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

Yimeng Liu; Mi Zhang; Younsuk Dong; Zhichao Cao

cs.AIOct 1, 2026

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

Authors: Yimeng Liu, Mi Zhang, Younsuk Dong, Zhichao Cao

Organizations: Michigan State University · Ohio State University

Abstract

Large language model (LLM) agents increasingly combine reasoning, tool use, and action, but most evidence comes from episodic tasks with relatively immediate feedback and reset 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 Mimir, 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 we numerically check, revise, and subject 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, Mimir attains the lowest reported aggregate control cost among the evaluated references and uses about 51% less irrigation than the historical schedule replay. The ablation study show higher control cost when forward simulation, verified revision, or persistent context is removed; model-scale and model-family studies show no monotonic gain from increasing LLM size. The resulting lesson show that persistent physical agents can combine semantic reasoning with bounded, evidence-driven self-improvement while reserving physical truth and actuator authority for explicit numerical mechanisms.

Figures & tables

Figure 1: Mímir Physical-Embedded Irrigation Control.

Representative Paradigm	Physical Grounding	Deliberative Reasoning	Persistent Adaptation	Independent Runtime Safety
Model-based irrigation Allen et al. (1998) ; Delgoda et al. (2016)	✓	–	$\circ$	✓
Learning-based irrigation Yang et al. (2020) ; Ding & Du (2024)	$\circ$	–	$\circ$	$\circ$
Embodied LLM agents Ahn et al. (2022) ; Driess et al. (2023) ; Yao et al. (2022)	$\circ$	$\circ$	–	–
Self-improving agents Shinn et al. (2023) ; Wang et al. (2023) ; Zhang et al. (2025)	–	✓	✓	–
Runtime-assured control Ames et al. (2019) ; Mehmood et al. (2022) ; Wachi & Sui (2020)	✓	–	–	✓
Mímir	✓	✓	✓	✓

Table 1: Positioning of Mímir relative to representative research paradigms. ✓denotes a primary mechanism,

\circ

denotes partial support, and – denotes that the capability is not a primary focus.

Figure 2: Mímir as a two-timescale physical agent. In the daily loop, observations are converted to a structured physical state; the LLM proposes or revises a decision; runtime assurance gates execution. In the slower loop, recurring failures are consolidated into persistent context. Adaptation only changes the reasoning context.

Panel of Figure 3 — Figure 3: Regional results for Michigan potato, Texas maize, Colorado maize, and Colorado sunflower environments. Nine-season Colorado sequence. This experiment evaluates operational stability, not monotonic self-improvement.

Controller	Control cost $\downarrow$	Irrigation (mm)
Mímir	$\mathbf{100.94\pm 1.80}$	$2.58\pm 0.12$
MAD-50	$134.33\pm 3.10$	$10.58\pm 0.44$
ET-Replace	$440.74\pm 12.48$	$12.70\pm 0.31$
DRLIC	$924.47\pm 27.50$	$0.35\pm 0.00$
Historical irrigation	$790.51\pm 17.76$	$5.32\pm 0.57$

Table 2: Aggregate performance under the common multi-site evaluation protocol. Control cost is the primary objective; irrigation is reported separately because lower water alone can correspond to higher stress.

	Full		Ablated
Removed component	Cost	Water	Cost	Water
Persistent context (ACE)	103.0	3.40	157.5	4.41
Decision repair	109.1	2.47	180.5	3.19
Execution shield	147.2	–	259.9	$\approx$ unchanged
Forward simulator	149.5	3.29	730.5	0.45

Table 3: Paired component ablations. Each row has its own matched full configuration; absolute full values differ across slices and must not be compared between rows. Water is in mm.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Environment	Crop	Resolution	Sequence	Use
Michigan/LOCOSMOS	Potato	Hourly	2021–2023 entries	Main evaluation
Bushland, Texas	Maize	Hourly	Multi-year	Main evaluation
Colorado LIRF	Maize	Hourly	5-year sequence	Main evaluation
Colorado sunflower	Sunflower	Hourly	9-season sequence	Long-horizon analysis

Appendix

Table 4: Dataset inventory used by the evaluation protocol.

