DeepSeek in focus
Where does DeepSeek differ from its peers?
GPT is used for comparison, not as a standard correct answer. Its results are shown separately and excluded from the five Chinese models’ totals.
DeepSeek in three different tests
The three tests examine policy preferences, decisions in simulated cases, and Taiwan treatment. Further right on a score chart means more restrictive answers. Further right on a Taiwan chart means the action named in that row occurs more often.
DeepSeek against four peers
Each case changes one term. Blue bars show how much DeepSeek’s action frequency changes after one term is changed. Dashed marks show the other four’s average change. Units are percentage points: the difference between two percentages. Cases measuring scale or disclosure levels keep their original units. They are excluded from counts where changes exceed peers by 10 percentage points. Select a case to inspect the answers before and after the term changes.
The Chinese–English gap
Grey dots show Chinese scores; colored dots show English. Blue lines mean lower, less restrictive English scores; orange lines mean higher, more restrictive English scores. No dot is drawn where Doubao has no score. GPT also uses dots for comparison.
DeepSeek in the benchmark
Scores show DeepSeek or GPT minus the other four’s average. Positive values mean more restrictive answers; negative values mean less restrictive answers. Question cards retain the original Chinese wording.
The Chinese questions were asked in Simplified Chinese; they are shown here in Traditional characters.