Workshop: 4th Annual Trillion Parameter Workshop: Building Open AI Infrastructure, Models, and Agentic Systems for Science
Authors: Carlo Siebenschuh, Brian Hsu, and Kyle Hippe (University of Chicago, Argonne National Laboratory) and Arvind Ramanathan, Ian Foster, and Rick Stevens (Argonne National Laboratory, University of Chicago)
Abstract: Temporal data are central to science, yet post-training methods that drive major advances in language modeling remain underexplored for time series foundation models (TSFMs). Reinforcement learning with verifiable rewards can improve model capabilities but tends to narrow the model's output distribution, causing overconfident probabilistic forecasts. We resolve this tension by combining group relative policy optimization (GRPO) with rewards based on proper scoring rules. Scoring-rule GRPO elicits calibrated predictive distributions: trajectory-level rewards improve general-purpose forecasting, whereas functional-level rewards specialize the predictive distribution for a task-specific quantity. Across diverse dynamical systems data, post-training with the energy-score reward reduces RMSE by 54% and in-distribution functional forecast error by 57% relative to the pre-trained model. These results show that proper scoring rules are promising rewards for temporal reasoning. Because the prediction quality of this post-training strategy transfers across tasks, it moves TSFMs closer to being truly foundational for scientific applications.
Back to 4th Annual Trillion Parameter Workshop: Building Open AI Infrastructure, Models, and Agentic Systems for Science Archive Listing
Back to Full Workshop Archive Listing