gap between pro- and anti-stereotypical accuracy (Δs of 35.7–40.3%), and StereoSet ss scores of 68.4–71.9 indicate a strong preference for stereotypical completions over anti-stereotypical ones.

Evidence
correlational
Key metric
LLaMA 7B: as=0.235, af=0.320, b=0.072, r²=0.494; Winobias acc=59.1%, Δs=40.3%, Δg=3.0%; StereoSet lms=95.5, ss=71.9, icat=53.7. LLaMA 65B: as=0.249, af=0.316, b=0.095, r²=0.490; Winobias acc=73.3%, Δs=35.7%, Δg=1.4%; StereoSet lms=94.9, ss=69.5, icat=57.9
Caveat
The linear model fit is moderate (r² > 0.35 but well below 1), so the coefficients capture only part of the prediction variance. The profession set is English-only and the prompt templates were filtered to exclude highly gender-skewed verbs.
Model
LLaMA
Concepts
Shortcut
Datasets
Winobias [eval], StereoSet [eval], Professions (Bolukbasi et al.) [eval]
Methods
Causal Tracing [supporting]
Related work
Causal mediation analysis / Vig et al. 2020 (causal mediation analysis) [builds-on]
Related findings
IC-1079
Extraction
automatic-extraction