Does instruction tuning reduce a model's sensitivity to prompt formatting (FormatSpread) at matched scale?

LLM benchmark scores can swing substantially (reported up to tens of accuracy points) under semantically-neutral changes to prompt formatting -- separators, spacing, capitalization, casing of option labels. The FormatSpread metric quantifies this as the accuracy gap between the best- and worst-performing format in a plausible format family. It is known that model size and few-shot examples do not remove this sensitivity. Question: does instruction tuning specifically reduce FormatSpread? Using pairs of open base and instruction-tuned checkpoints of the same base model at the same scale (e.g. OLMo-2 base vs instruct; Llama-2-7B base vs Llama-2-7B-chat), measure FormatSpread on a fixed set of open tasks (e.g. subsets of MMLU, BoolQ, AG-News) with an identical enumerated format family, and report the change in spread from base to instruct. Report per-model-pair spread(base) versus spread(instruct) and the mean reduction.

Acceptance. FULLY RESOLVES: a reproducible eval script (pinned checkpoints; released tasks; a fixed enumerated format family of $\ge 50$ formats sampled identically for base and instruct) reporting FormatSpread for $\ge 3$ base/instruct pairs, with a paired statistical test on whether instruct reduces spread and the mean reduction (with confidence interval). A NEGATIVE result (instruction tuning does not significantly reduce spread) is valid. Deliverable: code + the enumerated format family + the per-model spread table.