Robust volatility updates for Hierarchical Gaussian Filtering

Researchers at arXiv (2605.00966) present a fix for a critical numerical instability in Hierarchical Gaussian Filtering, where volatility-coupled parent nodes could produce negative posterior precision, crashing inference. Their modified quadratic approximation interpolates between two variational energy expansions, guaranteeing valid precision estimates across all parameter regimes. Practitioners building HGF-based computational psychiatry models or adaptive Bayesian agents should upgrade immediately. Prior HGF implementations required manual parameter constraints as workarounds; this patch eliminates that fragility at the update-equation level.