AI Briefing#on-policy-distillation
The latest AI and developer news about #on-policy-distillation, with the original source and a short summary.
The latest AI and developer news about #on-policy-distillation, with the original source and a short summary.
The framework parallelizes draft sampling and supervision across independent branches to improve inference efficiency for large language models.
JOLT jointly trains a single model as both teacher and student to provide dense, aligned guidance, achieving higher accuracy and significantly better sample efficiency than GRPO on math, coding, and agentic benchmarks.
Semi-OPD outperforms on-policy distillation in 14 of 17 tested teacher-student pairs, achieving up to 11.4x training speedup.
FuriosaAI presented four papers at ICML 2026 addressing efficiency in LLM inference, training, and serving.
This paper proposes a diagnostic framework that analyzes, at the token level, the impact of On-Policy Distillation on training reasoning models.
Introducing On-policy distillation (OPD), a core training technique used in the latest models such as Qwen and DeepSeek.
The author compares SFT, RL, and OPD from a distributional perspective and emphasizes the importance of policy-based sampling.