AINewsnow

On-policy distillation loses to GRPO on compute efficiency

Researchers compared vanilla reverse-KL on-policy distillation with GRPO across task difficulty, student and teacher size, domain, and optimisation choices, measuring held-out accuracy against training FLOPs. With an external post-trained teacher, OPD generally lost to GRPO on accuracy per training…

Read the full story at DEV Community — Machine Learning ↗

Timeline · 1 report

  1. 2026-11-08 07:01 · DEV Community — Machine Learning
    On-policy distillation loses to GRPO on compute efficiency

More stories

  1. Election backlash against data centers rattles AI stocks and expansion plans — TechCrunch AI
  2. Anthropic strikes $3.4bn deal for gas-fired generators to power AI data centres — Financial Times AI
  3. OpenAI to keep expanding AI data centers despite election results favoring restrictions — DigiTimes Asia
  4. Tencent AI chief warns AI will soon master biology, raising new safety risks — SCMP China Tech
  5. Sources: John Ternus wants more urgency in Apple's AI efforts, pushing Craig Federighi to move beyond Siri AI tweaks to agentic and other advanced capabilities — Techmeme
  6. US AI Safety Center Staff Now Report to White House, Raising Independence Concerns — AP News
  7. US voters reject AI data centers in nearly all local ballot measures — GeekWire
  8. Developer's 48-Hour Field Notes Detail Free-Tier LLM Batch Job Failures — DEV Community — AI

Get the daily brief of stories like this at 6:30 every morning →