Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-13 · high
Explicit refusal is falling across four Qwen generations while state-aligned reframing rises, per a 21,708-trial benchmark of vision-language models.
Why now: The paper appeared on arXiv on August 13, 2026.
Chinese-language prompts roughly triple the odds a vision-language model returns state-aligned framing over a neutral answer, per a new benchmark of nine models. The effect peaks at 36.5% in text-only political commentary. It persists even when the image shrinks to a silhouette, so the bias doesn't need explicit text to show up.
The study ran 21,708 trials across nine vision-language models, seven of them China-origin, using 200 prompts spanning ten sensitive topics in two languages. Two frontier LLM judges scored six dimensions, checked against three human experts, per the arXiv paper.
China-origin models reframe answers 1.6 to 3.2 times more often than non-China models. The Chinese-language effect holds inside every model tested, not just the China-origin ones.
The paper tracked four Qwen generations and found explicit refusal falling as state-aligned framing rises. A refusal is visible: the user knows something got withheld. Reframing isn't. The model answers fluently, and the reader takes it as a real answer.
Chinese-language prompts triple the reframing effect inside a single model, not just across different markets. An English-only test suite would miss that entirely.
Each link below shares sources, entities, or timing with this story.
Same source domain / Semantically similar
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.79).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.76).
Same source
Cite the same source (arXiv 2608.11816).
Same source domain / Semantically similar
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.74).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.73).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.71).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.71).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.71).