Paper summary: DAPD: Dual-Anchored Policy Distillation
On-policy self-distillation with a privileged teacher can cause "privilege illusion": the student acts as if it still had information it can't see at inference. DAPD anchors training along matched-information paths to prevent this.
Problem
On-policy self-distillation (OPSD) strengthens the teacher by giving it privileged information (for example, a reference solution). The authors identify a failure: privilege illusion. The student learns privilege-dependent behaviour it can’t reproduce from its own inference-time context, but behaves as if that information were still available, so performance degrades. The root cause is the information asymmetry between teacher and student.
Method
DAPD anchors training at two levels:
- Dual-Path Anchoring (DPA): a self-conditioned bridge aligns reference and rollout behaviour along two matched-information paths, so privilege-dependent behaviour isn’t transferred.
- Dual-Source Anchoring (DSA): applies these paths in both directions (reference-to-rollout and rollout-to-reference), reducing reliance on privileged guidance while keeping correctness supervision.
Key results
From the abstract: DAPD significantly reduces privilege illusion, beating OPSD on Qwen3-4B by +2.00 points on average, with gains holding across scale (+2.69 at 4B, +2.78 at 32B).
Limitations
There’s no dedicated limitations section in the extracted text. Results are on the Qwen3 family.
#llm#distillation#post-training
Sources
- DAPD: Dual-Anchored Policy Distillation (arXiv:2608.01735) — arXiv
- Hugging Face Daily Papers, 2026-08-04 — Hugging Face
- DAPD: Dual-Anchored Policy Distillation (arXiv HTML full text) — arXiv
- Code repository (linked in the paper) — GitHub
This is a summary written from the linked public sources. Figures and numbers are as reported there.