Skip to content

Back to the catalog

Mimir: Physics-Grounded LLM Agents for Long-Horizon Irrigation Control
arxiv.orgpaper

Mimir: Physics-Grounded LLM Agents for Long-Horizon Irrigation Control

Mimir is a physics-grounded LLM agent for long-horizon irrigation control that pairs a fast-timescale loop — where a structured physical interface and deterministic simulator verify, revise, and bound LLM-proposed actions before execution — with a slow-timescale loop that consolidates recurrent failure patterns into persistent contextual principles, while keeping the physical model, evaluator, and execution constraints immutable. Evaluated across multiple sites, crops, and years with a common retrospective evaluator, Mimir achieves the lowest aggregate control cost among tested baselines and uses about 51% less irrigation than historical schedule replay; ablations show forward simulation, verified revision, and persistent context are all necessary, and scaling LLM size or switching model families gives no monotonic improvement.

TL;DR A physics-grounded LLM agent for season-long irrigation control that verifies actions with a deterministic simulator and accumulates persistent failure-derived principles, cutting irrigation use ~51% versus historical schedules.
Open original ↗
Document

What it claims

In Mimir, an LLM proposes irrigation decisions but never issues the actuator command itself. Numerical tools and deterministic logic decide what actually runs. The evaluation is retrospective, using field-observed data from Michigan, Texas and Colorado across multiple crops and years, all scored by one common evaluator. Mimir reaches an aggregate control cost of 100.94 ± 1.80 with 2.58 ± 0.12 mm irrigation. Historical schedule replay uses 5.32 ± 0.57 mm, which is where the ~51% saving comes from. The authors say outright that they are not asking whether an LLM should replace MPC, FAO-56 scheduling or RL. Their question is which responsibilities can be adaptive and which must stay numerical and externally enforced.

How it works

The state comes from a fixed water balance: S(t+1) = S + P + I − ET − D. Here S is root-zone storage, P precipitation, I irrigation, ET evapotranspiration and D drainage. The LLM sees a structured context containing storage, field capacity, wilting point, crop stage, ET and the forecast. Units and admissible quantities are fixed before the LLM is queried.

Hard action bound. I_max = max(0, FC − S − forecast effective rain). The final command is projected into [0, I_max], and the LLM cannot widen that interval.

Fast loop (daily).

  1. A "Farmer" LLM role proposes an action.
  2. A "Physicist" role verifies it programmatically.
  3. The Farmer revises.

The loop stops when verification no longer changes the proposal or a fixed round budget runs out. The repaired proposal is only an anchor. A deterministic selector builds a finite candidate set around it, always including zero irrigation and the refill bound, and rolls each candidate out over the horizon. It then picks lexicographically: fewest modeled physical violations first, lowest rollout cost second. Runtime assurance and a trusted fallback path come after that.

Slow loop (per episode/season). Persistent context changes only when a configured evidence trigger fires on recurring failures, such as repeated over-reaction to forecasts or repeated verifier corrections. Several things never change:

  • model weights
  • the transition model
  • the evaluator
  • the action bounds
  • the execution gate
flowchart TD
  O[Observations] --> S[Physical state + action bound]
  S --> P[LLM propose]
  P --> V[Simulator verify]
  V -->|changed & budget left| R[LLM revise] --> V
  V -->|stable or budget out| C[Candidates around anchor + zero + refill]
  C --> SEL[Lexicographic select: violations, cost]
  SEL --> RA{Runtime assurance}
  RA -->|pass| A[Project to bound, actuate]
  RA -->|fail| F[Trusted fallback]
  A -.season outcomes.-> T{Evidence trigger?}
  T -->|yes| M[Update persistent context]
  M -.-> P

Evidence

  • Multi-site comparison: Mimir has the lowest aggregate control cost among the evaluated references (figures above).
  • Ablations: removing forward simulation, verified revision or persistent context each raises control cost. The fetched text gives no per-ablation numbers.
  • Nine-season sequence: results vary substantially from season to season but show no sustained divergence.
  • LLM scale and family: performance stays "in the same broad range" across sizes and families, with no monotonic gain from bigger models. The authors take this to mean the surrounding architecture drives behaviour.

Limits

  • Retrospective only. The evaluation replays field data through a common evaluator, and no live deployment is reported. Because the controller's own water balance is described as "lightweight", check how far the evaluator's model departs from it before trusting the cost figures.
  • Truncated text. The page cuts off partway through §4.3, so it does not include:
  • the slow-loop trigger details
  • the ablation tables
  • the specific LLMs and sizes tested
  • the list of baselines
  • Appendix B (the physical abstraction's assumptions)
  • Unclear irrigation unit. The page does not say what the mm figures are normalised over.
  • Nothing to adopt directly. No code, prompts, licence or run instructions appear in the fetched text.
  • Narrow scope by design. Learning is limited to context edits. The physical model is never revised online, so the approach depends on the fixed model being adequate for the site.

Technologies named here

  • Mimir

Techniques named here

  • Fast-timescale physical repair
  • Slow-timescale contextual consolidation
  • Physics-grounded agent architecture

Related in agents