Zephyrus-R1: Agentic Reinforcement Learning for Weather Reporting

Fengfei Yu; Marshall Fisher; Isadora White; Sumanth Varambally; Yasaman Jafari; Shivam Yadav; Samuel M. Bartlett; Shawn Roj

cs.LGOct 19, 2026

Zephyrus-R1: Agentic Reinforcement Learning for Weather Reporting

Authors: Fengfei Yu, Marshall Fisher, Isadora White, Sumanth Varambally, Yasaman Jafari, Shivam Yadav, Samuel M. Bartlett, Shawn Roj,

Organizations: University of California, San Diego

Abstract

Reinforcement learning (RL) has emerged as a key method for training AI models in scientific applications, but RL remains difficult when outputs cannot easily be verified against a known ground-truth answer. Natural-language weather forecast reports are one such domain: they translate atmospheric data into assessments of weather evolution, hazards, and uncertainty for both the general public and key decision-makers. Unlike numerical weather-analysis questions, they admit many valid outputs, which limits deterministic verification. To address this, we develop Zephyrus-R1, a practical framework for RL on scientific weather forecast reporting. We focus on three key components of the agentic RL pipeline: reward design, curriculum learning, and an analysis of human-expert alignment. To encourage both report legibility and factual accuracy, we design a reward function that combines a meteorological rubric with bidirectional claim entailment, both evaluated against National Weather Service reference reports. Our framework optimizes this reward in a simulated tool-calling environment, enabling vision-language agents to gather and analyze forecast evidence across regions and lead times before producing a report. We use annotations from three meteorologists to audit the claim labels produced by our LLM judges. Finally, a curriculum helps small models first develop tool use and scientific reasoning through RL on verifiable weather-analysis questions, then optimize the open-ended report reward. After training Qwen3.5-4B with Zephyrus-R1, we improve the mean across six judge-metric scores by 26% over the base model and achieve results competitive with frontier open-weight models. For Qwen3.8-27B, we improve the same mean by 13% over the base model, by 6.7% over Kimi-K3, and by 2.7% over GPT-5.6-Luna.

Figures & tables

Figure 1: Zephyrus-R1 framework. Agents call weather tools, inspect returned data or errors, and produce answers or reports. Stage 1 rewards verified answers; Stage 2 scores reports on a frozen rubric and Discussion F1 via an LLM judge. Curriculum RL transfers the Stage-1 policy to report generation.

Comparison	Acc.	$\kappa$	Within-1
GPT-OSS-Expert 1	0.370	0.097	0.920
GPT-OSS-Expert 2	0.630	0.346	0.960
GPT-OSS-Expert 3	0.570	0.226	0.960
Expert 1 -Expert 2	0.480	0.224	0.900
Expert 1 -Expert 3	0.460	0.214	0.860
Expert 2 -Expert 3	0.570	0.306	0.890

Table 1: Claim-label agreement on 100 claim-report pairs. Accuracy (Acc.) is exact label agreement; Within-1 excludes only SUPPORTED-REFUTED reversals. For majority-vote comparisons, three-way ties are assigned NEUTRAL.

Figure 2: Verifier accuracy on held-out validation weather-analysis questions for the base policy and the RSFT, RL-only, and RSFT-RL fine-tuned policies: (a) Qwen3.5-4B; (b) Qwen3.8-27B.

Figure 3: Tool-use diagnostics on the report-stage test set during report RL. Failure rate is shown for 4B (a) and 27B (b), and Tools/exec. for 4B (c) and 27B (d). Appendix C.7 describes how both diagnostics are computed.

