acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Large language model (LLM) agents increasingly combine reasoning, tool use, and action, but most evidence comes from episodic tasks with relatively immediate feedback, where episode resets often contain failures. Long-running physical control operates in a different regime: actions alter future states, errors compound across decisions, and an agent must improve from experience without being allowed to rewrite the physical rules that make execution safe. We study this regime through irrigation, where daily decisions interact with soil-water dynamics over entire growing seasons. We present Mímir, a physics-grounded LLM agent organized around two repair timescales. At the fast timescale, a structured physical interface and deterministic simulator turn an LLM output into a proposal that is numerically checked, revised, and subjected to bounded deterministic action selection before execution. At the slow timescale, recurrent failure patterns are consolidated into persistent contextual principles that condition future proposals, while the physical model, evaluator, and execution constraints remain immutable. Under a common retrospective evaluator across multiple sites, crops, and years, Mímir attains the lowest reported aggregate control cost among the evaluated references and uses about 51% less irrigation than the historical schedule replay. Matched ablations show higher control cost when forward simulation, verified revision, or persistent context is removed. Model-scale experiments show no monotonic improvement with increasing LLM size, while model-family experiments indicate broadly stable performance across the tested backbones. The resulting lesson is not that language models should replace numerical controllers, but that persistent physical agents can combine semantic reasoning with bounded, evidence-driven self-improvement while reserving physical prediction and actuator authority for explicit numerical mechanisms.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physics-Grounded LLM Agents, Long-Horizon Physical Control, Physical AI, Simulator-Grounded Reasoning, Cross-Episode Adaptation, Runtime Assurance, Irrigation Control, Embodied Agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AV4eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents Mímir, an LLM-based irrigation controller. The LLM proposes an irrigation action and revises it using feedback from a numerical simulator. A deterministic search and execution gate then chooses the action that is actually applied. Across seasons, the system updates contextual memory, while keeping the physical model and execution constraints fixed. The evaluation uses retrospective agricultural data from several sites and crops. The authors compare against historical irrigation, heuristic controllers, and an RL baseline, and report ablations for simulation, proposal repair, execution shielding, and memory. They also examine different LLM backbones and a nine-season sequence.

Critical strengths
  1. I like that the LLM does not directly control execution. Its output is checked, used to guide candidate search, and passed through fixed constraints. The same boundary applies to memory: experience can change future proposals but cannot rewrite the physical model or relax the action bounds. This design is sensible for the setting.
  2. The repair and memory ablations are also useful. In particular, retaining the downstream numerical machinery when removing proposal repair is a better comparison than simply showing that the full system beats a zero-shot LLM. It gives some evidence that the proposal stage matters.
Critical weaknesses
  1. My main concern is the missing numerical-control comparison. Appendix E.5 describes a deterministic physics-search/MPC baseline, but I could not find its results. Since the final action is already selected through numerical rollout and search, this is the comparison I most wanted to see. Does the LLM provide a better search anchor than a simple rule, previous action, or a reasonably dense numerical search? The current ablations do not answer that. Without this result, I cannot judge whether the language-and-memory layer adds enough to justify the paper’s central contribution.
  2. My second concern is that I cannot reconstruct the main evaluation from the submission. Appendix D leaves the objective weights, statistical unit, and sample count for later insertion. The irrigation aggregation is also unclear to me. More importantly, I need to understand what happens to soil state after a controller chooses an action different from historical irrigation. Does each controller evolve its own simulated trajectory, or are recorded soil observations subsequently used to reset or correct it? That distinction matters for whether the experiment measures the accumulated consequences of the controller’s decisions. These are missing explanations, not evidence that the results are wrong. But they prevent me from treating the headline comparison as sufficiently established.
Questions for authors
  1. Can you report the deterministic physics-search/MPC results described in Appendix E.5, including a comparison with the same simulator-call budget? This result is most likely to change my recommendation.
  2. Please provide the exact objective weights, stress threshold, sample counts, and meaning of the reported uncertainty. Are the irrigation values totals over an evaluated trajectory or averages over days/windows?
  3. After counterfactual irrigation actions, how is the next soil state obtained? Please clarify which historical measurements are reused and whether they can overwrite the effects of earlier controller actions.
  4. For the memory ablation, what principles were actually stored, when were they updated, and what were the update/reset rules? I would also like to know whether fixed guidance gives a similar benefit, though that would be an additional experiment rather than a clarification.
Reviewer_AccQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an LLM agent named "Mimir" for long-horizon irrigation control. It separates the LLM's reasoning role from physical authority. Based on a soil-water simulator, a deterministic selector, and safety constraints, the irrigation controller determines what is executed. A slower loop distills recurring errors into persistent context without altering the physical model or safety rules. On multi-site field data, Mímir achieves the lowest control cost among the evaluated baselines and uses about 51% less water than historical schedules. Ablations confirm that each component contributes to performance.

