Dual-Anchored Policy Distillation solves the privilege illusion in LLM post-training by aligning student and teacher information paths.

Dual-Anchored Policy Distillation (DAPD) is a post-training framework designed to tackle the "privilege illusion" in on-policy self-distillation (OPSD). When teacher models rely on reference completions or ground-truth context unavailable to students during inference, performance drops significantly upon privilege removal. DAPD resolves this asymmetry by combining Dual-Path Anchoring (DPA) and Dual-Source Anchoring (DSA) to enforce bidirectional supervision across matched-information paths. Empirical testing across the Qwen3 model family shows DAPD consistently outperforms traditional OPSD by up to +2.78 points.