Can LLMs Introspect? A Reality Check
Shashwat Singh · Tal Linzen · Shauli Ravfogel
Claims of LLM introspection are confounded: labels in self-report tasks are predictable from the input alone, and detection of activation steering is indistinguishable from generic anomaly detection.
摘要
Can large language models detect and report their own internal states? A number of recent studies have argued that they can. Drawing on lessons from human metacognition research, we argue that this conclusion may be premature. We identify two conditions that a paradigm needs to meet in order to establish introspection. First, the test needs to require \emph{privileged access}: it should not be solvable using cues available in the input. Second, it needs to require \emph{second-order computation}: second-order, meta-representations of first-order, task-related representations. This condition cannot be satisfied by task performance alone: it requires designs under which second-order and first-order accounts make divergent predictions. We re-examine two paradigms that have been used to argue for model introspection in light of these conditions. In the first, models must predict labels derived from their own hidden states; we find that classifiers that can only access the input match the models' in-context predictions, indicating that the original results do not demonstrate privileged access to internal representations. In the second paradigm, models must detect whether their internal states have been tampered with; we find they cannot reliably distinguish such interventions from manipulations of the input, suggesting that their success reflects generic anomaly detection rather than sensitivity to internal interventions in particular. We conclude that current evidence is insufficient to establish metacognitive monitoring in LLMs.