Agency Under Partial Observability: Prediction, Compression, and Empowerment
A recent study published on arXiv (2605.06346) introduces a theoretical model addressing agency in deterministic environments, where perceived randomness arises from uncertainties related to initial conditions, fixed law components, and external noise. The researchers conceptualize sensing and actuation as intermediary interfaces divided between parameters controlled by the agent and the environmental channel state, resulting in a deterministic POMDP. They demonstrate a distinction among prediction, compression, and empowerment: achieving perfect prediction necessitates either discovering the hidden quotient or exercising control, while high empowerment alone does not suffice. With adjustable interfaces and adequate memory, action-conditioned observation-compression diminishes posterior uncertainty regarding the latent quotient.
Key facts
- arXiv paper 2605.06346
- Announce type: new
- Studies agency under partial observability in deterministic worlds
- Apparent randomness arises from uncertainty over initial conditions, fixed law bits, and unrolled exogenous noise
- Models sensing and actuation as bridge interfaces
- Bridge interfaces split between agent-controlled parameters and environment-controlled channel state
- Induces a deterministic POMDP through prior over latent microstates and many-to-one observation coarsening
- Proves separation between prediction, compression, and empowerment
- Perfect prediction requires identifying hidden quotient or overwrite control
- High empowerment alone insufficient for perfect prediction
- Under refinable interfaces and sufficient memory, action-conditioned observation-compression reduces posterior uncertainty about latent quotient
Entities
Institutions
- arXiv