Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations
Citations
0
Open access
No
Source
arxiv
OpenAlex
Not enriched
arXiv
2603.18353
Abstract
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.
Collections
Add to collection
Paper intelligence
Analysis has not been completed yet.
No graph connections yet.
Sync citations or add papers to shared collections to build this network.
Knowledge graph
Citation network
Explore references, papers that cite this work and related papers in your Codex library.
References
0No references have been linked yet.
Cited by
0No saved paper is currently linked as citing this work.
Related papers
0Add papers to shared collections or enrich their topics to find related work.
Research workspace
Attach the paper PDF, extract its text, classify its contents and create semantic embeddings.
Paper resources
No PDF assets have been attached to this paper yet.