What it is
Builds a controlled multi-turn planning gym to study how long-horizon planning is learned, improved by post-training, and merged across teachers. Instead of treating planning gains as opaque, it uses a controllable environment to isolate pre-training, RL post-training, and multi-teacher distillation effects. It shows CoT state-transition world modeling and on-policy distillation matter more than atomic skills or GRPO in hard settings.
Key results
- Long-horizon accuracy 93.1 with world modeling vs 0.3 without
- Adding 5% long-horizon data raises long-task accuracy from 0.0 to 47.2
- Suboptimal data drops long-horizon pass@8 from 71.8 to 0.0
Long-horizon planning depends on explicit world-modeling and clean long-horizon supervision, while on-policy distillation is more reliable than RL credit assignment in harder regimes.