Probing an Embodied LLM: When Higher Observation Fidelity Hurts Problem Solving

Large Language Models are increasingly proposed as cognitive components for robotic systems, yet their opaque decision processes make it difficult to explain success or failure in closed-loop embodied tasks. The gist: Agents perform best under raw RGB input and worst under perfect ground-truth observations, suggesting that measured performance may reflect the interaction between perceptual errors and reasoning failures rather than robust problem solving.

The Problem and Methodology

Analyzing embodied LLM agents requires moving beyond outcome-based evaluation toward methods that reveal the behavioral mechanisms underlying success and failure. Because LLMs are largely opaque, their reasoning processes remain inaccessible, and aggregate success rates provide only a limited view of performance, as they do not reveal whether an agent solved a task through coherent inference, accidental exploration, perceptual error, or an interaction between these effects. To address this opacity, the researchers adopt a behavioral probing approach aligned with empirical AI methodology [4], inspired by system identification: rather than estimating an explicit dynamical model, they perturb an opaque agent and measure how its behavior changes. They use observation fidelity to denote how faithfully an agent’s observations represent the true state of its environment, using this as an intervention variable because it affects the information on which the model bases its decisions while leaving the underlying task structure unchanged.

The Lockbox Task and Observation Modalities

The study uses a sequential mechanical puzzle called the Lockbox, which is widely used to study problem-solving behavior across various domains. The task captures key challenges of real-world problems: actions have state-dependent outcomes, and optimal actions depend on the integration of past interactions. The researchers present this task through three different observation channels:

1. Raw RGB inputs.

2. RGB-D inputs, which augment the RGB modality with an aligned depth channel, providing additional geometric information about the scene.

3. Ground-truth symbolic state descriptions, which bypass the vision pipeline entirely and provide explicit labels for each Lockbox joint and its corresponding state.

Results on Observation Fidelity

The results reveal a counterintuitive pattern: increasing observation fidelity does not improve performance. Agents perform best with raw RGB input and worst when given symbolic ground-truth state. In simulation, this effect was probed by randomly flipping perceived action outcomes, finding that moderate noise improves performance, peaking at a 40% flip probability with a 2.85-fold success rate increase over the noise-free baseline. This gain is linked to a reduction in repetitive action loops, which are identified as a failure mode where an agent repeats near-identical action subsequences despite making little or no progress.

Perturbing Observations Reveals Non-Monotonic Response

To test whether lower-fidelity inputs improve performance by introducing stochasticity into the perceived Lockbox state, the researchers conducted a controlled experiment in simulation. They introduced noise directly into the state observations through random state-flips applied at controlled probabilities, mimicking misinterpretation of a visual observation. The fit revealed a "non-monotonic relationship between observational noise and task performance: starting from an average success rate of 23.3% under ground-truth observations, performance rises to a peak at a 40% state-flip probability before declining at higher noise levels. This supports the hypothesis that moderate observation noise can improve measured task performance without necessarily reflecting improved reasoning."

Repetitive Action Loops as a Candidate Failure Mode

A recurring behavioral pattern observed across trials is the tendency of the LLM to produce repetitive action subsequences within a single trial. These loops substantially degrade efficiency and frequently cause agents to exhaust their interaction budget before solving the Lockbox. The researchers found that perceptual noise is associated with fewer action loops, which are associated with reduced performance. Specifically, Figure 5(b) shows that the probability of engaging in repetitive action sequences decreases with increasing perceptual noise, reaching a minimum at a 40% state-flip probability, suggesting that the observed increase in success rates under moderate noise is consistent with the reduction in repetitive action sequences.

Conclusion

The findings suggest that success rates alone are insufficient for evaluating LLMs, as measured performance may reflect an accidental interplay between perceptual errors, reasoning failures, and environmental dynamics rather than robust problem solving. The conclusion is that accurate feedback may sustain repetitive action loops, whereas erroneous observations can disrupt them. Therefore, what appears as better performance under lower fidelity inputs should not be interpreted as evidence for stronger reasoning but may instead reflect a side effect of compounding perceptual errors that happen to be beneficial for this task. The work concludes that LLMs are still subject to weaknesses at both the perceptual and reasoning level, and rigorous evaluation requires controlled behavioral interventions.

--- Page 12 ---

References

1. Agrawal, P., Goyal, P.: Addressing llm diversity by infusing random concepts (2026), https://arxiv.org/abs/2601.18053

2. Auersperg, A.M.I., Kacelnik, A., von Bayern, A.M.P.

Here are the specific improvements to AI systems based on the findings in this paper, categorized by their impact:

)Improved System Capabilities: Probing and Robustness of Embodied LLMs

The core improvement suggested by this research is a shift from evaluating LLM performance solely on final success rates to developing behavioral probing methodologies that reveal the underlying cognitive mechanisms. This leads to systems that are not just better at tasks, but better at understanding why they succeed or fail.

Here are the specific improvements and what these improved AI systems can do:

1. A system capable of diagnosing its own failure modes by analyzing observational fidelity effects.

2. A system capable of dynamically adjusting its internal perception strategy based on environmental uncertainty to optimize performance in real-time.

3. A system that minimizes inefficient, repetitive action loops during complex sequential problem-solving tasks (like the Lockbox).

)Specific Implementation Details and Expected Outcomes:

1. The improved AI system can be equipped with a Fidelity Sensitivity Module that monitors the input modality (e.g., RGB vs. RGB-D vs. Ground Truth State) and dynamically adjusts its internal inference weighting or attention mechanisms based on real-time environmental uncertainty metrics derived from the observation channel quality.

2. The system can implement a Stochastic Perturbation Engine that, when encountering low-fidelity input or ambiguous state information, deliberately introduces controlled noise (mimicking state flips) into its perceived action outcomes to test for robustness and improve exploratory behavior, rather than relying solely on perfect (but potentially misleading) structured inputs.

3. The system can incorporate a Loop Detection and Pruning Subsystem that uses a combinatorial optimization solver (like the ILP described in the paper) to identify and eliminate redundant or cyclical action subsequences before they exhaust interaction budgets, thereby increasing overall solution efficiency by preventing repetitive failure modes.

)Overall System Capability Enhancement:

The improved AI system will transition from a purely reactive problem-solver to a more sophisticated, self-aware agent capable of:

1. Identifying the trade-off between perceptual accuracy and reasoning success.

2. Adapting its perception strategy to navigate uncertainty (i.e., knowing when good enough noisy information is superior to perfect, misleading data).

3. Executing solutions with higher efficiency by avoiding redundant planning cycles that plague current LLM agents in closed-loop tasks.

In essence, the improved system will be more resilient and mechanistically informed, reducing reliance on perfect inputs and mitigating common LLM failure modes (like getting stuck in loops) through controlled perceptual noise injection.