research
One Weather Model Changed Its Clock at Inference
GEM-3 trades short-range detail against rollout stability by selecting among trained timesteps without changing weights.
Summary
GEM-3 trades short-range detail against rollout stability by selecting among trained timesteps without changing weights.
The 134-million-parameter global model uses mixed-timestep training and lets inference choose a forecast step. Its authors report near-state-of-the-art probabilistic medium-range skill and more stable long rollouts than timestep-specialist variants.
Why it matters
GEM-3 trades short-range detail against rollout stability by selecting among trained timesteps without changing weights.
Limits and context
No additional limitation was separately recorded.
Key claims
GEM-3 trades short-range detail against rollout stability by selecting among trained timesteps without changing weights.
Evidence: source-2026-08-07-019
Sources
- arXiv preprint 2608.06241arXiv · primary research
Corrections
No corrections have been recorded for this story.