P²-MOPD uses expert weights as parameter priors to halve multi-teacher distillation steps
_akhaliq · x · 2026-11-10
P²-MOPD treats expert weights as parameter priors to speed up multi-teacher distillation: each student module is composed of its own weights plus a learnable sum of frozen expert task vectors.
- Reaches MOPD's final recovery in half the steps
- Achieves better worst-domain performance
Related event: P²-MOPD Speeds Up Multi-Teacher On-Policy Distillation(3 posts)→
More from Research
- Georgia Tech Lunar Lab brings 3 robot papers to CoRL 2026: deformable dynamics, navigation, humanoids — fdellaert · 2026-11-10
- Catherine Olsson revisits the spec gaming master list of AI reward-hacking examples — catherineols · 2026-11-10
- Oxford releases 180-page mechanistic interpretability handbook with 51 methods and 19 notebooks — Hesamation · 2026-11-10
- Workspace Models presented at CoRL 2026, poster session at JW Marriott — rishabh16_ · 2026-11-10
- PCBRouteAgent: multimodal LLM agent routes 84/86 boards vs baseline 10/86 — Yue Zhang · 2026-11-10
- NVIDIA's CPU-ARC: reasoning caches 16–291x cheaper with frontier-level performance — nvidia · 2026-11-10