A new prototype lets teams of AI models flag their own confusion, then decide together whether to answer fast or slow down and argue it out.
Researchers built a proof-of-concept metacognition framework for ensembles of large language models, letting the system track its own uncertainty rather than just generate an answer. The system computes what the authors call a Metacognitive State Vector, scoring each query across five dimensions borrowed from cognitive psychology: emotional response, correctness evaluation, experiential match, conflicting information, and problem importance. Based on those scores, the framework automatically routes simple queries through fast System 1 processing and harder ones through slower, deliberative System 2 processing that recruits multiple model nodes. In System 2 mode, graph-theoretic algorithms assign each node a role - domain expert, critic, evaluator, synthesizer, or generalist - based on its own metacognitive state, and the demo visualizes the whole process with real-time radar charts.
LLMs are bad at knowing what they don't know, which is a big reason they hallucinate confidently instead of flagging doubt. This framework is a rare attempt to bolt an actual self-monitoring layer onto an ensemble, rather than just prompting a model to 'rate your confidence' and hoping the number means something. If self-awareness signals can reliably trigger slower, more scrutinized reasoning, that's a plausible path toward LLM systems that know when to ask for help.
For now it's a demo built on illustrative examples, not a benchmark run against hallucination-heavy tasks - the harder test is whether the same routing holds up once the queries stop being hand-picked.