Procedural instruction following is essential for controllable language‑model systems, especially when generated trajectories are inspected or repaired downstream. We introduce instruction duplication, a minimal black‑box inference‑time control that simply repeats the procedural instruction without any model retraining or decoding changes. Experiments span seven instruction‑tuned models, 300 medical multiple‑choice questions, eight placement conditions, and 16,800 scheduled generations. Raising the duplication count from one to two copies increases the deterministic All‑8 diagnostic‑response rate (passing all eight observable tests) from 90.22% to 93.17%, a gain of 2.95 percentage points, eliminating 30.2% of the failures that remained after a single copy. Pre‑provisional TF‑IDF recall improves from 73.44% to 74.81% (+1.38 points, Holm‑adjusted significance).
Review