Stochastic sampling fails to replace structured control in language agents

A recent investigation available on arXiv (2605.09692) examines whether stochastic sampling can replace structured mechanisms such as reason, memory, self-state, and inhibition in the action selection of language agents. Analyzing a baseline lesion matrix comprising seven datasets with 74,352 calls, the high-stochasticity comparator was found to be more erratic than the structured-control counterpart across all datasets. In contrast, lesions targeting reason and inhibition diminished expected structured-control profiles in each dataset. Furthermore, in a matched-interface control involving 26,946 generations, the structured agent exhibited stronger action-field coupling compared to all stochastic, post-hoc, scrambled, and verbosity controls throughout the datasets. The main behavioral assessment eliminated free-form trace wording, demonstrating that unpredictability does not imply control.