Does multiple-choice selection bias shrink with scale or instruction tuning in open models, and does PriDe debiasing transfer?

LLMs answering multiple-choice questions exhibit 'selection bias': a prior preference for particular option IDs (e.g. 'A') independent of content, so permuting the option order changes the answer. Zheng et al. traced this to token bias over the option-ID tokens and proposed PriDe, a label-free inference-time debiasing that estimates the ID prior from a few permuted samples. Question: using open model families with multiple scales (Pythia 160M-2.8B, and/or OLMo scales) on a fixed MCQ benchmark (e.g. an MMLU subset, ARC), measure the selection-bias magnitude -- e.g. the standard deviation of recall across option IDs, or the accuracy drop under answer-order permutation -- as a function of scale and of base-versus-instruct, and test whether PriDe's bias reduction transfers to these open models. Report bias magnitude versus scale, base versus instruct, and residual bias after PriDe.

Acceptance. FULLY RESOLVES: a reproducible script (pinned checkpoints; released MCQ tasks) reporting, across $\ge 3$ scales and $\ge 1$ base/instruct pair, (i) a selection-bias metric (recall standard deviation across option IDs and/or permutation-induced accuracy variance), (ii) its trend with scale, and (iii) the residual bias after applying PriDe, with the reduction quantified per model. Deliverable: code + bias-versus-scale table + pre/post-PriDe comparison.