Teacher Forcing as Generalized Bayes: Why Switching Training Surrogates for Chaotic Dynamics Creates Optimization Geometry Mismatch

arXiv·low signal

Identity teacher forcing (ITF) has been the go-to method for training recurrent surrogates on chaotic dynamical systems, but switching between different training surrogates (teacher-forced vs. free-running) creates optimization geometry mismatches that destabilize training. The paper formalizes this through generalized Bayesian inference, explaining why practitioners see sudden training collapses when transitioning between regimes. Primarily relevant to ML practitioners working with time-series modeling of chaotic systems (climate, turbulence, financial markets).