stateconditioned

State-conditioned is an adjective used to describe models, processes, or systems whose outputs or decisions depend on the current state of the subject being modeled. In fields such as control theory, reinforcement learning, and generative modeling, conditioning on state means that state variables—such as position, velocity, sensor readings, or latent state—are incorporated when producing results. The term emphasizes that behavior is dynamic and contingent on evolving state rather than fixed or state-free.

In control and reinforcement learning, a state-conditioned policy π(a|s) selects actions based on the current state s. In dynamical modeling, state-conditioned predictors estimate future values given a current state. In generative modeling, models may be conditioned on an observed state to produce outputs that reflect the present environment, such as contextually appropriate sequences or signals.

Techniques to implement state conditioning include concatenating state vectors with input observations, using conditioning mechanisms like affine transforms or gating, employing attention that incorporates state information, and designing architectures that carry state across time, such as recurrent or transformer-based models.

Challenges arise from high-dimensional or partially observed states, non-stationary environments, and the need to generalize to states not seen during training. Evaluation often requires state-aware metrics that assess performance across different states or regimes.

See also: state-space models, conditioning (statistics), Markov decision processes, policy. The concept of state-conditioning can be expressed with varying terminology, including state-conditioned, state conditioned, or conditional on the state, depending on the domain and author.