researchAI Research1d ago
SWE-OPD Research Introduces On-Policy Distillation for Agentic Coding
Facts
Researchers have developed SWE-OPD, a methodology for transferring software engineering capabilities from large teacher models to smaller student models using on-policy distillation. The study demonstrates that this approach improves resolve rates on SWE-bench benchmarks while significantly reducing computational costs and GPU requirements.
- ▸SWE-OPD method transfers coding capabilities from teacher LLMs to smaller student models via on-policy supervision.
- ▸Distilling Qwen3.6-27B into Qwen3.5-9B increased SWE-bench Pro resolve rates from 35.9% to 39.4%.
- ▸Offline variant of SWE-OPD reduces end-to-end GPU-hours by 59.4% to 72.6% and halves peak GPU allocation.
- ▸Cross-tokenizer distillation of GLM-4.5-Air from DeepSeek-V4-Flash-0731 showed performance gains of up to 4.6% on multilingual benchmarks.
- ▸Study identifies that decreasing loss does not always correlate with improved performance, citing potential issues with support-mass gradient terms.
- Occurred
- 2026-11-16 12:39 UTC
- First reported
- 2026-11-16 12:39 UTC
- Last updated
- 2026-11-17 08:16 UTC
- Event type
- research
- Region
- global
Entities
Categories
Source timeline · 1
- arXivFirst-party2026-11-16 12:39
Community vote
Sign in to vote on importance and sentiment.
Platform Signals
Impact score
62
AI sentiment
PositiveEvent significance30%1
Entity importance20%1
Source credibility20%1
Novelty15%1
Corroboration15%1
Confidence1
Novelty1
Corroboration1
Community
ImportanceNo votes
Strategic sentiment
Positive 0%Neutral 0%Negative 0%
Platform signals are algorithmic and community-derived indicators. They are not investment recommendations and do not predict valuation or performance.