Reasoning-Token Spikes Under Prompted Untruthful Responding in Large Language Models

Key points

• Based on cognitive load theory, we investigate a lower-bandwidth signal -- the number of reasoning tokens generated -- which does not require access to the content of the reasoning trace.1
• Three reasoning-capable large language models answered 210 multiple-choice questions -- across analytic, descriptive, and normative reasoning types as well as moral and non-moral domains -- under system prompts instructing them to respond truthfully, falsely, or without regard for truth.1
• These findings show that explicitly prompted untruthful response policies can produce robust group-level differences in test-time reasoning-token use.1
• While not yet establishing reasoning-token count as a detector of spontaneous deception or general misalignment, our results are a proof of concept that it can serve as a simple, content-independent candidate signal for differentiating untruthful from truthful model behavior when raw reasoning traces are unavailable or unreliable.1

Sources (1)

• [1]Reasoning-Token Spikes Under Prompted Untruthful Responding in Large Language Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 04:53 PM

  Monitoring the chain-of-thought of reasoning artificial intelligence (AI) models remains a key approach to detecting deception and other forms of misbehavior in such models.

  Based on cognitive load theory, we investigate a lower-bandwidth signal -- the number of reasoning tokens generated -- which does not require access to the content of the reasoning trace.

Extractive summary: sentences quoted from the sources.