SIRIS: Can LLMs Tell When Simulation Errors Matter in AI-Driven Systems Research

SIRIS. Can LLMs Tell When Simulation Errors Matter in AI-Driven Systems Research. By Jinwoo Hwang and five others, from KAIST and Microsoft Research Asia. Published on the sixth of November.

LLM agents are increasingly the ones driving systems research, and they lean on simulators to evaluate candidate designs at scale. Simulators are imperfect by design. A modeling limitation can distort the predictions and still leave the recommended design correct. Or it can change which design ought to be chosen. Those are very different situations, and nothing in a simulator's output says which one an agent is in.

Siris tests whether LLMs can tell the two apart. The question is whether a model, drawing on its own understanding of systems, can judge when a limitation actually invalidates the simulator's recommendation, as opposed to merely bending its numbers.

The construction works by perturbing a simulator. From that, Siris builds pairs of workloads that share the same candidates, the same limitation, and the same simulator recommendation. The perturbation changes the predictions on both workloads. A reference simulator, though, supports the recommendation on only one of them. So the two cases look alike on the surface, and only an understanding of the system separates the one where the recommendation holds from the one where it fails.

Thirteen frontier LLMs were evaluated on design problems in LLM serving. Spotting a limitation, it emerges, does not ensure a correct choice. Even the strongest models sometimes abandon recommendations that were valid. Adding a review turn for self-correction helps those models restore valid choices, but in several others it introduces new errors.

Knowing that a simulator has a flaw and knowing whether the flaw matters are separate skills, and Siris is built to pull them apart. The authors draw the lesson that systems understanding should be evaluated through the design decisions a model makes under an imperfect simulator, not only through whether it can name what is wrong.