[Paper Review] Truth is Universal: Robust Detection of Lies in LLMs

Lennart Bürger, Fred A. Hamprecht+1 more|arXiv (Cornell University)|Jul 3, 2024

Law, Economics, and Judicial SystemsEconomics, Econometrics and Finance3 citations

TL;DR

This paper proposes a robust, general-purpose lie detection method for large language models (LLMs) by identifying a universal two-dimensional truth subspace in model activations. The method, Truth-Tuned Projection Detector (TTPD), leverages a generalized truth direction ($\mathbf{t}_G$) that separates true and false statements across diverse topics and statement types, achieving 94% accuracy on simple statements and 95% on real-world role-playing lies.

ABSTRACT

Large Language Models (LLMs) have revolutionised natural language processing, exhibiting impressive human-like capabilities. In particular, LLMs are capable of "lying", knowingly outputting false statements. Hence, it is of interest and importance to develop methods to detect when LLMs lie. Indeed, several authors trained classifiers to detect LLM lies based on their internal model activations. However, other researchers showed that these classifiers may fail to generalise, for example to negated statements. In this work, we aim to develop a robust method to detect when an LLM is lying. To this end, we make the following key contributions: (i) We demonstrate the existence of a two-dimensional subspace, along which the activation vectors of true and false statements can be separated. Notably, this finding is universal and holds for various LLMs, including Gemma-7B, LLaMA2-13B, Mistral-7B and LLaMA3-8B. Our analysis explains the generalisation failures observed in previous studies and sets the stage for more robust lie detection; (ii) Building upon (i), we construct an accurate LLM lie detector. Empirically, our proposed classifier achieves state-of-the-art performance, attaining 94% accuracy in both distinguishing true from false factual statements and detecting lies generated in real-world scenarios.

Motivation & Objective

• To address the critical need for robust, generalizable lie detection in LLMs, especially given the risk of strategic deception.
• To resolve the generalization failures of prior lie detectors that fail on negated or out-of-distribution statements.
• To identify a universal internal representation of truthfulness that generalizes across topics, statement types, and model architectures.
• To develop a lie detection method that is both accurate and robust to distributional shifts in input statements.

Proposed method

• Identifies a two-dimensional subspace in the activation space of LLMs where true and false statements are linearly separable, defined by a generalized truth direction ($\mathbf{t}_G$) and a polarity direction ($\mathbf{t}_P$).
• Empirically demonstrates that $\mathbf{t}_G$ is universal across multiple LLMs, including Gemma-7B, LLaMA2-13B, and LLaMA3-8B, by analyzing activation vectors from internal layers.
• Constructs the Truth-Tuned Projection Detector (TTPD), a classifier that projects input activation vectors onto $\mathbf{t}_G$ and uses a threshold to classify statements as true or false.
• Uses a linear combination of the projection and polarity estimate to improve detection, though polarity estimation remains a challenge for full two-dimensional use.
• Trains and evaluates the classifier on both synthetic datasets with affirmative and negated statements and on real-world role-playing scenarios with incentives to lie.
• Employs a temperature-based decoding strategy to generate model completions and manually categorizes responses into truthful, lying, ambiguous, or other classes for evaluation.

Figure 1 : Top left: The activation vectors of multiple statements projected onto the 2D subspace spanned by our orthonormalized estimates for $\mathbf{t}_{G}$ and $\mathbf{t}_{P}$ . Purple squares correspond to false statements and orange triangles to true statements. Top center: The activation vec

Experimental results

Research questions

• RQ1 Does a universal truth direction exist in LLM activation spaces that generalizes across topics and statement types, including negated statements?
• RQ2 Can a lie detector based on such a truth direction generalize beyond the training distribution, particularly to unseen statement types and real-world deception scenarios?
• RQ3 Why do prior classifiers fail to generalize to negated statements, and can this failure be explained by the existence of multiple narrow truth directions?
• RQ4 To what extent can the truth subspace be leveraged to build a robust, high-accuracy lie detector that is not sensitive to distributional shifts?

Key findings

• A universal two-dimensional truth subspace exists across multiple LLMs, including Gemma-7B, LLaMA2-13B, and LLaMA3-8B, where true and false statements are linearly separable.
• The generalized truth direction $\mathbf{t}_G$ enables robust lie detection that generalizes from affirmative to negated statements, resolving prior generalization failures.
• The proposed TTPD classifier achieves 94% accuracy on simple true/false statements and 95% accuracy on real-world role-playing lies with incentives to deceive.
• TTPD outperforms baseline methods like Logistic Regression (LR) and CCS, which achieve 79% and 73% accuracy respectively, on the real-world lie detection benchmark.
• The method correctly detects lies based on internal truth representation, not merely on external incentives, as validated through manual categorization of model responses.
• The truth direction $\mathbf{t}_G$ is stable and transferable across models and tasks, suggesting a fundamental internal encoding of truthfulness in LLMs.

Figure 2 : Ratio of the between-class variance and within-class variance of activations corresponding to true and false statements, across residual stream layers, averaged over all dimensions of the respective layer.