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

Source facts

The preprint studies long-running irrigation decisions where actions change future states and errors compound. Mimir combines a structured physical interface, deterministic simulation, verified revision, bounded action selection, and persistent contextual principles while keeping the physical model and execution constraints fixed. The abstract reports lower aggregate control cost than evaluated references and about 51% less irrigation than historical schedule replay.

Analysis

Mimir is an example of constraining an LLM with explicit numerical mechanisms when the agent acts in a persistent physical environment. It separates semantic proposal and learning from physical truth and actuator authority, a useful pattern for safety-critical applications.

Uncertainty/Doubt

• This is an arXiv preprint. The retrospective evaluator, site and crop coverage, and reported irrigation reductions were not independently reproduced here.