AI
Overcoming Scaling Limits in On-Policy Self-Distillation for LLM Reasoning
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 01/10/2026, 10:51:06 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 09:18:24 UTC.
Briefings
Research on On-Policy Self-Distillation (OPSD) reveals that scaffold correctness impacts downstream accuracy more than context correctness. Verified on-policy trajectories improve performance by 3.05 points at 8B parameters, whereas OPSD's gain drops to 0.14 at 8B. The study introduces OASIS, which uses verified labels and unverified model attempts as teacher context, achieving 3.2-3.8 average improvements across Qwen3-1.7B, 4B, and 8B models on AIME 2024, AIME 2025, and HMMT 2025.
Key points
- Verified on-policy trajectories improve performance by 3.05 points at 8B parameters, whereas OPSD's gain drops to 0.14 at 8B.
- Scaffold correctness has a stronger effect on downstream accuracy than context correctness in On-Policy Self-Distillation.
Context and limitations
Results are limited to Qwen3-1.7B, 4B, and 8B models on AIME 2024, AIME 2025, and HMMT 2025; no data on smaller or larger models.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 10:51:06 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiehzyzyfqyxiesirk57yrhvojdk2spzmldk57dkn2e5467fptqq7m