Abstract
This work investigates whether LLMs understand and deploy strategic vagueness, with implications for AI safety. Steven Pinker’s theory of indirect speech posits that humans use ambiguous language as a rational strategy when facing audiences with conflicting interests, enabling coordination with allies while maintaining plausible deniability against adversaries. We investigate whether LLMs exhibit similar capabilities: can they produce language that evades a trusted monitor while successfully coordinating with an accomplice?
Bio
Pramod Kaushik is a research associate in FBK Italy and a student researcher at the University of Trento. Previously, he has worked at TRDDC Pune, Inria Bordeaux and Columbia Uni. His recent work on the theory of LLM Sampling won the best paper award at ACL last year. He has previously worked on decision making agents and on understanding human decision making and building neurocomputational models of the brain.