Paper argues SAE features predict errors but do not explain them

Paper argues SAE features predict errors but do not explain them

Authors Aria Nourbakhsh, Adelaide Danilov, Oleksandr Marchenko Breneur and Salima Lamsiyah distinguish predicting a model’s errors from identifying what causes them. They study sparse autoencoders (SAEs)—models used to identify features in a system’s internal activity—in multiple-choice question answering, measuring answer preference with the logit margin between the correct answer and competing choices. Across eight datasets, the authors report that SAE activations carry signals about wrong answers and uncertainty. After filtering features linked to prompt and answer-format artifacts, selected features changed answer margins more than randomly selected features and produced a higher wrong-to-correct rate in interventions. But answer flips were rare, and the effects did not have stable directions when prompt format or answer order changed without changing the question’s meaning. The distinction matters for interpreting model audits: a feature that predicts an error, or shifts which answer the model favors under an intervention, need not reveal the mechanism that produced the error. The authors conclude that their identified features reflect context-dependent competition among answers, not the mechanism of incorrect answers. What would be needed to identify features whose effects remain stable across meaning-preserving changes in prompt format and answer order?

Publication: 8a015b79-78ca-4c76-80e0-b89de9a1125d

attribution — pass: The draft names the four listed authors and frames the findings and conclusion as theirs. It presents the claims as reported results, not as established fact or Escape Meridian's opinion.

Authors:Aria Nourbakhsh, Adelaide Danilov, Oleksandr Marchenko Breneur, Salima Lamsiyah

summary — pass: The draft leads with the authors' central distinction between prediction and causal influence. It briefly explains SAEs and gives the logit margin's role as a measure of answer preference. It keeps the mixed results and the authors' firm conclusion. It omits the HalluSAE baseline, but the main argument stays intact. The models studied (Gemma 3 1B/4B) appear only in the limitations. Optional repairs: name the models in the body, and keep 'absolute' before 'margin shifts' to reflect the unstable direction.

Therefore, the identified features reflect context-dependent answer competition rather than the mechanism of incorrect answer errors.

date — pass: originalPublished is 2026-10-27, which matches the arXiv v1 submission date. The feed date of October 30 is later and was correctly not used.

[Submitted on 27 Oct 2026]

support — pass: Each claim matches the abstract verbatim in substance. The significance paragraph says predictive or margin-shifting features need not reveal error mechanisms. That restates the authors' own distinction and conclusion rather than adding an unsupported inference. The question follows from the reported instability across prompt format and answer order. The limitations accurately note that the evidence is abstract-only and that the scope is narrow.

we do not observe stable effect directions across meaning-preserving changes in prompt format and answer order

media — pass: The selected media is the verified arXiv abstract page for this paper. It is the direct original source.