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

2026-06-12

AI Summary

This study addresses latent failures in large language model (LLM)–generated plans for household environments—errors that do not immediately halt execution but silently undermine task objectives or cause irreversible damage. The work presents the first systematic definition and evaluation of such latent planning failures and introduces an explicit state-reasoning approach augmented with counterfactual forward simulation. Using a symbolic kitchen world model encompassing 77 actions, 262 objects, and approximately 46,800 semantically valid interactions, plans are validated via a state-machine executor. Experiments reveal that only 17% of plans produced by state-of-the-art LLMs are error-free, with 56% containing latent failures. Incorporating the proposed method reduces latent failures by 72% and irreversible failures by 75%.

Abstract

Large language models (LLMs) are increasingly deployed as planners for autonomous agents in household environments. While existing benchmarks evaluate whether LLM-generated plans execute successfully, they overlook a critical type of failure: latent failures. Unlike immediate failures that trigger instant feedback at execution time and enable timely correction, latent failures do not immediately halt plan execution but silently compromise goal achievement. In severe cases, they cause irreversible harm. To address this gap, we introduce SIMMER, a benchmark for evaluating latent failures in LLM planning through a human-curated symbolic world model grounded in the kitchen domain. SIMMER defines a world model comprising 77 actions, 262 unique objects, and approximately 46,800 possible interactions that are semantically realistic, derived from real-world cooking scripts. It then leverages a state machine executor that validates plans against the world model and detects immediate precondition violations, latent hazards, and irreversible failures. Experiments across six LLMs show that even frontier models achieve at most 17% error-free plans. Moreover, up to 56% of plans contain latent failures, the majority of which lead to irreversible consequences. We further demonstrate that explicit state reasoning via counterfactual foresight simulation can reduce latent failures by up to 72% and irreversible cases by up to 75%, suggesting a promising direction for more robust LLM planners.

Problem

• latent failures
• LLM planning
• executable planning
• autonomous agents
• world model

Innovation

• latent failures
• world model
• executable planning
• counterfactual foresight
• LLM benchmarking