{"slug": "auditing-learned-dynamics-in-weather-models", "title": "Auditing Learned Dynamics in Weather Models", "summary": "A study of GraphCast, DeepMind's autoregressive weather forecasting model, found that sparse autoencoder features need not carry conventional physical meaning to be causally relevant, according to arXiv paper 2611.25119v1. Ablating a feature locked to GraphCast's computational mesh increased global geopotential height error, while scaling up convection-associated features cut tropical-cyclone intensity error by a median of 51% across seven storms. The authors say the results separate interpretable physical association from causal utility and establish sparse features as experimental coordinates for auditing and intervening on learned weather dynamics.", "body_md": "arXiv:2611.25119v1 Announce Type: new \nAbstract: Learned weather models are experimentally accessible computational systems whose internal representations can be inspected and manipulated. Using sparse autoencoder features from GraphCast, DeepMind's autoregressive weather forecasting model, we test whether causal relevance requires physical semantics, a question that matters for auditing whether AI forecasters used in high-assurance settings rely on reliable mechanisms. We also look at whether physically associated latent directions can be steered to improve extreme-event forecasts, where interventions are applied to the model's internal activations during the rollout. We find that ablating a feature locked to GraphCast's computational mesh increases global geopotential height error despite lacking conventional physical semantics. Second, suppressing convection- and low-level-vorticity-associated features weakens predicted tropical-cyclone deepening (the fall in central pressure during intensification), while amplifying them reduces intensity error, with amplification of convection features through feature scaling achieving a median of 51\\% error reduction over seven storms. Together, these results separate interpretable, physical association from causal utility and establish sparse features as experimental coordinates for auditing and intervening on learned weather dynamics.", "url": "https://wpnews.pro/news/auditing-learned-dynamics-in-weather-models", "canonical_source": "https://arxiv.org/abs/2611.25119", "published_at": "2026-11-20 05:00:00+00:00", "updated_at": "2026-11-20 05:14:41.569420+00:00", "lang": "en", "topics": ["machine-learning", "ai-research", "ai-safety", "artificial-intelligence"], "entities": ["GraphCast", "DeepMind", "arXiv"], "also_reported_by": [], "takeaway": "Sparse autoencoder features in GraphCast are causally relevant without physical semantics: ablating a mesh-locked feature raised global geopotential height error, and scaling convection features cut tropical-cyclone intensity error by a median of 51% over seven storms.", "recommendation": {"advice": "Audit learned weather models by ablating and scaling sparse autoencoder features during rollout, not by physical labels alone; scaling convection features cut intensity error by a median of 51% across seven storms.", "playbook": "evals-testing", "playbook_url": "https://wpnews.pro/playbooks/evals-testing"}, "alternates": {"html": "https://wpnews.pro/news/auditing-learned-dynamics-in-weather-models", "markdown": "https://wpnews.pro/news/auditing-learned-dynamics-in-weather-models.md", "text": "https://wpnews.pro/news/auditing-learned-dynamics-in-weather-models.txt", "jsonld": "https://wpnews.pro/news/auditing-learned-dynamics-in-weather-models.jsonld"}}