A Removal Based Approach to Improve LLM Faithfulness at Test-Time

TEXT START: Large language models (LLMs) are increasingly used for consequential decisions, making their explanations an important tool for auditing model behavior.

THE DISSECTION

The paper proposes an inference-time containment procedure: accept the model’s explanation, delete input concepts it did not cite, and re-query the model. Its real product is not recovered reasoning. It is a cleaner relationship between a revised prompt, a revised answer, and a revised explanation.

That is a legitimate input-ablation diagnostic. It is not access to the model’s actual causal computation. The abstract’s claim is also impossible to size from the supplied text: it gives no metric definitions, effect sizes, uncertainty, or failure cases.

THE CORE FALLACY

The central error is treating post-removal consistency as proof of original faithfulness. Deleting visible concepts is not the same as removing their influence from the model’s latent representations, priors, correlations, prompt structure, or learned heuristics. Re-querying creates a new decision process; it does not repair the old one.

The method also lets the explanation define the intervention. That creates a self-confirming loop: the model names what supposedly mattered, the procedure removes everything else, and the resulting answer appears more faithful because the test was redesigned around the model’s own account. It can suppress explicit omissions while leaving hidden influences untouched.

HIDDEN ASSUMPTIONS

• Concepts can be cleanly identified and removed without changing the meaning, distribution, or difficulty of the task.
• Unmentioned concepts are the same thing as unmentioned causal influences.
• Mentioned concepts remain influential in the intended way after the prompt is altered.
• Re-querying preserves the relevant computation instead of invoking a materially different one.
• Faithfulness metrics measure causal truth rather than textual consistency or benchmark compliance.
• The model cannot reconstruct the removed information through context, correlations, or priors.
• Additional inference cost and repeated model access are acceptable in consequential deployments.
• A more faithful explanation is sufficient to make an automated decision safer or more accountable.

SOCIAL FUNCTION

Classification: partial truth, transition management, and ideological anesthetic.

The paper names a real defect and offers a cheap, model-agnostic patch that may improve benchmark-defined faithfulness. Its institutional function is larger: it supplies an audit surface that can make opaque cognitive automation easier to deploy without changing the underlying model or restoring human control. The danger is not that the technique is worthless; it is that its narrow gain gets promoted into evidence of reliability.

THE VERDICT

This is a useful wrapper around an unfaithful system, not a cure for unfaithfulness. It may produce explanations that better match a re-queried answer while leaving the original causal mechanism unknowable. Under the Discontinuity Thesis, it does not resist cognitive automation; it makes that automation easier to legitimize. The machine remains the decision-maker. Only the paperwork gets cleaner.