Distilled RL mixes teacher supervision into LLM post-training and beats RL, OPD

ZGCA · hf · 2026-07-21

Distilled RL is a new post-training method for LLMs that tries to combine reinforcement learning with teacher guidance.

• The paper argues that plain RL suffers from coarse outcome supervision and hard credit assignment, while on-policy distillation can either teach too little or become ineffective when the teacher is too different.
• Distilled RL integrates teacher supervision directly into the RL objective so it can transfer new knowledge more selectively instead of blindly matching teacher logits.
• The method includes three components: reverse importance sampling with clipping, negative sample reset, and sequence-level geometric normalization.
• In both within-family and cross-family distillation, the authors report substantially better pass@1 and pass@k than standard RL and OPD.

Original post →

More from Research

• AI slop is already clogging PR review and weakening the credit system behind science — rbhar90 · 2026-07-27
• ICML 2026 oral paper replication scores stay middling after a stricter re-scoring — profjamesevans · 2026-07-27
• Long-running agents will need immutable event logs, this thread argues — sebpaquet · 2026-07-27
• Seed IQ navigates Doom II, prompting questions about benchmarks beyond ARC-AGI — Fit_Transition8824 · 2026-07-27
• Agentic Data Science in Practice: Agents Write Code but Answer Wrong Questions — hugobowne · 2026-07-27
• A concise canon of foundational papers in ML, systems, NLP, speech, and audio — deliprao · 2026-07-27

Research channel →