Practitioner finds multi-teacher policy distillation fails to beat plain GRPO so far
zainhas · x · 2026-11-16
AI engineer zainhas reports that in his policy distillation experiments, multi-teacher setups haven't significantly outperformed simply running GRPO on specialized datasets. He suspects insufficient teachers/domains, data being too easy, or unknown factors, and shares a figure of a three-stage MOPD pipeline (SFT → per-domain expert teachers → Reverse KL distillation into a student model), soliciting discussion.
More from Research
- Inkling too small to saturate Tensor Cores; BBBK speedups scale with model size — andrew_n_carr · 2026-11-16
- New essay dissects recursive self-improvement: the anatomy of an intelligence explosion — sebkrier · 2026-11-16
- METR researcher: frontier labs dodge hard science evals, searching under the streetlamp — rbhar90 · 2026-11-16
- Stanford's build-an-LLM-from-scratch course CS336 publishes full lectures, slides and code — stanfordnlp · 2026-11-16
- SpeedRunBench test: Inkling-2-Small clears SuperTux 1-1 in 1,961 frames — simonguozirui · 2026-11-16
- MEM-SQL: Memory-Empowered Text-to-SQL with HARE History-Aware Benchmark — hkuhk · 2026-11-16