Model	Route	Rub.		F1		Rub.		F1		Rub.		F1		Avg
		GPT-OSS [-1pt] 120B				Gemma-3 [-1pt] 27B-IT				Mistral-Small [-1pt] 3.2-24B
Qwen3.5-4B	Base	0.082		0.138		0.507		0.621		0.391		0.342		0.347
	Direct RL	0.123	+51%	0.217	+57%	0.575	+13%	0.693	+12%	0.453	+16%	0.477	+40%	0.423	+22%
	Curriculum RL	0.129	+57%	0.231	+67%	0.514	+1%	0.768	+24%	0.441	+13%	0.548	+60%	0.438	+26%
Qwen3.8-27B	Base	0.136		0.224		0.735		0.715		0.690		0.537		0.506
	Direct RL	0.167	+23%	0.318	+42%	0.742	+1%	0.770	+8%	0.719	+4%	0.631	+18%	0.558	+10%

Table 2: Weather-report results on the test events for the base policies and two Zephyrus-R1 training routes. Every score averages three complete scoring passes; higher is better for both metrics, and bold marks the within-scale best. Green percentages give the relative change from the same-scale Base, computed from unrounded scores.

#	Model	Rub.	F1	Rub.	F1	Rub.	F1	Avg
		GPT-OSS [-1pt] 120B		Gemma-3 [-1pt] 27B-IT		Mistral-Small [-1pt] 3.2-24B
1	Qwen3.8-27B + Zephyrus-R1	0.260	0.312	0.778	0.736	0.728	0.615	0.571
2	GPT-5.6-Luna	0.170	0.297	0.713	0.801	0.698	0.660	0.556
3	Kimi-K3	0.168	0.286	0.736	0.732	0.721	0.570	0.536
4	DeepSeek-V4-Flash-Vision-Exp	0.153	0.262	0.706	0.729	0.689	0.570	0.518
5	GLM-5.3-Flash	0.156	0.224	0.728	0.726	0.699	0.567	0.517

Table 3: Leaderboard on the test events, ordered by Avg (higher is better). Every score averages three complete scoring passes. Shaded rows are Zephyrus-R1 models trained with Curriculum RL. Avg is the mean of the six unrounded scores.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Question template	Answer	Required tool
Geographic feature with the highest or lowest time-mean field value	location	dataset
Minimum, maximum, mean, or median field value over a named geographic feature	numerical	dataset
Sublocation within a named parent feature with the highest or lowest recorded field value	location	dataset
Hours from the initial time at which a named feature reaches its maximum or minimum field value	time	dataset
Minimum, maximum, mean, or median field value over a named feature at a specified time	numerical	dataset
Geographic features above or below a specified climatological percentile	anomaly list	climatology

Appendix

Table B.1: The eight ZephyrusBench question templates used in Stage 1 and the tool required by each template.

Type	Required content	Score 1.0	Score 0.5	Score 0
Numerical	one JSON number	relative error $\leq 1\%$ (absolute error $\leq 0.05$ if $\|\mathrm{gold}\|<1$ )	relative error $\leq 5\%$ (absolute error $\leq 0.15$ if $\|\mathrm{gold}\|<1$ )	otherwise
Time	one JSON number in hours	exact	within 6 h	otherwise
Location	JSON object with place name and value	normalized name equals or contains the gold name	wrong name, value inside the numerical 1.0 band	otherwise
Anomaly	JSON list of names (empty if none)	set F1 over normalized names, graded in $[0,1]$	(graded)	no overlap

Appendix

Table B.2: Stage-1 answer formats and verifier scores. Numerical questions use absolute-error bands when the gold magnitude is below one. Anomaly lists receive graded set F1.

Figure B.1: Stage-1 accuracy on the 125-question held-out validation set for Qwen3.5-4B and Qwen3.8-27B. Filled circles show evaluated checkpoints, while open circles at step 0 mark each route’s initial policy. RSFT and RL-only start from the base policy, whereas RSFT-RL starts from the selected RSFT checkpoint. Gray dotted lines show base-policy accuracy. In the RSFT panels, the red open circle and dotted line mark the best checkpoint.

