Reasoning-Token Spikes Under Prompted Untruthful Responses in Large Language Models
These findings show that explicitly prompted untruthful response policies can produce robust group-level differences in test-time reasoning-token use, and are a proof of concept that reasoning-token count can serve as a simple, content-independent black-box candidate signal for differentiating untruthful from truthful...
Maverick Morales, T. Dominik, Vermut Gao et al.
· 0 citations