Emergent Latent-State Computation under Stochastic Volatility

Xiaoyu Huang, Lulu Wang

Mechanistic interpretability has largely focused on language models and deterministic toy tasks. Much less is known about how sequence models internally represent latent stochastic…