StressWeb Benchmark Tests LLM Web Agent Robustness Against Realistic Interaction Variability

A new diagnostic benchmark called StressWeb has been introduced to evaluate the robustness of large language model-based web agents. Existing evaluations often occur in stable, well-behaved interaction conditions, potentially overestimating agent performance. StressWeb addresses this by creating realistic, controllable web environments that serve as reference baselines with clean, stable workflows. The framework then introduces structured, controlled perturbations that emulate real-world interaction variability. These perturbations include shifting layouts, altered interaction semantics, and execution disruptions. By systematically comparing agent behavior between these clean baseline settings and the perturbed environments, the benchmark enables a diagnosis of robustness under various "what-if" scenarios. The research, documented in arXiv:2604.16385v1, highlights that high task success in idealized settings may not reflect performance in realistic web interactions. This work aims to provide a more accurate assessment of how these agents would perform when faced with the unpredictable nature of actual web use.