Semi-OPD: offline student rollouts beat on-policy distillation in 14 of 17 model pairs
siyan_zhao · x · 2026-10-13
Studying whether on-policy sampling is always best for distillation, researchers tested 17 teacher-student pairs and found that Semi-OPD—distilling on offline rollouts from the initial student—outperforms OPD in 14/17 cases, faster and better. Gains correlate strongly with initial teacher-student alignment: poorly aligned pairs can suffer from online student rollouts.
Related event: NVIDIA Study Challenges On-Policy Distillation Superiority(3 posts)→
More from Research
- Convolution: the decades-old engineering math that CNNs later learned to automate — techNmak · 2026-10-13
- DeepMind open-sources 'cryptids': Turing machines with Lean proofs of famous conjectures — gaganghotra_ · 2026-10-13
- TERRA team's cancer samples arrive for spatial transcriptomics, targeting ~200M cells — mo_lotfollahi · 2026-10-13
- Base Models Detect Hidden Intentions (AUC 0.94) That Prompted Classifiers Miss — burny_tech · 2026-10-13
- Learning robot manipulation from human videos alone, no teleoperation demos needed — burkov · 2026-10-13
- Nature 2024: Repeated Training on AI Output Causes 'Model Collapse' — CurieuxExplorer · 2026-10-13