Quantum Discovery Lab
After OpenAI’s proprietary GPT model solved an award-winning maths problem, I wanted to try something similar myself. I am currently an undergraduate student studying Nanotechnology at Sivas Cumhuriyet University, and building on this, I wanted to develop an AI model capable of generating hypotheses in areas that have not yet been explored, such as material synthesis, quantum structures and atomic particles. As I needed a source problem, I opted for Unsloth and, amongst the Unsloth-supported models, chose the Phi-4 model—one of the most successful models in the fields of mathematics and natural sciences—and trained it using GRPO.
If you require the simulation code and reward function, you can find them in the project files at https://oguzyesil.dev.
Charts
Although the results show that I was able to teach the model something, this does not mean that the model can discover things that do not exist, or that the hypotheses it discovers will work 100 per cent in the real world or are physically correct. However, using a novelty checker, extra points were awarded in the reward function for physical hypotheses that the model had not yet discovered, thereby encouraging the model to generate new hypotheses.
This repository contains only the c250 (the model saved at step 250). If you wish to access and try out the other models, you can find them in the project files at https://oguzyesil.dev.
- Downloads last month
- -