Model selection problem. Mainly used to solve mathematics and physics problems

API
1

Currently using the GPT-4o model. It answers most questions correctly. But the performance on some complex questions is poor. The difficulty level of the questions is American high school Algebra 2. Which model can be used to answer these questions better and more accurately?

2

If you look at the benchmarks GPT-4o is the best model you can overall get for the category of questions you are hoping to answer. Look into RAG or LLM call functions to enhance the current capabilities of the Model. A way I would go about is to ask it to extract information from a problem through an llm call function and then send the variables to an api call to a math api to solve the question. Wolfram Alpha API is a good math api for example.

3

Thanks for the help. 1. My purpose is to help students improve their learning efficiency. Therefore I want GPT to be able to read printed question stems and handwritten answers (most of which has already been implemented). 2. The problem currently encountered is that the GPT answer is incorrect. For example, geometry questions and function graph questions.
3. I hope the correct answer rate is above 95%.
I am a beginner in using GPT and would like some guidance.
Below are some examples of incorrect GPT answers.
Thanks for the guidance.

image
image3419×985 195 KB

4

Okay so you are prompting GPT to answer the questions directly and the problem is that it is not reaching the 95% answer rate you are hoping for. What I suggest to do instead is this. GPT-4O is really good at extracting information from small texts like a math question so what you could do instead is ask it to extract some variables and then use some web API or other function to do the calculation to guarantee the answer is going to be correct. Here is an Example for a quadratic equation problem:

Problem:
Use the given parameters to find the values of aaa, bbb, and ccc in the quadratic equation at2+bt+c=0at^2 + bt + c = 0at2+bt+c=0.

Chat GPT Response:
{
“a”: “-16”,
“b”: “64”,
“c”: “90”
}

Final Step:
You can the use this JSON file to calculate the answer with a hard coded quadratic formula solver function.

If you are new to chat gpt look at the llm call function link. This will be very helpful for the GPT response.

5

Thank you for your help.The link you gave is very helpful

  1. The purpose of this project is to help some students with ordinary economic conditions. They lack enough money to hire private teachers. Most of these students are in economically underdeveloped countries such as Asia and Africa. So the cost issue is also an important factor.
  2. Using general prompt words, GPT can answer more than 80% of the questions.
    In elementary geometry, linear functions and quadratic functions, the performance is not very good.
  3. These questions are relatively personalized, and general prompt words cannot be used to guide GPT.
  4. The difficulty I encountered is to improve the accuracy of answers when the interaction cost is relatively low.
    Anyway thanks for the guidance.