	Verifier			Accuracy by answer type
Checkpoint	Acc.	Acc. reward	Tool	Num.	Loc.	Time	Anom.
Qwen3.5-4B	0.394	0.394	0.996	0.334	0.400	0.392	0.522
RSFT (step 546)	0.498	0.496	0.988	0.467	0.521	0.457	0.595
RL-only (step 44)	0.572	0.572	0.998	0.549	0.562	0.526	0.709
RSFT-RL (step 32)	0.521	0.521	1.000	0.473	0.512	0.491	0.688
Qwen3.8-27B	0.648	0.648	0.998	0.571	0.637	0.685	0.785

Appendix

Table B.3: Stage-1 results on the held-out validation set. Acc. averages the per-answer score in Table B.2 . Acc. reward averages the reward in Equation 3 , including the tool check and turn penalty. Tool is the required-tool access rate. The remaining columns give mean scores by answer type. Bold marks the within-scale best.

Setting	Value	Role
Rejection sampling
Base-policy rollouts per question	8	candidate trajectories for RSFT
Sampling	temperature 0.8	diversify successful trajectories
Retention cap	2 per question	limit repeated successes
Shared RL optimization
Sampling	temperature 0.6, top- $p$ 1.0	rollout sampling

Appendix

Table C.1: Training and evaluation settings. Shared RL settings apply to Stage-1 RL and report RL, and report-stage settings apply to report RL under both routes.

Figure C.1: Online rubric and Discussion F1 weights during report RL for both routes and model scales. Each run extends through step 163, and the two weights sum to one.

Criterion	Definition
Setup Fidelity & Key Feature Retention (Additions AND Omissions)	Accurately represents the specific synoptic/mesoscale features and stated hazards/corridors in the setup, without introducing unsupported details and without dropping any primary stated feature/hazard that changes the forecast story.
Spatiotemporal Specificity (Progression When Present; Persistence When Static)	Communicates where and when impacts occur with specificity appropriate to the setup: describes progression tied to moving features when the setup is progressive, or explicitly states persistence/limited movement when the setup is quasi-stationary.
Impact/Hazard Extraction & Prioritization (Reference-Weighted)	Identifies the setup’s main hazards and impact corridors (and any stated relative maxima/minima) and ranks them appropriately, using practical, decision-relevant framing without adding new hazards.
Mechanism-to-Weather Reasoning (Sufficient, Setup-Consistent)	Uses meteorologically correct, setup-consistent causal links (forcing, moisture transport, instability, terrain/onshore/upslope, deformation/lake-effect) to explain why impacts occur where/when described. Depth is rewarded but not required for a high score if the forecast is a faithful paraphrase.
Thermal/Airmass & Precip-Type Plausibility (No Out-of-Pattern P-Type)	Maintains a coherent temperature/airmass narrative consistent with the setup and seasonally plausible precipitation type statements (snow/ice/convective vs stratiform), avoiding thermodynamic regime shifts not supported by the setup.
Constraint-Respecting Detail & Calibration (No Invented Timing/Numbers/Pattern Labels)	Uses specificity only when supported: avoids invented cutoffs (“ends by Saturday evening”), ungrounded pattern labels (“split flow”) or pseudo-precision, and matches the setup’s confidence level (possible/expected/likely).

Appendix

Table C.2: The six meteorological rubric criteria supplied to the judge. Anchored integer scores are normalized to

[0,1]

Panel of Figure 14 — Figure C.2: Complete scoring anchors for the six meteorological rubric criteria. The judge receives the anchor descriptions shown for levels 1, 3, 5, 8, and 10. Intermediate integer scores are allowed.

Figure D.1: Report-RL F1 and rubric-score trajectories for (a) Qwen3.5-4B and (b) Qwen3.8-27B across the three evaluation judges. Each point averages three scoring passes on the 197 test events.

Table 16

Panel of Figure 17 — Figure E.1: Report from Figure with every Gemma claim labeled SUPPORTED ( S ), REFUTED ( R ), or NEUTRAL ( N ). SUPPORTED and REFUTED claims are shaded green and red, and the reference sentences that support or refute them carry the same tag and shading.

