Towards Computational Provenance: Carrying Causal-State Evidence in Generated Text
The paper tests whether generated text can preserve detectable evidence of a verified internal state, even when the visible answer is unchanged.
Benjamin Belay frames this as “computational provenance” and tests it in a modular feed-forward network and a transformer on a controlled arithmetic task. The models are forced through one of two discrete intermediate states, then the verified state is encoded as a subtle statistical signal in the output text. Detectors recovered that signal across all 128 matched pairs in both public and sealed evaluations, with replication across independently trained models. A separate answer-only transformer experiment did not yield a naturally recoverable intermediate state by linear probing. ArXiv · AI/CL/LG's note
Benjamin Belay frames this as “computational provenance” and tests it in a modular feed-forward network and a transformer on a controlled arithmetic task. The models are forced through one of two discrete intermediate states, then the verified state is encoded as a subtle statistical signal in the output text. Detectors recovered that signal across all 128 matched pairs in both public and sealed evaluations, with replication across independently trained models. A separate answer-only transformer experiment did not yield a naturally recoverable intermediate state by linear probing. ArXiv · AI/CL/LG's note
score 5