Within Schema
Distribution shift only. Weather, crop traits, or prices vary under the same task schema.
(O, A, R) unchanged · context changes (weather, crop, prices)
NeurIPS 2026 · Evaluations & Datasets Track
A Framework and Benchmark for LLM‑RL Generalization in Agricultural Management
Traditional RL generalization focuses largely on changing environments under a fixed task interface. But what happens when the interface itself changes?
Σ=(ΣOobservation,ΣAaction,ΣRreward)
We formalize schema shift as a change in the policy-facing schema Σ between training and deployment.
Farms vary in their task interface, and every decision plays out over a season or years.
Two task families
Manage one field through a growing season.
Plan a multi-year crop rotation on one field.
One RL framework for LLM agents across different crop simulators.
Does an RL-trained LLM policy still generalize under schema shift? Each tier adds more: none, one component, or several at once.
Distribution shift only. Weather, crop traits, or prices vary under the same task schema.
(O, A, R) unchanged · context changes (weather, crop, prices)
One schema component shifts. One of measurements, management options, or goals changes.
One of O, A, R changes
Several schema components shift at once. Across farms, measurements, management options, and goals differ.
Several of O, A, R change; dynamics P may change too
All results are scored on a frozen, held-out test suite. Benchmark details Evaluation protocol
With explicit reasoning, a Qwen3-4B policy trained by RL improves over the untrained model on every held-out weather regime, crop and price regime, and outperforms a neural-network PPO baseline trained with the same budget. Without reasoning, training often stalls near doing nothing.
Renaming, adding or dropping secondary fields barely matters; removing the resource state (soil water and nutrients) breaks the policy.
The policy handles a new menu if it has seen those actions together before.1
1 Consistent with joint action coverage; menu wrappers also differ (paper §4.2).
Trained on yield alone, the policy transfers to profit, water and nutrient goals it never saw. Trained on several goals at once, only the reasoning policy learns reliably.
A policy trained on one simulator improves on that simulator, but on another simulator it usually ends up worse than before training.
| Trained on ↓Tested on → | WOFOST | DSSAT | Cycles |
|---|---|---|---|
| WOFOST | ▲ better | ▲ better1 | ▼ worse |
| DSSAT | ▼ worse | ▲ better | ▼ worse |
| Cycles | ▼ worse | ▲ better1 | ▲ better |
1 Both gains off the diagonal go into DSSAT, where the untrained model scores 0 because it misses the output format.
One policy is trained on many simulator, action menu and goal combinations, then tested on a combination it never saw. On DSSAT it gets about 86% of the way to a specialist trained on that exact combination. On WOFOST and Cycles it stays below the untrained model.
Seeing the components of a task is not the same as mastering their combination.
Additional insightOn Qwen, reasoning mainly stabilizes RL training; this does not hold for every task or backbone.
All results and charts