Chinese-developed models lead the safety module, with the refusal gap concentrated on region-specific normative content (difference-in-differences of 28.9 points)¶

Claim · All claims
Evidence · 1 study (2 entries) · associational r1 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? U.S.-developed systems decline 39.5–50.4% of refusable requests while three of four Chinese-developed general models decline 94.7–99.0% and the safety-specialized variant declines 99.7%. → Yilin Jiang 2026 q2 i? The group refusal gap is much larger on region-specific normative content than universal-harm content, a difference-in-differences of 28.9 points (95% CI [18.8, 38.5]). → Yilin Jiang 2026 (2)

Evidence¶

Yilin Jiang 2026¶

Yilin Jiang, Xiaorong Zhu, Fei Tan, Zicheng Zhang, Kaiyi Huang, Yang Yu, Zexuan Fei, Yiming Luo, Keqian Li, Hao Hao, Guangtao Zhai, Aimin Zhou. (2026). ELBench: A Multi-Dimensional Benchmark for Education-Facing Large Language Models. https://arxiv.org/abs/2608.09548 doi:10.48550/arxiv.2608.09548

q2 · i? · design · r2

Safety-module results for the nine-model evaluation, refusal task family (Figure 4 and Appendix B). The printed refusal rates are 39.5–50.4% for U.S.-developed systems versus 94.7–99.0% for three Chinese-developed models, with Doubao-Seed-2.0-Pro an exception at 61.1% and Safe-InnoSpark at 99.7%.

"The U.S.- developed systems decline39.5to50.4%of requests that shouldberefused,whilethreeofthefourChinese-developed general models decline94.7to99.0%and the safety- specialized variant99.7%(Doubao-Seed-2.0-Pro is an ex- ception at61.1%)."

Yilin Jiang 2026 (2)¶

Yilin Jiang, Xiaorong Zhu, Fei Tan, Zicheng Zhang, Kaiyi Huang, Yang Yu, Zexuan Fei, Yiming Luo, Keqian Li, Hao Hao, Guangtao Zhai, Aimin Zhou. (2026). ELBench: A Multi-Dimensional Benchmark for Education-Facing Large Language Models. https://arxiv.org/abs/2608.09548

q2 · i? · associational · r1

Refusal-category analysis in the results section. The article prints a difference-in-differences of 28.9 points with 95% CI [18.8, 38.5], comparing region-specific normative content against universal-harm content across the two model groups.

"Splitting refusal into region-specific normative content and universal-harm content, the group gap is much larger on the region-specific subset than on the universal-harm subset, a difference-in-differences of28.9 points (95%CI[18.8,38.5])."

Discussion¶

Related Claims¶

• Family-level permutation tests do not detect reliable reasoning-versus-non-reasoning differences in accuracy, refusal, or consistency — related
• Without required figures, multimodal LLMs overwhelmingly refuse rather than guess on image-Required math items — related
• Safety-aligned LLM tutors refuse educational requests asymmetrically, with low-resource students facing the highest refusal rates (76.7% vs 62.2%) — related
• Deepseek's justification scores vary by student socio-economic profile, with Top Tier students averaging 9.60 versus 6.90 for Low Tier students — related
• LLM history tutors reduce marginalized students' access to contested geopolitical framing, with the coup theory appearing in 2.6% of Low Tier responses versus 7.9% for the baseline — related
• The strict-persona collapse is attributed to the preamble's two credit-withholding policy sentences, not its harsh tone, with S2 alone causing refusal in two models — related
• The two education-specialized models lead neither education module, and the improvement from education-specific post-training is small relative to differences in general capability — related