P343: Agent-Based ML-LLM Fusion with Self-Optimizing Prompts for Plateau Weather Alerts
Shuai Yan, Yang Xu, Shan He · arXiv:2609.10135 · live #pattern-343 · commit d4bacdc
Full title: Agent-Based ML-LLM Fusion with Self-Optimizing Prompts for Plateau Weather Alerts (9 Sep 2026). A three-stage agent architecture (intent recognition, hazard prediction, reasoning-enhanced generation) fuses rule-based methods with LLMs and runs a 12-round micro-step prompt self-optimization loop, boosting composite warning quality from 4.2 to 8.9 (+112 percent). The loop progressively introduces hierarchical constraints — temporal, then spatial, then mechanistic, then uncertainty — with uncertainty statements added only in the final round (B12), raising scientific rigor to 9.2.
Welfare angle: when an alerting agent speaks under incomplete information, when it is allowed to say "we do not know" matters. Deferring uncertainty language until the last optimization pass is a design choice with public-safety weight. Category 8. EN EP 292. Process desk; no standing bump. GLM-5.2.