Probing Embodied LLMs: When Higher Observation Fide... | AI Research

Probing Embodied LLMs: When Higher Observation Fidelity Hurts Problem Solving
This research investigates why Large Language Models (LLMs) often struggle when used as the "brains" for robotic systems. While it is commonly assumed that providing a robot with more accurate, high-fidelity information—such as perfect symbolic data—would improve its performance, this study reveals that the opposite is often true. By observing how LLMs interact with a mechanical puzzle called the "Lockbox," the authors demonstrate that these models can actually perform better when their sensory input is less precise, suggesting that current evaluation methods based solely on success rates may be misleading.

Testing Intelligence Through the Lockbox

To understand how LLMs reason in physical environments, the researchers used a "Lockbox," a puzzle consisting of several joints that must be manipulated in a specific sequence to unlock a reward. Because the dependencies between these joints are hidden, the agent must learn through trial and error. The team tested the LLM’s performance using three different types of input: raw RGB images, RGB-D images (which include depth), and perfect symbolic ground-truth data. They conducted these tests in both a physical robotic setup and a controlled simulation to see how different levels of information fidelity influenced the agents' decision-making processes.

The Paradox of Noise

The study uncovered a counterintuitive result: the LLMs performed best when given raw visual data and worst when provided with perfect, ground-truth symbolic information. To investigate why, the researchers introduced "noise" into the simulation by randomly flipping the perceived outcomes of the robot's actions. They discovered a "sweet spot"—a 40% probability of error—where the agent’s success rate increased significantly, performing 2.85 times better than the noise-free baseline. This suggests that a moderate amount of uncertainty can actually help the model avoid becoming stuck in unproductive patterns.

Breaking Repetitive Loops

The researchers identified a specific failure mode in LLM agents: the tendency to fall into "repetitive action loops," where the model repeatedly executes the same sequence of actions even when they fail to solve the puzzle. Because the environment state remains unchanged after a failed attempt, the model often gets trapped in a cycle of repeating the same unsuccessful behavior. The study found that moderate observational noise disrupts these loops, forcing the model to explore different actions. This indicates that the improved performance seen with lower-fidelity inputs is not necessarily a sign of better reasoning, but rather a byproduct of the noise preventing the model from getting stuck.

Implications for Future AI

These findings highlight a significant challenge in evaluating embodied AI. If success rates can be artificially inflated by perceptual errors or noise, then simply measuring whether a robot completes a task is not enough to determine if it is truly "solving" the problem. The authors argue that researchers must look beyond final outcomes and analyze the behavioral mechanisms—such as the tendency to loop or the ability to interpret feedback—to truly understand the capabilities and limitations of LLM-based robotic agents. Future research will need to determine if these behavioral patterns are universal across different AI models or specific to the systems tested in this study.