OASIS: on-policy self-distillation on verified scaffolds¶
L1 selection and loss mechanism. Implements shortest verified scaffold selection, independent same-question teacher context, and clipped forward KL.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv v1 |
| Organization | First-author affiliation: North South University |
| First public release | 2026-09-29 (arXiv v1) |
| Authors' code | No author repository was found as of 2026-09-30. |
| Adapter | oasis-opsd |
| Local reproduction code | src/auto_research/post_training/latest_20260930_followup.py |
Motivation and method¶
OASIS separates which trajectory to supervise from which context the teacher sees. For each question, select the shortest verified on-policy rollout as a scaffold and preferentially use a different unverified attempt as teacher context. Questions without a verified rollout produce no training signal. This requires only a final-answer verifier, not a gold solution.
Key figure from the paper¶
Figure 1: Teacher advantages concentrate on failed trajectories, and OASIS retains gains as scale increases. From the original paper; copyright remains with the authors.
Local reproduction¶
Tests check that short unverified trajectories are not chosen, teacher context differs from the scaffold, and masked-out tokens have no gradient. See the three-seed mechanism results. Qwen3 1.7B–8B rollout and training were not run. This is a mechanism implementation and claims no AIME/HMMT gains.