Technical Summary: SIMMER - Benchmarking Latent Failures in LLM Executable Planning

Problem Statement

Large Language Models (LLMs) are increasingly deployed as planners for autonomous agents in household environments. However, existing evaluation paradigms suffer from a critical blind spot: they fail to detect latent failures.

Current benchmarks typically fall into two categories:

1. Virtual Environments (e.g., TextWorld, ALFWorld, VirtualHome): These rely on formal state tracking but are often overly simplified, lacking the complexity to model implicit state changes like contamination, temperature propagation, or chemical transformations. They primarily detect immediate precondition violations.
2. Semantic Similarity Metrics: These evaluate plans based on surface-level similarity to reference texts. They cannot detect errors where individual actions appear reasonable in isolation but violate implicit state dependencies (e.g., cross-contamination) that only manifest later.

The paper argues that latent failures—errors that satisfy all explicit preconditions but propagate silently through implicit state changes—pose significant risks, particularly when they lead to irreversible consequences (e.g., bacterial contamination that cannot be undone). Unlike immediate failures, these do not halt execution, making them invisible to standard "plan-execute-replan" loops.

Methodology: SIMMER

To address this gap, the authors introduce SIMMER, a benchmark designed to evaluate LLM planning through execution against a symbolic world model. SIMMER consists of three integrated components:

1. Symbolic World Model

Grounded in the kitchen domain, this model is derived from real-world cooking scripts collected from wikiHow and Instructables.

• Scale: It defines 77 actions and 262 unique objects, supporting approximately 46,800 semantically realistic interactions.
• Structure: Following the PDDL paradigm, actions are defined as tuples of <args, preconditions, effects>, and objects as <properties, states, location>.
• Richness: Unlike prior environments (e.g., VirtualHome with ~2,100 interactions), SIMMER models fine-grained state dependencies, such as how a "dirty" state propagates from raw meat to a cutting board and then to vegetables.

2. Failure Taxonomy

The benchmark categorizes planning failures into three distinct types:

• Immediate Failures: Violations of explicit preconditions that block execution (e.g., attempting to chop without holding a knife).
• Latent Failures: Errors that do not violate preconditions but compromise the goal later (e.g., forgetting to add baking powder).
• Irreversible Latent Failures: A critical subset of latent failures where the consequence cannot be undone by subsequent actions (e.g., cross-contamination of food).

3. State Machine Plan Executor

A custom executor simulates plan execution step-by-step to detect failures invisible to conventional metrics.

• Phase 1 (Execution): Checks preconditions and applies state transitions, including propagating latent state changes (e.g., contamination spread).
• Phase 2 (Post-Execution Audit): Scans the final state for unresolved issues (e.g., food left uncooked, appliances left on) that were not flagged during individual steps.

Key Contributions

1. SIMMER Benchmark: The first benchmark to explicitly evaluate latent and irreversible failures in LLM planning using a large-scale, human-curated symbolic world model in the kitchen domain.
2. Systematic Analysis: A comprehensive evaluation of six LLMs (including frontier proprietary models like GPT-5.4 and open-weight models like Llama 3.3) revealing pervasive planning failures.
3. Mitigation Strategy: The proposal and evaluation of Counterfactual Foresight Simulation, a prompting strategy that enforces explicit state reasoning before each action.

Experimental Results

The authors evaluated six LLMs on 100 cooking tasks spanning 12 techniques.

• High Failure Rates: Even the best-performing frontier models produced error-free plans in fewer than 17% of tasks (average across all models was ~7.2%).
• Latent Failures: Up to 56% of plans contained latent failures. Crucially, a significant portion (20–45% of plans) contained irreversible failures.
• Failure Modes:
  • Latent Failures were primarily caused by State Propagation (inability to track how states like contamination transfer across objects) and Implicit Precondition Ignorance (missing safety-critical steps not explicitly stated in instructions).
  • Immediate Failures were dominated by Embodiment Grounding Deficiency (ignoring physical constraints like holding capacity) and State Tracking Errors.
• Mitigation Efficacy:
  • Self-Refine (a post-hoc critique pipeline) reduced total failures by over 40% for frontier models but offered marginal or negative benefits for weaker models.
  • Counterfactual Foresight Simulation (step-by-step state prediction during generation) proved most effective. It reduced latent failures by up to 72% and irreversible failures by up to 75% on frontier models. This suggests that externalizing state tracking via explicit reasoning significantly improves planning reliability.

Significance and Claims

The paper claims that current evaluations substantially overestimate LLM planning capabilities by focusing on surface-level correctness while missing errors that propagate silently through world states.

The significance of SIMMER lies in its ability to:

1. Expose Critical Gaps: Demonstrate that even state-of-the-art models struggle with implicit state dependencies, leading to safety violations that are often irreversible.
2. Provide a Diagnostic Tool: Offer a structured taxonomy and executor to distinguish between immediate, latent, and irreversible failures, enabling more targeted improvements.
3. Guide Future Research: Suggest that explicit state reasoning (via counterfactual foresight) is a promising direction for building more robust LLM planners, particularly for high-stakes domains where silent failures can cause irreversible harm.

The authors conclude that while LLMs show potential, their deployment in safety-critical household environments requires rigorous evaluation against symbolic world models that account for the full complexity of state propagation and irreversible consequences.