Towards Computational Provenance: Detectable Causal-State Evidence in Generated Text

Summary

This paper (arXiv:2608.16868) studies computational provenance in language models: whether generated text can carry detectable evidence about which causally relevant internal state produced it. The authors test a bounded form of the idea in two controlled architectures—a modular feed-forward neural network and a transformer-based model—both trained on the same arithmetic task with a mandatory pathway through two discrete intermediate states, so different internal paths can yield identical answers. They deliberately switch between paths, authenticate the state actually used, and let that verified state determine a subtle statistical pattern in the generated output that can later be detected. In both public and separately sealed end-to-end evaluations, the feed-forward and transformer systems passed all 128 paired matching tests, with detectors recovering the signal associated with the authenticated internal state. The results provide a controlled proof of concept that information about verified, causally relevant internal states can be preserved in generated text without changing the answer.