BenchmarkThree tiers · eight settings

A Three-Tier Generalization Benchmark

AgriManager systematically evaluates how reinforcement learning policies generalize across changing environments and task interfaces. Eight settings, organized into three tiers, give a structured testbed for understanding where policy generalization succeeds and where it breaks.

Tier 1

Within Schema

Can a policy generalize to unseen environments when the task interface remains fixed?

Observation, action and reward schemas remain fixed. Context changes.

SettingTrainTestSimulatorRows (test)
Weather regimeNormal weatherDrought, wet, hot, coldWOFOST1,280
Crop variety4 / 8 / 16 seen crops4 held-out cropsWOFOST1,024
Price regimeTypical pricesHigh corn or high soybean priceCycles384
Tier 2

Single-Axis Schema

Can a policy adapt when one part of the task interface changes?

One schema axis is targeted at a time. The native action menus also involve wrapper differences.

SettingTrainTestSimulatorRows (test)
Observation schemaFull schemaSubset, renamed, superset, anonymized fieldsWOFOST1,024
Action menuA: lnpkw, lnpk · B: lnpk, lwA: lnw, ln, lw · B: lnpkw, lnw, ln (+ each family's seen menu)WOFOST512
Reward formulationMulti-objective (yield, profit, water) or yield onlyYield, profit, water, nutrientWOFOST512
Tier 3

Compositional Schema

Can learned policies transfer across simulators or recombine previously seen task components?

Cross-simulator transfer changes dynamics alongside the schema. Unified-policy tests withhold complete tuples while retaining component coverage.

SettingTrainTestSimulatorRows (test)
Cross-simulator transferOne source simulator3 × 3 source–target pairsall three386
Unified policySeen schema tuplesFour held-out schema combinationsall three1,156

Settings and row counts as in the paper's Table 1; the eight test sets sum to the 6,278-row test suite.

Examples

Real scored test episodes from AgriManager policies: what the policy observed, what it wrote, and what it did.

Full episodes in DSSAT, WOFOST and Cycles

Four real scored test episodes, decision by decision. Scrub the season, press play, or click the chart. The same backbone before and after RL on a Cycles crop-planning scenario is on the Results page →

What a schema shift looks like in the prompt

Excerpts from real AgriManager prompts for one WOFOST maize season. An illustration of the interface, not a result.

Before
After shift

Evaluation protocol

6,278test-suite rows: every policy is scored on the same frozen, held-out evaluation rows (rows, not independent physical scenarios)
3training seeds per method, each scored at its terminal checkpoint with no selection on validation
2reference policies: no-action (0) and a scripted agronomic reference (1) that anchor normalized return

Frozen held-out test suite

All headline scores come from a prespecified suite of 6,278 evaluation rows, built after submission and disjoint from each experiment's training and validation data. Rows are not independent physical scenarios.

Three seeds, terminal checkpoints

Each LLM-RL arm and neural baseline trains three seeds. Every run is scored at its terminal checkpoint with no selection on validation; we report mean ± SD over seeds.

Normalized return

R̃(π) = (R(π) − Rno-action) / |Rref − Rno-action|

No-action sits at 0 and the scripted agronomic reference at 1. The gap varies across targets, so policies are compared by R̃ only within a target.

Baselines

The untrained model (the same backbone before RL), neural-network PPO with the interaction budget of LLM-RL on the Tier 1 settings, the scripted agronomic reference, and no-action.

Confirmatory and exploratory results

Three-seed test-suite results are confirmatory. Results only on validation or from a single seed (training curves, backbone comparisons) and the per-crop breakdown are exploratory and labeled as such.

Further details are in the documentation →