Does Alignment Debias or Just Suppress? Evaluating Stereotypical Bias Across Base-Instruct Pairs
Year: 2026
Type: conference-paper
Abstract: Warning: This paper contains offensive stereotypes about gender, race, and religion.Alignment has been widely reported to reduce stereotypical bias in large language models, but it is unclear whether these reductions reflect debiasing or suppression at the output level.Across 27 base-instruct pairs from four open-weight families and four bias benchmarks, three patterns provide substantial evidence that the measured reduction is largely suppression.(i) The chat template drives about half the alignment effect on CrowS-Pairs and essentially all on StereoSet, where 18 of 27 improve without it (p = 0.061) versus 26 of 27 with it (p < 0.001).Per-pair effects remain below d = 0.2 in both conditions.IAT lies at the weight-level end of the spectrum, with template contribution indistinguishable from zero.(ii) Decomposing BBQ's ambiguous-context score shows instruct models defer rather than debias (28% → 53%) and are more biased when they answer (conditional bias 0.53 → 0.59).Suppressed bias is recoverable: the most aggressive adversarial framing restores CrowS-Pairs bias to near base levels, while on BBQ it partly reverses deferral without reactivating conditional bias.(iii) Across eight probed pairs, base and instruct gender directions remain geometrically aligned (median cosine 0.86 vs. 0.006 randompair baseline).Together, these findings suggest benchmark improvements after alignment may overstate genuine bias reduction.
Authors: Sigurd Kornelius Havstein, Anders Søgaard
Institution: University of Copenhagen
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7224499501