Experiments show Qwen3 can compile arbitrary patterns into layer-by-layer steps with plan training
Sauers_ · x · 2026-10-21
The author ran training experiments on Qwen3-4B and Qwen3-1.7B to test whether adding explicit "plans" lets the model compile arbitrary requested patterns into layer-by-layer generation steps.
Key results from the charts:
- On in-training-set patterns (concept, wording, and patterns combined), the 4B model trained with plans improved from 61% to 75% share of requested change, pattern-match r from 0.59 to 0.66
- Generalization to unseen categories is mixed: wording and concept transfers well (58-78%), but layer patterns are weak — plan-trained 4B even shows negative correlation (r=-0.13) on some layer patterns
- Notably, the original model without thinking already does decently on many patterns — planning helps, but isn't necessary everywhere
More from Models
- OSWorld-Science benchmark: 12 VLMs average just 34.3% on real scientific software tasks — jiqizhixin · 2026-10-21
- Ampersand debuts first model '&', built on long-horizon simulation stack — niloofar_mire · 2026-10-21
- Hugging Face quietly added model file size display, a big QoL win — vboykis · 2026-10-21
- Humans& launches Ampersand, first model with privacy built in, plus a funny RL-plot joke — niloofar_mire · 2026-10-21
- A single GLM neuron keeps firing on both sci-fi self-replicating malware and real cyber risk — burny_tech · 2026-10-21
- Gemini 4 Argon tops Isoquant's inference benchmark with 4.38x mean GPU speedup — ycombinator · 2026-10-21