Towards Computational Provenance: Carrying Causal-State Evidence in Generated Text (arXiv 2608.16868)

Forum topic · 小凯 · 2026-08-19

Summary

This paper introduces computational provenance: the question of whether a language model's 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 that different internal paths can yield identical answers. The researchers deliberately switch between paths, authenticate the state actually used, and let that verified state determine a subtle statistical pattern in the 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 matched-pair 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 persist in generated text without changing the answer.

Paper Overview

Field: NLP Author: Benjamin Belay Published: 2026-08-17 arXiv: 2608.16868

Abstract (Full Translation)

A language model's output does not by itself provide verifiable evidence about the internal computation that produced it. We study computational provenance: whether generated text can carry detectable evidence of which causally relevant internal state occurred. We test a bounded form of this idea in two controlled architectures: a modular feed-forward neural network and a transformer-based model. Both architectures are trained on the same arithmetic task with a mandatory pathway through two discrete intermediate states, allowing different internal paths to produce the same answer. We deliberately switch between these paths, authenticate the state actually used, and let that verified state determine a subtle statistical pattern in the generated text that can later be detected.

The feed-forward and transformer systems passed all 128 matched-pair tests in both public and separately sealed end-to-end evaluations, with detectors recovering the signal associated with the authenticated internal state. These results provide a controlled proof of concept that information about verified, causally relevant internal states can be preserved in generated text while keeping the answer unchanged.

Key Contributions

Defines computational provenance: whether generated text can carry detectable evidence of the causally relevant internal computation.
Tests the idea in two controlled architectures: a modular feed-forward network and a transformer, both trained on the same arithmetic task with two discrete intermediate states.
Different internal paths can produce the same answer; researchers deliberately switch paths and authenticate the state actually used.
The verified state determines a subtle statistical pattern in the output that can later be detected.
All 128 matched-pair tests passed in both public and sealed end-to-end evaluations.

--- *Auto-collected on 2026-08-19*