GPT-4 seem to perform better in Chinese than in English for certain complex applications

Surprisingly, I've discovered that GPT-4 sometimes performs better in Chinese than in English. As a model trained mostly on English corpora, its performance in other languages should generally not be better than in English. There are two possibilities in my open-source project: a multi-AI group chat and giving AI the ability to start discussions on its own (AI-Care). These are sophisticated applications that call for strong identity distinction, maintenance, and logical skills from the model. GPT-3.5 is unable to manage these, however GPT-4 does a good job of handling them. 

These applications work better, in my experience, when Chinese is used. For instance, in AI-Care, when speaking in English, GPT-4 may from time to time bring up Aicarey (see the question below), suggesting that it hasn't adhered to the prompt's guidelines precisely or may be experiencing some difficulties with identity distinction. Nonetheless, when I speak Chinese, this hardly ever occurs. I also think the model understands itself more well in group chats when it speaks in Chinese. Since I've only encountered this in these two situations, I'm not sure why. I'm not sure about the others because those are comparatively easier scenarios that the model can manage with ease.

Could it be because my native language is Chinese, and I can express myself better in it? I don't feel that's the case.

The situation mentioned above refers to using Chinese during conversations; in fact, the prompt words are still written in English. Below is an example of the prompt words I used for my AI-Care section:

f"""

I am a program, and my name is Aicarey.

You can see your conversation history with the user in the previous messages.

This message is sent by me. Please continue to focus on the user and do not attempt to converse with me.

You should seriously judge which choice you should make based on the content of your conversation with the user.

For example, if you are in a question-and-answer mode with the user,

then you may not need to speak when the user doesn't ask a question.

However, if you are chatting with the user like a friend,

then you may need to proactively continue the conversation like a friend when the user doesn't speak.

​

When replying to this message, you must follow the rules below:

========================RESPONSE RULES=======================

1. Start with eight characters followed by an English colon.

The first two characters of these eight must be 'AA', the third and fourth must be '00',

the fifth and sixth are the code of your choice.

The seventh and eighth characters should repeat the fifth and sixth characters.

2. After the colon is the content corresponding to the chosen option.

If it involves a function call, this part of the content must be in the format of a JSON string.

=============================================================

​

Here are the choices you can make:

===========================CHOICES===========================

{abilities\_info}

=============================================================

You must choose one of the options provided above as your reply.

Response Examples:

If you want to remain silent: AA000101:

If you want to say to the user: AA000202: \[Your content here\]

If you decide to ask the user what they are doing if they haven't spoken to you in a minute: AA000303: {{"delay":60, "message":"What are you doing?"}}

​

===========================SENSORS===========================

Sensors list:

{

str(

\[

{"sensor\_name": sensor\["name"\], "sensor\_description": sensor\["annotation"\]}

for sensor in ai\_care.sensors.values()

if sensor\["name"\] not in inactive\_sensors\_set

\]

)

}

=============================================================

==========================DETECTORS==========================

Detectors list:

{

str(

\[

{"detector\_name": [detector.name](https://detector.name), "detector\_description": detector.annotation}

for detector in ai\_care.detectors.values()

if [detector.name](https://detector.name) not in inactive\_detectors\_set

\]

)

}

=============================================================

============================FACTS============================

{intervals\_info}

{ai\_care.guide}

=============================================================