Metric-Aware Hybrid Forecasting for the CTF4Science Lorenz Challenge

Researchers have developed a system for the CTF4Science Lorenz challenge, a benchmark for chaotic-system prediction. The system achieved a score of 83.83551 on the public leaderboard, with a subsequent submission reaching 83.85529^[1].

The CTF4Science Lorenz challenge evaluates a system's ability to perform short-horizon forecasting, long-time distribution matching, and trajectory reconstruction. According to arXiv^[1], the challenge mixes these tasks across nine task pairs. However, another arXiv submission^[2] described the CTF-4-Science Lorenz benchmark as evaluating chaotic-system prediction across twelve hidden scores and five scenario families. The discrepancy highlights differing descriptions of the challenge's scope. The winning system employed a metric-aware hybrid approach, assigning different predictors to various metric families. A representative submission scored 83.83551 on the public leaderboard, while a follow-up submission reached 83.85529^[1]. In contrast, a different submission reported a final public score of 79.63^[2]. The submissions were made on June 2, 2026, as per the submission date mentioned in one of the arXiv papers^[1].