Panel of Figure 18 — Figure E.2: Report from Figure with every Mistral claim labeled SUPPORTED ( S ), REFUTED ( R ), or NEUTRAL ( N ). SUPPORTED and REFUTED claims are shaded green and red, and the reference sentences that support or refute them carry the same tag and shading.

RL for VLMs

Reinforcement learning for vision-language models aligns multimodal reasoning, tool use, and action with task-specific feedback.

See all 472 papers

AFDBench: A Reasoning-First AI Scientist for NationalWeather Service Forecast Discussions
Aug 25, 2026Manmeet Singh, Somnath Luitel, Prabhjot Singh +3Forecasting Benchmarks Language Model Generation Evaluation
Department of Earth, Environmental, and Atmospheric Sciences, Western Kentucky University, Bowling Green, KY, USA · Department of Computer Science, University of Texas at Austin, Austin, TX, USA · RediMinds Inc., Southfield, MI, USA.
WeatherSyn: An Instruction Tuning MLLM For Weather Forecasting Report Generation
May 8, 2026Zinan Zheng, Yang Liu, Nuo Chen +3Multimodal Large Language Models Weather Forecasting
Hong Kong University of Science and Technology (Guangzhou) · The Chinese University of Hong Kong · HY Founda- tion Model Team, Tencent +2
Agent-Based ML-LLM Fusion with Self-Optimizing Prompts for Plateau Weather Alerts
Sep 9, 2026Shuai Yan, Yang Xu, Shan HePrompt Optimization Weather Forecasting
Chengdu Jincheng College College of Computer and Software Chengdu, Sichuan, China

Aug 25, 2026J/K move · Enter open · S save

AFDBench: A Reasoning-First AI Scientist for NationalWeather Service Forecast Discussions

Manmeet Singh, Somnath Luitel, Prabhjot Singh, Manraaj Banga+2

Department of Earth, Environmental, and Atmospheric Sciences, Western Kentucky University, Bowling Green, KY, USA · Department of Computer Science, University of Texas at Austin, Austin, TX, USA · RediMinds Inc., Southfield, MI, USA.

Forecasting Benchmarks Language Model Generation Evaluation Weather Forecasting Scientific Reasoning in Language Models

Open paper AskPDF arXiv

Large language models (LLMs) hallucinate numerical values when generating high-stakes meteorological text, posing risks for weather communication. We present AFDBench, an AI meteorologist that generates professional Area Forecast Discussions (AFDs) by reasoning through structured AI weather forecast data from Google's WeatherNext 2. We introduce AFDBench, the first benchmark for evaluating generative meteorological reasoning, comprising 7,732 expert written discussions from 13 National Weather Service (NWS) offices paired with real AI weather forecast inputs, and three complementary metrics: Met-Align (numerical accuracy), Style-Align (professional dialect adherence), and Input-Grounding (fidelity to source weather data). Zero-shot evaluations reveal that open-source LLMs achieve low Style-Align (~0.33) and moderate Input-Grounding (~0.88), failing to write in the professional NWS register or faithfully use their input data. We apply Group Relative Policy Optimization (GRPO) with domain-specific rewards targeting temperature accuracy, synoptic correctness, and format compliance. On 1,033 held-out samples from two unseen NWS offices, GRPO nearly doubles Style-Align from 0.318 to 0.619 and improves Input-Grounding from 0.881 to 0.940, demonstrating that reinforcement learning teaches a 7B-parameter model to write like a professional meteorologist and faithfully interpret AI weather data.

Abstract

RL for VLMs

Explore similar work

AFDBench: A Reasoning-First AI Scientist for NationalWeather Service Forecast Discussions

WeatherSyn: An Instruction Tuning MLLM For Weather Forecasting Report Generation

Agent-Based ML-LLM Fusion with Self-Optimizing Prompts for Plateau Weather Alerts