A Coupled Explanation Is Not a Readout: What Self-Explanation Training Shows When a Model's Explanations Follow Its Own Drift, Why an Explanation Computed in a Separate Forward Pass Can Only Replay What It Explains, and the Interventions That Would Separate Replay From Access

Abstract

Life of Research

Training language models to explain their own behaviour is increasingly offered as a scalable route to oversight, and a 2026 result made the route look stronger: models fine-tuned on fixed counterfactual explanations of an earlier checkpoint, or of a different model, come to explain their own current behaviour better than the behaviour their labels describe, and their explanations follow behaviour that later post-training shifts. The authors call this introspective coupling and read it as evidence for a potentially stronger form of privileged self-access. In this analysis the author reviews and synthesises published results to show that the coupling result is real and that it licenses a narrower conclusion than access. In the designs that produce it, the explanation is generated in a forward pass separate from the behaviour it describes, with the behaviour's input restated in the query, so the only channel from behaviour to explanation is the shared weights. An explanation computed that way can track the model's drift by re-running the relevant computation on the restated input, which this paper calls replay. Replay predicts the signatures reported for such training: beating a peer model trained on the same labels, beating the source of the labels, correlated effects of patching the shared input, and failure on long or complex properties. It also predicts that the advantage holds only against observers who cannot run the model; for the questions these studies ask, an observer who can run the model computes the ground truth exactly, because that is how the labels are made. Only designs that set an internal state to a value the tokens would not produce can separate replay from a readout of internal state, and those designs report detection that is real, unreliable and largely content-agnostic. The paper consolidates 33 rows of published values in a table and a figure, sorts the self-advantage literature by route and by the observer it is measured against, states a reading procedure for self-explanation claims, and names six experiments that would settle the open part.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 144 papers by Pranay Mahendrakar, or his ORCID record.