Answer Engineering: Local Trajectory Editing for Protocol-Constrained Decision Making in Large Language Models

The paper introduces "Answer Engineering," a deterministic runtime intervention method for large language models that enhances protocol compliance in decision-making without requiring model retraining. Evaluated on a clinical benchmark for sudden sensorineural hearing loss (SSNHL), the approach improved compliance from 54.5% to 83.5% and increased balanced accuracy from 42.0% to 80.7% by applying localized rule-guided interventions during autoregressive generation. This method is significant for practitioners as it enables more reliable outputs in critical applications by addressing procedural adherence through auditable control mechanisms.