AppSim

We present AppSim, a family of learned world models for API simulation. Building on the model-based reinforcement learning successes in vision-based games and text-driven environ- ments, we consider two approaches. In the first, we finetune a TinyLlama model on a custom dataset of synthetic trajectories from the Ap- pWorld engine, combining semi-random API exploration with ChatGPT-driven user requests. In the second, we use a powerful off-the-shelf LLM in a zero-shot setting. Evaluating on held- out trajectories using object-match accuracy, BLEU, and ROUGE, we find that the general- purpose LLMs exhibit superior performance.

You can find the details about our findings in the final report and see our code here.