Does different GPT API differ in the performance of customerized prompts?

[Joan-YAO]

I have tried GPT-4o-mini in the LLM agent but get a sub-optimal decision like the following screenshot. The decision (phase 2) provided by the embedded RL is more reasonable than the one (phase 1) provided by LLM agent. Is this problem caused by the version of API version? Or is it caused by the imcomplete logic judgement in the chain-of-thought or prompt engineering? Which file contains the prompt for me to make some adjustment or modification?

[wmn7]

Hello,

Thank you very much for your interest in this project.

You can modify the prompt in the following file: https://github.com/Traffic-Alpha/LLM-Assisted-Light/blob/main/TSCPrompt/llm_prompt.py

I used GPT-4 for testing. I've reviewed most of the outputs, and the logic and final decisions are consistent.

Bests, Maonan