DAPD Beats OPSD Distillation Across Qwen3 1.7B to 32B

A new distillation paper posted to Hugging Face makes a claim that lands harder if you have watched on-policy self-distillation scale poorly at 8B and above. The authors argue OPSD quietly teaches students to lean on a 'privilege illusion': the training-time teacher can peek at a reference completion, so the student learns to behave as if that reference were still there at inference, even though it isn't. As models get bigger, this becomes more expensive. The gain OPSD offered over the base policy at 1.7B (+5.19 points) collapses to at most +0.28 from 8B through 32B.

Their fix, Dual-Anchored Policy Distillation (DAPD), attacks the information asymmetry directly. Instead of letting a privileged teacher supervise a student without the same information, DAPD builds anchor distributions where both sides share the same context. Dual-Path Anchoring introduces a 'Self' distribution that conditions the student on the completion it is predicting, giving a bridge under matched information availability. Dual-Source Anchoring then applies the same construction in both directions, using the reference to guide the rollout and the rollout to guide the reference, so the student is not stuck learning from a single guidance source.

The reported numbers are cleaner than typical distillation papers. On Qwen3-4B, DAPD beats OPSD by +2.00 points averaged over six benchmarks (AIME24, AIME25, HMMT25, LCB v5, BFCL v3, IFBench), reaching a task average of 57.34. It also beats two more sophisticated OPSD variants, Purified OPSD and DOPD, by +1.09 and +3.85 points. The scaling story is the more interesting one: DAPD retains gains of +2.41, +2.13, and +3.06 points at 8B, 14B, and 32B respectively over base policies. A behavioral probe backs the mechanism: wrong claims, where the model asserts an unsupported answer, drop 73% relative to OPSD over training steps 250 to 300.

The honest caveat is that this is a single-lab result on one model family (Qwen3, 1.7B to 32B), trained with LoRA on OpenThoughts data, so how much of the benefit transfers to Llama, DeepSeek, or Mistral is untested. The method also adds training-time computation for the extra anchored distributions and requires per-scale recalibration, with the paper's own reference-guidance weight dropping from 0.5 at 1.7B to 0.2 at the larger sizes. What the reporting does not give you is wall-clock overhead or how DAPD composes with reinforcement learning after distillation.

For teams working on reasoning models at 8B and above, where OPSD flattens, this reads as a specific technique to try rather than a general framework, and the authors have released code on GitHub for independent replication.