DUET Improves LLM Compliance with Runtime Prohibitions

Large Language Models (LLMs) deployed in real-world settings often need to adhere to dynamic, runtime-injected prohibitions, such as enterprise policies or PII redlines. Existing training methods like SFT and DPO are not well-suited for these token-localized, variable constraints. Researchers have developed DUET (Dual-Teacher On-Policy Distillation), a novel method for improving LLM compliance. DUET employs two identical-weight teachers: one that sees the prohibition and one that does not. Their per-token disagreement effectively isolates the prohibition's causal effect, providing a clean supervision signal. This disagreement drives signal cleaning and preference-directed learning, pushing the student model away from non-compliant tokens and towards compliant ones. Evaluated on an industrial benchmark, DUET achieved 72.3-85.2% violation compliance across Qwen models while retaining 88-93% normal utility, significantly outperforming other distillation baselines.