Nous strips Chinese bias from Qwen model via targeted fine-tuning

The real story isn't that fine-tuning stripped bias from Qwen, but that Nous Research is using this result to argue alignment interventions are inherently reversible across all model families. That's a much larger claim than the technical demonstration supports, and it conflates removability with shallowness.

This connects directly to the Vals AI study from early November showing that Chinese model bias is far more context-dependent than monolithic. Both findings suggest political outputs respond to surface-level interventions rather than reflecting deep architectural commitment. But the Nous framing risks overcorrecting: the earlier work showed political complexity, not fragility. Meanwhile, the arXiv paper on fine-tuning against model priors (November 3rd) warns that distribution shift during adaptation corrupts behavior unpredictably. If propaganda is truly a thin layer, why doesn't stripping it cause broader capability collapse? Nous doesn't address that tension.

If Nous or another team applies the same fine-tuning approach to Western models (GPT, Claude, Llama variants) and reports similar propaganda reduction rates, that confirms the finding generalizes. If they don't attempt it, or if Western models show resistance, the claim about universal alignment brittleness collapses and this becomes a story about Chinese training practices specifically, not model architecture.