Data-free On-policy Distillation¶
L1 mechanism diagnostic. A diagnostic of policy features on self-generated questions, without producing a natural-language dataset.
Paper information¶
| Field | Value |
|---|---|
| Paper | Data-free On-policy Distillation |
| Affiliations | Institute of Automation, Chinese Academy of Sciences / Tencent |
| First public release | 2026-09-12 (arXiv v1) |
| Authors' code | No public author repository found as of 2026-09-16. |
| Adapter / method | df-opd |
| Local code | src/auto_research/post_training/latest_20260916.py |
Original paper¶
Background and contribution¶
Have the teacher generate questions in its own policy space, reducing dependence on external post-training corpora.
flowchart LR
I[Teacher policy] --> Q[Question-generation mechanism] --> S[Student policy features] --> O[Auditable diagnostics]Key figure from the paper¶
Original Figure 1: data-free distillation workflow. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
The reference preserves defining gate, weighting, and state transitions and reports auditable intermediates under fixed budgets. Three seeds test mechanism invariants, not paper-scale training.
Paper results¶
The authors' capability, speed, and cost are not local measurements or online gains.
Local reproduction¶
Three-seed results are marked diagnostic_only=true and do not establish full capability.
Reproduction limits¶
No natural-language training set is generated locally.