A paper made me stop and read for half an hour. S0 Tuning

Core idea: Without changing the model weights, just tuning an initial state matrix can significantly improve the model's coding ability.
On Qwen3.5-4B, using only 48 HumanEval training samples (not 48K, but 48), S0 tuning increased pass@1 by 23.6 percentage points.
Compared to LoRA, S0 outperformed by 10.8 percentage points. p-value < 0.001, statistically significant.
On FalconH1-7B, S0 achieved 71.8%.
This means that after tuning, the model's speed and size remain unchanged, only the "starting position" is better.
For those deploying local models, this opens a door: take a general model, tune it into a specialized model with just dozens of domain samples, with no performance cost.
The paper is on arXiv: 2604.01168. Anyone working on model adaptation should read it.