Large language model as reservoir operator: a generative agent bridging human judgment and optimization in water resource systems

Abstract Human judgment drives reservoir water systems globally, but computational models struggle to capture it. Data-driven approaches replicate patterns without revealing decision logic, and optimization requires objectives that are hard to fully specify. Large language models (LLMs) may offer an alternative, as they replicate human decision patterns in economic and cognitive tasks and show emerging multi-step reasoning under uncertainty. We test this by embedding an LLM as an operator of Folsom Lake reservoir in California, tasked with monthly water supply allocation decisions over a 21 year simulation. Benchmarking against historical operations, a neural network regressor, and dynamic programming reveals that the LLM exhibits risk-averse hedging behavior without task-specific training. The LLM’s performance falls between naive release policies and formal optimization, spanning a range comparable to varying risk aversion in stochastic dynamic programming. Models that generate intermediate chain-of-thought traces consistently outperform those that output decisions directly, and sensitivity and ablation analyses partially validate the LLM’s self-reported rationale. These results demonstrate that LLM generative agents can approximate human-like water management decisions from contextual information alone, bridging quantitative optimization and qualitative judgment for decision support, policy exploration, and human–water system research.