Cognitive Dissonance: Why Do Language Model Outputs Disagree with Internal Representations of Truthfulness?
This addresses the issue of interpreting language model behavior for researchers and practitioners, clarifying that disagreements are not necessarily due to lying but involve nuanced mechanisms, though it is incremental in refining existing understanding.
The paper tackled the problem of why language models' outputs disagree with their internal representations of truthfulness, finding that this disagreement arises from confabulation, deception, and heterogeneity, with probes often being more accurate due to better calibration rather than higher correctness rates.
Neural language models (LMs) can be used to evaluate the truth of factual statements in two ways: they can be either queried for statement probabilities, or probed for internal representations of truthfulness. Past work has found that these two procedures sometimes disagree, and that probes tend to be more accurate than LM outputs. This has led some researchers to conclude that LMs "lie" or otherwise encode non-cooperative communicative intents. Is this an accurate description of today's LMs, or can query-probe disagreement arise in other ways? We identify three different classes of disagreement, which we term confabulation, deception, and heterogeneity. In many cases, the superiority of probes is simply attributable to better calibration on uncertain answers rather than a greater fraction of correct, high-confidence answers. In some cases, queries and probes perform better on different subsets of inputs, and accuracy can further be improved by ensembling the two. Code is available at github.com/lingo-mit/lm-truthfulness.