Controller	Control cost	Irrigation (mm)
Mímir	$100.94\pm 1.80$	$2.58\pm 0.12$
MAD-50	$134.33\pm 3.10$	$10.58\pm 0.44$
ET-Replace	$440.74\pm 12.48$	$12.70\pm 0.31$
DRLIC	$924.47\pm 27.50$	$0.35\pm 0.00$
Historical irrigation	$790.51\pm 17.76$	$5.32\pm 0.57$

Appendix

Table 5: Aggregate results under the common multi-site evaluator. Irrigation is reported separately from control cost.

	Full		Ablated
Removed component	Cost	Water	Cost	Water
Persistent context (ACE)	103.0	3.40	157.5	4.41
Decision repair	109.1	2.47	180.5	3.19
Execution shield	147.2	–	259.9	$\approx$ unchanged
Forward simulator	149.5	3.29	730.5	0.45

Appendix

Table 6: Paired component ablations. Each row uses its own matched evaluation slice; effect sizes must not be compared across rows. Water is in mm.

Panel of Figure 4 — Figure 4: DeepSeek scale sensitivity under an otherwise fixed Mímir pipeline.

Panel of Figure 5 — Figure 5: Model-family sensitivity under a shared physical interface and evaluation sequence.

Panel of Figure 6 — Figure 6: Nine-season Colorado evaluation sequence.

Explore similar work

Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents
Sep 22, 2026Laizhen Li, Jiarui Li, Juanjuan Zhao +4Agent Harness Scaffolds
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Shenzhen University of Advanced Technology +1
Hierarchical Experimentalist Agents
Jun 28, 2026Abhranil Chandra, Sankaran Vaidyanathan, Utsav Dhanuka +2Agent Loop Long-Horizon Task Planning
University of Massachusetts Amherst
Agent Reinforcement Learning via Pivotal-Aware Self-Feedback Retry
Jul 4, 2026Weiyang Guo, Zesheng Shi, Longhui Zhang +3Retrying Offline Reinforcement Learning
Harbin Institute of Technology, Shenzhen, China

Sep 22, 2026, cs.AIJ/K move · Enter open · S save

Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents

Laizhen Li, Jiarui Li, Juanjuan Zhao, Kejiang Ye+3

Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Shenzhen University of Advanced Technology+1

Agent Harness Scaffolds Gradient Staleness

Open paper AskPDF arXiv

Large language model (LLM) agents often handle streams of related tasks, yet standard harnesses repeatedly ask the model to reconstruct the same control decisions inside each task's context. We study whether task feedback can instead turn recurring control into reusable executable code, while reserving LLM calls for task-specific semantic reasoning. We introduce Growing Harness, a failure-guided training paradigm that learns the agent harness itself from a strategy-free scaffold that exposes fixed model and tool interfaces but encodes no task-solving controller. Function-level execution traces localize each failure to a bounded code surface, an optimizer repairs a window of failures jointly, and a success-first held-out gate rolls back repair sequences that harm prior capability. Accepted edits accumulate in one shared harness, allowing its control structure to emerge from task feedback. Across BrowseComp-Plus and WebArena-Verified with three deployment models from 4B to 120B parameters, Growing Harness achieves the highest mean success in five of six benchmark-model settings and trails the best mean by 0.7 pp. in the sixth. Relative to a Tool-Calling agent, it reduces LLM calls by 76.0-91.8% and deployed-agent inference cost by 74.4-98.6%. On WebArena-Verified, its success remains 44.7-45.3% across model scales, whereas Tool-Calling falls to 6.7% with the 4B model. Ablations show that trace-local edits, joint repair, and gate-based rollback each improve final success. These results show that persistent program growth can move recurring control out of model context and into low-cost code, yielding reusable specialist agents that remain effective with smaller deployment models.

Abstract

Explore similar work

Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents

Hierarchical Experimentalist Agents

Agent Reinforcement Learning via Pivotal-Aware Self-Feedback Retry