AtomicGuard and Active Inference

Having fun reading up on Friston’s Active Inference. Has me thinking that AtomicGuard is an inversion of Active Inference.

Agent alignment to the world via Active Inference

IIUC Friston posits that we model:

• the world P(s)
• observation (perception/sensing) P(o)
• an expectation of how the new observations update our world model Q(s)

And then we then calculate the surprise associated with our new world model given our expectation. Which is the KL Divergence between Q(s) and P(s'|o). It will continue until it converges on a world model that aligns to its observations.

Aligning the world to agent specification

The AtomicGuard framework (Dual State Action Pairs) does the inverse.

• it knows the desired state (i.e. the specification)
• it observes an approximation of the specification (i.e. the response from the LLM)
• it has a requirement of how the observation must be (i.e. the guard function)

If the requirement is satisfied then the agent progresses, otherwise it will request a new approximation from the LLM until the environment converges on the desired state*.

*This is dependent on the learning loop that updates the generative model.