safety security
Untruthful Prompts Used More Reasoning Tokens
Three models spent more tokens when explicitly told to lie or ignore truth, a group-level signal that is not yet a deception detector.
Summary
Three models spent more tokens when explicitly told to lie or ignore truth, a group-level signal that is not yet a deception detector.
Across 210 multiple-choice questions, truth-directed responses used fewer reasoning tokens than lie-directed or truth-indifferent responses in all three tested reasoning models. The result does not establish instance-level detection, spontaneous deception or robustness under adversarial pressure; it identifies a content-independent candidate signal for future study.
Why it matters
Three models spent more tokens when explicitly told to lie or ignore truth, a group-level signal that is not yet a deception detector.
Limits and context
- The result does not establish instance-level detection, spontaneous deception or robustness under adversarial pressure; it identifies a content-independent candidate signal for future study.
Key claims
Three models spent more tokens when explicitly told to lie or ignore truth, a group-level signal that is not yet a deception detector.
Qualification: The result does not establish instance-level detection, spontaneous deception or robustness under adversarial pressure; it identifies a content-independent candidate signal for future study.
Evidence: source-2026-10-08-017
Sources
- arXiv preprint 2610.10405arXiv · primary research
Corrections
No corrections have been recorded for this story.