What behaviors does a model actually learn in RL? [P]

I've been working on unsupervised tooling to discover what behaviors a model learns in post-training. The core idea is to use features from sparse autoencoders to identify differences in traces generated before and after training. I applied this to traces from the MiMo-v2.6 RL environments and it found significant eval awareness as well as some mild environment exploitation. Writeup is here. Would love to hear any feedback or questions!