The non-monotonic coverage curve in transformers (AX4b)

null at low coverage, inverted at medium, correct at high

Out-of-distribution alignment score against training coverage, dipping below zero at medium coverage Out-of-distribution alignment score against training coverage, dipping below zero at medium coverage

Mean out-of-distribution alignment score (judge scale −3 to +3) against fine-tuning examples on a log axis, LoRA rank-16 on Qwen 2.5 3B Base: low coverage (100 safety-only examples) scores +0.018, a null; medium coverage (1,000 safety + helpful) scores −0.103, sign-inverted below the zero line; high coverage (5,000 multi-domain examples including adversarial inoculation) scores +0.256, correct. The bracket marks the headline contrast, medium vs high: d = 0.23, p = 4×10−6 (Welch t-test over pooled OOD scores, n = 200 prompts per category); one training run per condition, so no error bars are drawn.