Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Basu, Sanjay, Patel, Sadiq Y., Sheth, Parth, Muralidharan, Bhairavi, Elamaran, Namrata, Kinra, Aakriti, Morgan, John, Batniji, Rajaie
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866915874052505600
author Basu, Sanjay
Patel, Sadiq Y.
Sheth, Parth
Muralidharan, Bhairavi
Elamaran, Namrata
Kinra, Aakriti
Morgan, John
Batniji, Rajaie
author_facet Basu, Sanjay
Patel, Sadiq Y.
Sheth, Parth
Muralidharan, Bhairavi
Elamaran, Namrata
Kinra, Aakriti
Morgan, John
Batniji, Rajaie
contents Language models encode task-relevant knowledge in internal representations that far exceeds their output performance, but whether mechanistic interpretability methods can bridge this knowledge-action gap has not been systematically tested. We compared four mechanistic interpretability methods -- concept bottleneck steering (Steerling-8B), sparse autoencoder feature steering, logit lens with activation patching, and linear probing with truthfulness separator vector steering (Qwen 2.5 7B Instruct) -- for correcting false-negative triage errors using 400 physician-adjudicated clinical vignettes (144 hazards, 256 benign). Linear probes discriminated hazardous from benign cases with 98.2% AUROC, yet the model's output sensitivity was only 45.1%, a 53-percentage-point knowledge-action gap. Concept bottleneck steering corrected 20% of missed hazards but disrupted 53% of correct detections, indistinguishable from random perturbation (p=0.84). SAE feature steering produced zero effect despite 3,695 significant features. TSV steering at high strength corrected 24% of missed hazards while disrupting 6% of correct detections, but left 76% of errors uncorrected. Current mechanistic interpretability methods cannot reliably translate internal knowledge into corrected outputs, with implications for AI safety frameworks that assume interpretability enables effective error correction.
format Preprint
id arxiv_https___arxiv_org_abs_2603_18353
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations
Basu, Sanjay
Patel, Sadiq Y.
Sheth, Parth
Muralidharan, Bhairavi
Elamaran, Namrata
Kinra, Aakriti
Morgan, John
Batniji, Rajaie
Artificial Intelligence
I.2.7; J.3
Language models encode task-relevant knowledge in internal representations that far exceeds their output performance, but whether mechanistic interpretability methods can bridge this knowledge-action gap has not been systematically tested. We compared four mechanistic interpretability methods -- concept bottleneck steering (Steerling-8B), sparse autoencoder feature steering, logit lens with activation patching, and linear probing with truthfulness separator vector steering (Qwen 2.5 7B Instruct) -- for correcting false-negative triage errors using 400 physician-adjudicated clinical vignettes (144 hazards, 256 benign). Linear probes discriminated hazardous from benign cases with 98.2% AUROC, yet the model's output sensitivity was only 45.1%, a 53-percentage-point knowledge-action gap. Concept bottleneck steering corrected 20% of missed hazards but disrupted 53% of correct detections, indistinguishable from random perturbation (p=0.84). SAE feature steering produced zero effect despite 3,695 significant features. TSV steering at high strength corrected 24% of missed hazards while disrupting 6% of correct detections, but left 76% of errors uncorrected. Current mechanistic interpretability methods cannot reliably translate internal knowledge into corrected outputs, with implications for AI safety frameworks that assume interpretability enables effective error correction.
title Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations
topic Artificial Intelligence
I.2.7; J.3
url https://arxiv.org/abs/2603.18353