Open-Weight LLMs Lack Self-Introspection Ability

This research investigates the ability of open-weight large language models (LLMs) to introspect and report on changes to their own internal computational processes. Using a novel framework called Open-Weight Masked Introspection (OWMI), researchers intervened on various internal components, such as residual-stream sites, attention heads, and sparse autoencoder features, across eight different open-weight models. The study's findings indicate that none of the tested models could reliably detect or report whether their internal computation had been altered, performing at chance levels (AUROC ~0.5007). This suggests a fundamental inability for these models to verbally introspect about their own internal states, despite recent claims to the contrary for more complex models. Interestingly, the necessary information for introspection *does* exist within the models' internal activations. A model fine-tuned specifically for this task achieved near-perfect recovery, and linear probes could recover intervention presence with high accuracy (75% to 95.8%) from the same activations. This implies that the failure lies not in the absence of internal signals, but in the path from these internal states to a coherent verbal report. One model even showed the signal in its confidence scores rather than its words. The research concludes that oversight mechanisms relying on an LLM's self-testimony require external validation against internal references.