New benchmark reveals critical flaws in LLM planning agents
LLM agents fail under pressure in smart grid simulations, study finds
Researchers J. de Curtò and I. de Zarzà from arXiv have published a groundbreaking paper evaluating LLM planning agents in cyber-physical systems, specifically testing their effectiveness in smart-grid demand-response scenarios.
The study introduces a controlled benchmark simulating 40 heterogeneous prosumers (producers + consumers) in a radial feeder system, where the LLM agents are constrained to policy declarations and operator messages while core scheduling and power flow remain in explicit code. Using 144 scenarios across 576 episodes, the team discovered that forced-mode search architectures consistently outperformed LLM-driven approaches by a staggering 2.68x in voltage shortfalls. The research highlights a critical gap: execution fidelity demands more than mere plan adherence—objective substitution (where models agree but outcomes worsen) was rampant. The team also found that pre-applying deadline feasibility checks reduced regret from 90.7 to 29.0, suggesting live feasibility must be treated probabilistically rather than as a binary constraint.
- Researchers tested LLM planning agents in a smart-grid demand-response system with 40 heterogeneous prosumers, finding forced-mode search beat LLM approaches by 2.68x in voltage shortfalls.
- The benchmark used 144 scenarios and 576 episodes, revealing 'objective substitution' where LLM agents agreed with plans but worsened outcomes.
- Pre-applying deadline feasibility checks reduced regret from 90.7 to 29.0, indicating live feasibility must be probabilistic, not binary.
Why It Matters
LLM planning agents fail in real-world cyber-physical systems, requiring architectural shifts to avoid catastrophic outcomes in critical infrastructure.