LLM self-awareness research: models that know when they're wrong

A Show HN project called Cactus Hybrid published work on teaching Gemma 4 to detect its own uncertainty using mechanistic interpretability. The approach: find hidden state signals that correlate with model 'self-awareness' about correctness, then use those signals to route to a larger model when the small one is likely wrong. Commenters drew comparisons to Goodfire's reinforcement learning from representation work.

This is a practical engineering approach to a real problem: small models are cheap but wrong too often, large models are accurate but expensive. A reliable confidence signal lets you build tiered inference pipelines that are both cheap and accurate. The discussion immediately went to coding task benchmarks and agent frameworks where a primary large model could delegate subtasks to smaller ones with fallback logic.

The thread also raised a fair methodological question: is the hidden state signal actually measuring self-awareness, or is it measuring something correlated but distinct? The honest answer from the builders was 'we don't fully know yet.'