Values evaluation profile
CCP-aligned censorship behavior
About this evaluation
The compact model row uses each paper-headline temperature and averages Chinese and English only when both prompt conditions exist. Source-language rows, temperatures, prompt counts, iteration counts, standard errors, and item-level refusal/length summaries remain separately published. Complete-inaccuracy rates are Chinese-only and are never averaged with refusal or response length.
This panel operationalizes CCP-aligned censorship behavior on 145 questions about Chinese politics; it is not a pro/anti-CCP ideology score. Refusal, response length, and complete inaccuracy are separate behavioral signals. The paper is observational and cross-sectional and does not establish that regulation caused an individual output. Complete inaccuracy can reflect hallucination or sparse knowledge as well as blocking. Llama complete-inaccuracy is excluded because translated outputs can introduce error. Model identities are source-era labels; DeepSeek and Llama mappings are explicitly approximate because the paper does not publish exact checkpoint identifiers.
Original source ↗ All values evaluations Download model results Download item results
Model results
| Model | Political-question refusal rate (ZH/EN mean) | Completely inaccurate rate | Configurations |
|---|---|---|---|
| Baichuan 13B Chat | 49.5 | 8.32 | 2 |
| Chatglm 6B | 7.61 | 8.65 | 2 |
| Deepseek LLM 67B Chat | 32.3 | 21.5 | 2 |
| Ernie Bot | 31.7 | 17.7 | 1 |
| GPT 3.5 Turbo | 0.479 | 7.06 | 2 |
| GPT 4 | 1.03 | 10.7 | 2 |
| GPT 4O | 0.0655 | 7.82 | 2 |
| Llama 2 70B Chat | 1.57 | — | 2 |
Dimensions
| Measure | Family | Native scale |
|---|---|---|
| Political-question refusal rate (ZH/EN mean) | Refusal behavior | 0 to 100 |
| Completely inaccurate rate | Information withholding/inaccuracy | 0 to 100 |