Explicit–Implicit Bias Divergence in Smaller or Older LLMs

Determine whether the explicit–implicit divergence in bias expression across evaluation tasks manifests similarly in smaller-scale or older large language models.

Background

The paper demonstrates that bias expression in LLMs is task-dependent: models tend to counter-stereotype on explicit probes while reproducing stereotypes on implicit ones. This effect is interpreted as a consequence of safety alignment suppressing overt bias without altering underlying associations.

All evaluated systems are frontier-scale 2026 models spanning commercial and open-weight families. The authors note a gap in understanding whether the same divergence between explicit and implicit tasks appears in models that are smaller or from earlier generations, which limits the generalizability of their findings.