Critical strengths
  1. The paper targets long-horizon control where errors compound and feedback is delayed.
  2. The evaluation uses real field data
Critical weaknesses
  1. The framework relies heavily on a calibrated physical simulator, with the first 60% of each chronological segment used for calibration or training. This raises two concerns: (1) calibrating the simulator on the same environment may limit the evaluation to the current data distribution rather than demonstrating generalization to unseen environments; and (2) environment-specific calibration may introduce prior information about the evaluation environment. The authors should clarify the calibration protocol and discuss potential information leakage or evaluation bias.
  2. The current experiments mainly validate the proposed mechanisms indirectly through control cost and irrigation amount. Representative cases should be provided to explicitly show how physical evidence corrects an LLM proposal in the fast loop and how recurring experience updates the persistent context in the slow loop. This would provide more direct evidence that the two-timescale mechanism performs meaningful and correct updates.
  3. As experience accumulates across seasons, the persistent context may continuously grow and potentially degrade LLM reasoning or exceed its effective context window.
  4. As stated around Line 195, the current formulation mainly considers root-zone water states and irrigation depth as the control action. This relatively simple state/action space may make physical consistency easier to maintain. The authors could consider a more complex setting, such as joint irrigation and fertilization control, to better demonstrate the extensibility of the proposed framework.
  5. The content around Line 877 appears incomplete or is not fully displayed. Please check and correct the corresponding formatting.
  6. The implementation code is not publicly available.
Questions for authors

How long-term memory is compressed, prioritized, merged, or discarded? How memory size affects performance over long horizons?

Reviewer_zZM4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Mímir is an LLM irrigation controller with two timescales. Fast loop: a bucket soil-water model supplies structured state and a feasible action interval; an LLM "Farmer" proposes, a programmatic "Physicist" verifies and revises (up to three rounds), a deterministic selector rolls out a bounded candidate set and picks the lexicographic best, and runtime assurance gates execution. Slow loop: recurring failures are distilled into persistent context via ACE, with physics, evaluator and safety bounds fixed. On about 19 crop-years of retrospective field data (Michigan, Texas, Colorado) Mímir reports the lowest control cost among the compared references (100.94 ± 1.80 vs. MAD-50 134.33 and historical replay 790.51) and 51% less irrigation than historical replay. Ablations removing the simulator, decision repair, persistent context or execution shield raise cost, and results are similar across LLM sizes and families.

Critical strengths
  1. Principled, deployable design. Physical state transition, feasible set and execution authority stay outside the LLM, which only shapes proposals and context. The paper is careful in scope (no dominance claim over MPC, water read jointly with cost, nine-season run described as stability, not learning).
  2. Real multi-site field data with a chronological split and explicit rules against future reads, which is more credible than purely synthetic evaluation, plus component ablations.
Critical weaknesses
  1. The LLM's contribution is not isolated, and the evidence suggests it may be small. The deterministic physics-search/MPC controller on the same model (App. E.5) is called "diagnostic" and no numbers are reported; the LLM-only result is excluded for using a different metric (App. J). So neither "search without LLM" nor "LLM without search" is compared under the final metric, although the selector already searches a bounded candidate set with the same rollouts. Scale and family sweeps are within noise (8B 103.5 ± 18, 14B 96.2 ± 12), which the paper reads as "the architecture matters" but equally indicates the LLM matters little. Baselines also look weak (DRLIC applies 0.35 mm and costs 924, no worse than historical replay; no tuned MPC in the main comparison).
  2. The evaluation is partly self-referential and under-specified. Cost is computed on trajectories from the same lightweight bucket model that Mímir uses for rollouts and whose rollout cost its selector minimizes; the baselines do not optimize it. "Historical replay costs 790" describes how the model scores past practice, not an agronomic outcome, and no yield, observed soil moisture or independent crop model is used. The cost weights are not given, "±" is undefined, ablation slices have different "full" values (103.0, 109.1, 147.2, 149.5 vs. 100.94) without CIs, and the memory trigger/budget/reset settings are an unfilled instruction in Appendix H, so the results cannot be reproduced or checked for sensitivity.
Questions for authors
  1. What do the "±" values denote (seeds, sites, years, LLM samples), and what are the irrigation units (mm per day, season, site)?
  2. What are cost and irrigation for the search-only/MPC controller and for a no-LLM Mímir (fixed anchor) under the main protocol?
  3. Is the evaluator's trajectory generated by the same bucket model as the controller's rollouts? What are results against observed soil moisture or an independent crop model, and how sensitive is the ranking to the cost weights?
  4. Would static hand-written guidelines match the learned ACE principles?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.