Research map / Multi-agent systems and games / LLM agents

HVR-Met: A Hypothesis-Verification-Replanning Agentic System for Extreme Weather Diagnosis

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HVR-Met, a seven-agent system with a Hypothesis-Verification-Replanning loop and a knowledge-base-driven benchmark for automated extreme weather diagnosis, evaluated with GPT-5 and other LLMs.

Problem

Extreme weather diagnosis requires multi-step meteorological reasoning, dynamic tool use, and expert prior knowledge, which general agents and deep forecasting models lack; existing agent evaluations do not cover atomic meteorological sub-tasks.

Approach

A multi-agent system with a guideline library distilled from 584 papers, where a Diagnostician forms testable hypotheses about anomalous figures/indices, verifies them with computed evidence, and replans via the library when anomalies are missing, looping until five anomalous factors are confirmed or 50 turns are reached.

Key results

• GPT-5 achieves 4.70 hypothesis, 4.92 data, 4.07 index, 4.13 figure and 3.94 final report scores on a 0-5 scale; removing the Guideline Library drops final report to 0.40.
• Reported pass rates are 71.86% for index calculation, 79.52% for figure generation, and 85% for final reporting.
• Automatic LLM scoring correlates with 30 human-scored cases at Pearson r between 0.81 and 0.98.

Surprising findings

• Removing the Index or Figure Knowledge Base lowers the Hypothesis score, which the authors attribute to error propagation in the verification loop.
• Reasoning-enhanced models (Gemini-3-pro-thinking, DeepSeek-R1) outperform GPT-5 on hard index calculations (46.43% and 42.31% vs 31.25%).

Limitations

• Relies on reanalysis data; real operational forecast biases are not evaluated.
• Benchmark is semi-synthetic, with QA pairs generated by GPT and a VLM.
• Figure evaluation relies on a VLM answering binary questions, which adds judge dependence.

Experimental setup

Datasets: WeatherBench2 584 extreme weather papers (guideline library construction) MetPy documentation Cartopy/Matplotlib plotting scripts

Benchmarks: 100 extreme weather full-pipeline cases 150 index calculation QA (30 indices) 100 figure generation QA (20 figure types)

Metrics: LLM-rubric score 0-5 per stage pass rate (relative error < 0.05) VLM binary QA accuracy Pearson correlation with human scores

Baselines: Gemini-3-pro-preview-thinking Deepseek-R1-0528 Qwen3-Coder-480B-A35B-Instruct

Builds on: AG2

Closest papers by meaning

• ForecastCompass: Guiding Agentic Forecasting with Adaptive Factor Memory (NeurIPS 2026 · similarity 0.83)
• HR-Extreme: A High-Resolution Dataset for Extreme Weather Forecasting (ICLR 2025 · similarity 0.82)
• WeatherGFM: Learning a Weather Generalist Foundation Model via In-context Learning (ICLR 2025 · similarity 0.82)
• TimeClaw: A Time-Series AI Agent with Exploratory Execution Learning (NeurIPS 2026 · similarity 0.82)
• Language-driven All-in-one Adverse Weather Removal (CVPR 2024 · similarity 0.81)
• RadarQA: Multi-modal Quality Analysis of Weather Radar Forecasts (NeurIPS 2025 · similarity 0.81)