Training Qwen3 with Explicit Plans Compiles Arbitrary Patterns into Layer-wise Steps
Experiments on Qwen3-4B and Qwen3-1.7B show that training with explicit plans lets models compile arbitrary requested patterns into layer-wise generation steps, achieving 49%–78% success on unseen edits, with activation heatmaps revealing corresponding layer-wise internal structure.
2026-10-21 ~ 2026-10-21 · 3 related posts
- Experiments show Qwen3 can compile arbitrary patterns into layer-by-layer steps with plan training — Sauers_ · 2026-10-21
- Layer-wise heatmaps show trained models 'compile' patterns into planning steps — Sauers_ · 2026-10-21
- J-lens evaluation: 49-78% success on never-seen edits, but well below trained concept vectors — Sauers_ · 2026-10-21