Faithfulness is a measurement problem

One paper defines faithfulness structurally as information flow. Another shows the score depends on which classifier you use. Both are about the same thing: the instrument needs calibrating before the readings mean anything.

Faithful reasoning routes answer-relevant information through the mediated path from prompt to chain-of-thought to answer — and measured faithfulness moves depending on the classifier doing the measuring.

The circularity being escaped

Most faithfulness work asks whether an explanation looks like it explains the answer. That is a plausibility judgement, and plausibility is exactly the thing a language model is optimised to produce.

An information-flow criterion asks a structural question instead: does the trace carry information the answer needed? If the answer can be produced without it, the trace is commentary — however sensible it reads.

Why the classifier result matters

Because faithfulness evaluations typically use a classifier to decide whether a trace acknowledges an influence. If classifier choice changes the score, the score is a property of the model-and-classifier pair, not of the model.

Two papers reporting different faithfulness numbers for the same model may not disagree about anything except instrumentation.

Where this has happened before

Sparse autoencoders, roughly a year ago: an instrument adopted fast, then found to have free parameters the readings inherited, then a wave of consistency and validity work. Fields appear to learn this once per instrument, and there is no shortcut.

It is not a debunking

The effect being measured is large enough to survive considerable noise — 87.5% against 28.6% is not a marginal finding. The instruction is to report the apparatus, not to stop measuring.

Which matters because chain-of-thought monitoring is one of the better oversight mechanisms available, and it is worth precisely as much as the trace's fidelity to the computation. That number needs to be knowable.

arXiv — Faithfulness as Information Flow → · arXiv — Measuring Faithfulness Depends on How You Measure → · arXiv — Measuring Chain-of-Thought Monitorability Through Faithfulness and Verbosity →