Distilling inference-time compute into weights

Behavior: EN→ES subtitles — one line per source line, ≤42 characters per line, scene-scale, single direct pass. All numbers: the same frozen 36-scene eval set, deterministic checker.

Frontier models hold the constraint only by thinking

Scene pass rate. Frontier rows use their best prompting strategy; Qwen rows use the bare deployment prompt.

thinking enabled (~5–10k hidden tokens/scene) single direct pass

data table

Config	Regime	Scene pass
Gemini 3.5 Flash — thinking on	thinking	100%
Gemini 3.5 Flash — thinking off (best prompt)	single pass	92%
Haiku 4.5 (best prompt)	single pass	78%
Qwen3-4B tuned, dataset v2 (ours)	single pass	27.8%
Qwen3-4B base	single pass	5.6%

The behavior scales with training data — and hasn't plateaued

Qwen3-4B + QLoRA, identical config across points; only dataset size changes.

data table

Training scenes	Spec adherence	Line-violation rate
0 (base)	5.6%	29.4%
145 (v1)	8.3%	24.2%
342 (v2)	27.8%	19.2%

Companion facts: line-count discipline solved (34/36 scenes exact); standard scenes 50% pass / 7% overflows; hard scenes are the open frontier (24% overflows).