Definition
On-policy self-distillation (SDPO) is a post-training technique where a model teaches itself from its own on-policy rollouts—used as the hidden reference innovation in innovationeval because it is recent, cited, and adopted in frontier training stacks.
Key Points
- Epoch’s InnovationEval withheld the paper name but supplied the same datasets/metrics as the human SDPO result; agents were asked for a novel method beating GRPO on Qwen3-8B
- GPT-5.6 Sol’s best attempt resembled prior self-imitation ideas and was partly out-of-scope; in-scope wall-clock-adjusted gains reached only 15% of SDPO (innovationeval)
- Serves as a stress test for algorithmic novelty vs. hyperparameter tuning or memorized implementations