NeurIPS 2026 · Evaluations & Datasets Track

AgriManager

A Framework and Benchmark for LLM‑RL Generalization in Agricultural Management

Chi Gui1, Junwen Zheng1, Ritwik Nigam1, Qianlan Yang1, Meagan Lang1, Vardhan Dongre1, Jeremiah Barr2, Yu-Xiong Wang1, Vikram S. Adve1 1University of Illinois Urbana-Champaign · 2Corteva Agriscience

Observe
Think
Act
Harvest
vs the untrained model on the same field

Beyond Distribution Shift: Schema Shift

Traditional RL generalization focuses largely on changing environments under a fixed task interface. But what happens when the interface itself changes?

Distribution shift

Same interface, different context

  • WeatherNormal→Drought
  • CropSeen→Unseen
  • PriceTypical→High corn price
Schema shift

The task interface itself changes

  • ObservationFull→Partial
  • ActionN + Water→N only
  • RewardYield→Profit

We formalize schema shift as a change in the policy-facing schema Σ between training and deployment.

Same field, different prompt

See the real prompts

Why Study Schema Shift in Agriculture?

Farms vary in their task interface, and every decision plays out over a season or years.

Varying task interfacesFarms differ in what they observe, can do, and aim for.
Long-term consequencesDecisions pay off at harvest, or years later.
Simulation-driven evaluationDSSAT, WOFOST and Cycles allow reproducible training and testing.

Two task families

Crop growthWOFOST · DSSAT · Cycles
SowingHarvestreward

Manage one field through a growing season.

Observation
Crop stage and biomass, soil nutrients and water, inputs applied so far, weather
Action
Apply N, P or K fertilizer, irrigate, or wait, every few days
Objective
Profit by default; also yield, water or nutrient stewardship
Crop planningCycles
Year 1Year 19revenue

Plan a multi-year crop rotation on one field.

Observation
Soil nitrogen state, current crop prices, years remaining, past crops and revenue
Action
Which crop to plant this year, and in which week
Objective
Gross revenue (yield × price), summed over the rotation
We support20+ cropsVaried seasons: drought, hot, cold, wetVaried crop pricesMultiple goals: profit, yield, water, nutrients

A Three-Tier Generalization Benchmark

Does an RL-trained LLM policy still generalize under schema shift? Each tier adds more: none, one component, or several at once.

Tier 1

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)

WeatherCrop varietyPrice
Fixed-interface NNLLM
Tier 2

Single-Axis Schema

One schema component shifts. One of measurements, management options, or goals changes.

One of O, A, R changes

ObservationAction menuReward
Fixed-interface NNLLM
Tier 3

Compositional Schema

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

Cross-simulatorUnified policy
Fixed-interface NNLLM

All results are scored on a frozen, held-out test suite. Benchmark details Evaluation protocol

Where Generalization Works — and Where It Breaks

Tier 1

Within a fixed interface, RL generalizes

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.

Tier 2

Different schema shifts, different bottlenecks

Observations

Renaming, adding or dropping secondary fields barely matters; removing the resource state (soil water and nutrients) breaks the policy.

Actions

The policy handles a new menu if it has seen those actions together before.1

Trained onNPKWNPK →NW Improves over the untrained model
Trained onNPKW →NW Falls below the untrained model
N, P, K: nitrogen, phosphorus, potassium fertilizer · W: irrigation water

1 Consistent with joint action coverage; menu wrappers also differ (paper §4.2).

Objectives

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.

Tier 3

Composition remains open

Switching simulators

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 →WOFOSTDSSATCycles
WOFOST▲ better▲ better1▼ worse
DSSAT▼ worse▲ better▼ worse
Cycles▼ worse▲ better1▲ better
Compared with the untrained model on the test simulator

1 Both gains off the diagonal go into DSSAT, where the untrained model scores 0 because it misses the output format.

Recombining components

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.

Held-out profit targets, one combination each
UntrainedOne policy, never saw this combinationSpecialist, trained on it
Each row has its own scale.

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