> I don't think we've seen any groundbreaking new research or architecture since the introduction of inference time compute ("thinking") in late 2024/early 2025 circa GPT-o4

It was model improvements, followed by inference time improvements, and now it's RLVR dataset generation driving the wheel.