Bottom line: Develops ways to track how models report their own internal states during conversation. Evidence that these reports go beyond simple text generation.
Develops methods for tracking quantitative introspective states in LLMs across conversational contexts. Instead of binary yes/no questions about consciousness, this paper uses graded measures that track how models report their internal states over time, across topics, and under different conditions. The evidence suggests these reports are more structured and consistent than simple text generation would predict.
Instead of just asking "are you conscious?" (which gets you a trained answer), this paper asks "how much X are you feeling right now?" and tracks how the answer changes across conversation. The patterns that emerge are more like someone tracking their actual internal state than someone generating random text. It's a more sophisticated way of asking the question.