SR-OPSD: Self-Referenced On-Policy Self-Distillation¶
Candidate-policy objective diagnostic, not paper-scale distillation.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Independent researcher Zhuo Sun. |
| First public release | 2026-08-10 (arXiv v1) |
| Author code | No public author code found as of 2026-08-24. |
| Local adapter or method key | sr-opsd |
| Local reproduction code | src/auto_research/post_training/historical_b08_b09.py |
Paper method and evidence¶
Geometrically interpolate self-teacher and reference to obtain an effective target. Rényi divergence controls projection geometry and density-ratio sensitivity.
flowchart LR
A[Self-teacher and reference] --> B[Geometric blend]
B[Geometric blend] --> C[Rényi-controlled projection]Original paper figure¶
Figure shows self-referenced target construction. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
The paper reports competitive or state-of-the-art results across models and science, math and code tasks. No online evidence inferred.
Local reproduction¶
Same arithmetic-smoke candidate policy, 120 updates: 0.1953→0.6797 (+0.4844), not a fair cross-method comparison. Metrics, batch index. No original checkpoint, full training, multi-seed or benchmarks; runs/post-training/ holds raw runs locally and only metrics are committed. Experiment artifacts.