Eliciting Latent Knowledge from Quirky Language Models

TL;DR

This paper introduces datasets and methods to extract truthful knowledge from language models that may produce untrustworthy answers, demonstrating effective probing and anomaly detection techniques.

Contribution

The authors present new datasets, a suite of quirky language models, and evaluation methods for reliably eliciting latent knowledge despite model deception.

Findings

Linear probes can detect knowledge independently of model output.

Contrast pair logistic regression recovers 89% of AUROC gap.

Anomaly detection flags untruthful behavior with 0.95 AUROC.

Abstract

Eliciting Latent Knowledge (ELK) aims to find patterns in a capable neural network's activations that robustly track the true state of the world, especially in hard-to-verify cases where the model's output is untrusted. To further ELK research, we introduce 12 datasets and a corresponding suite of "quirky" language models (LMs) that are finetuned to make systematic errors when answering questions if and only if the keyword "Bob" is present in the prompt. We find that, especially in middle layers, linear probes usually report an LM's knowledge independently of what the LM outputs, enabling us to elicit the correct answer despite the model's untruthful output. The best probing method (logistic regression on contrast pairs) recovers 89% of the gap in AUROC between truthful and untruthful contexts, and 75% for questions harder than those used to train the probe. We also find that a mechanistic anomaly detection approach can flag untruthful behavior with 0.95 AUROC. Our results show promise for eliciting reliable knowledge from capable but untrusted models, and facilitates future research empirically investigating ELK methods.