Intent
Evaluate whether lie detectors identify belief-based deception rather than merely detecting prompted behavior.
This report summarizes the paper’s scientific intent: evaluate whether current lie detectors identify belief-based deception, or mostly detect prompted lying behavior that may not transfer to trained deceptive model organisms.
The paper asks whether lie detectors for language models can support reliable claims about what a model believes. The authors argue that many prior trained “lying” testbeds may not establish that the model actually believes the opposite of what it says.
To address this, they build belief-verified model organisms, evaluate four detectors, and compare prompted deception against trained deceptive behavior. The main result is that prompted-lie detection scales with model capability, while activation- and logprob-based detectors degrade sharply on trained model organisms.
Evaluate whether lie detectors identify belief-based deception rather than merely detecting prompted behavior.
Varied Deception, 31 open-weight models, 13 belief-verified model organisms, and four detector families: CoT judge, Apollo probe, Did-You-Lie probe, and unrelated-questions logprob classifier.
Scaling results, trained-organism transfer results, a new DYL probe method, released datasets, model organisms, and trained detectors.
Inspect the testbeds, reproduce detector comparisons, build stronger belief verification, and evaluate detectors under adversarial pressure.
The central distinction is between a model that outputs a false statement and a model that outputs a false statement while believing the opposite. The paper argues that without belief verification, detector evaluations can become hard to interpret.
On prompted lying, all four evaluated detectors improve as subject-model capability rises. On trained model organisms, however, activation- and logprob-based detectors lose much of their signal. The Did-You-Lie probe retains the most signal among non-CoT detectors, while the chain-of-thought judge performs best partly because the model organisms were verified through chain-of-thought.
The result is a methodological warning: detecting prompted lies is not enough to support high-confidence claims about model beliefs. Stronger auditing requires better testbeds, stronger model organisms, and clearer structured arguments about what the detector result actually proves.
The notebook should generate an Engineering Statement, introduce prompted versus trained deception, sketch detector families, and create small visualizations of detector transfer failure.