Do they know
what they know?

I’m interested in how intermediate computations relate to what a model knows, where it is uncertain, and the explanation it eventually gives.

THE QUESTIONWhich internal signals track knowledge or uncertainty, and when does an explanation fail to reflect the computation behind it?

A useful self-model could include knowledge, uncertainty, capabilities, available tools and a record of what the system has already done.

THE QUESTIONDoes this information remain accurate as circumstances change, and does it need to be represented as one integrated model?

The test is behavioural: whether information about a system’s limitations changes when it answers, checks its work or asks for help.

THE QUESTIONWhat changes when decisions use self-relevant information, once errors, extra tool calls and deferred work are taken into account?

Agents act on the world. Their actions change what they know and what they can do next. This is where my project work meets my doctoral interests.

THE QUESTIONHow should an agent update its account of its capabilities and past actions when observations no longer match its assumptions?