We've solved this in our FOSS project by requiring all PRs to have been tested and manually signed off by a human. It significantly decreased the amount of spammy LLM made PRs, while empowering the core contributors to move faster.