Instruction Duplication Eliminates 30% of Residual Instruction-Following Failures With No Retraining or Decoding Change
arXiv 2609.04024 introduces a minimal black-box inference-time control that simply repeats the procedural instruction. Across seven instruction-tuned models, 300 medical multiple-choice questions, eight placement conditions and 16,800 scheduled generations, moving from one copy to two raised the deterministic All-8 diagnostic from 90.22% to 93.17%, eliminating 30.2% of failures remaining after one copy, while final-answer accuracy stayed exactly 60.21% and premature commitment rose from 1.52% to 2.30%. The authors are careful that a blinded audit gave 10/30 directional confirmations with 20/30 ties and did not meet its prespecified 28/30 criterion, so the value is specifically for downstream systems that act on the exposed trajectory rather than on the final answer.
↳ Follow the thread