State-Centric Decision Process Framework for Language Environments

The newly introduced State-Centric Decision Process (SDP) tackles the absence of runtime structure in language environments such as web browsers and code terminals. In contrast to traditional MDP analysis, these environments produce raw text without defined state spaces, mappings from observations to states, verified transitions, or criteria for termination. SDP fills these gaps by enabling the agent to construct them predicate by predicate during its actions. At each stage, the agent selects a natural-language predicate that represents the intended world state, performs an action to achieve it, and verifies the observation against the predicate. Successful predicates are recognized as certified states, leading to a trajectory that encompasses all four absent components: a task-induced state space, observation-to-state mapping, certified transitions, and a termination criterion. This framework was tested across five benchmarks, as outlined in the paper arXiv:2605.12755.