SIMMER: Benchmarking Latent Failures in LLM Executable Planning with a World Model

A new benchmark called SIMMER reveals that large language models produce plans riddled with latent failures — errors that do not halt execution but silently compromise safety, often causing irreversible harm in simulated household tasks [1]. The benchmark, detailed in a paper posted to arXiv on June 12, evaluates LLM planning through a human-curated symbolic world model grounded in the kitchen domain [1]. The world model encompasses 77 actions, 262 unique objects, and roughly 46,800 possible interactions derived from real-world cooking scripts [1]. A state machine executor validates each plan against this model, detecting immediate precondition violations, latent hazards, and irreversible failures [1]. Researchers tested six LLMs and found that even frontier models achieved at most 17% error-free plans [1]. Up to 56% of generated plans contained latent failures, and the majority of those led to irreversible consequences [1]. The dominant latent failure mode was the inability to track how states propagate implicitly across objects through contact — for instance, when an agent slices raw meat and then handles other ingredients without recognizing contamination spread [4]. Existing planning benchmarks have focused on whether a plan executes successfully, overlooking failures that do not trigger instant feedback [1]. Latent failures do not immediately halt execution but silently compromise goal achievement [1]. The SIMMER framework addresses this gap by running a post-execution audit that scans the final state for unresolved conditions such as uncooked contamination on food items or appliances left on [3]. The paper also tested a mitigation strategy: explicit state reasoning via counterfactual foresight simulation. This approach reduced latent failures by up to 72% and irreversible cases by up to 75% [1]. The results suggest that prompting models to reason explicitly about state changes can substantially improve plan robustness [1]. arXiv, where the paper was posted, is an open-access repository of electronic preprints that are moderated but not peer-reviewed [9]. As of November 2024, the repository was receiving about 24,000 new articles per month [9].