Microsaccade-Inspired Probing: Positional Encoding Perturbations Reveal LLM Misbehaviours

We draw inspiration from microsaccades, tiny involuntary eye movements that reveal hidden dynamics of human perception, to propose an analogous probing method for large language models (LLMs). Just as microsaccades expose subtle but informative shifts in vision, we show that lightweight position encoding perturbations elicit latent signals that indicate model misbehaviour. Our method requires no fine-tuning or task-specific supervision, yet detects failures across diverse settings including factuality, safety, toxicity, and backdoor attacks. Experiments on multiple state-of-the-art LLMs demonstrate that these perturbation-based probes surface misbehaviours while remaining computationally efficient. These findings suggest that pretrained LLMs already encode the internal evidence needed to flag their own failures, and that microsaccade-inspired interventions provide a pathway for detecting and mitigating undesirable behaviours.

Key Contributions

Microsaccade-inspired probing method using lightweight position encoding perturbations to elicit latent misbehaviour signals from LLMs
Unified detection framework requiring no fine-tuning or task-specific supervision that generalises across factuality, safety, toxicity, and backdoor failure modes
Empirical demonstration that pretrained LLMs internally encode evidence sufficient to flag their own failures, surfaced via positional perturbations

🛡️ Threat Analysis

Model Poisoning

Explicitly detects backdoor attacks as one of the core evaluated threat types; the probing method surfaces backdoor-induced misbehaviours using position encoding perturbations.