, debating controversial statements), Simhi et al.
Aggregating internal signals reveals hidden agreement dynamics in LLM conversations
By pooling internal representations across conversations, the authors uncover a Markov process governing agreement and a power-law coupling between internal and external measures.
Big Tech
Adi Simhi · Yaniv Nikankin · Jonathan Herzig · Idan Szpektor · Yonatan Belinkov · Shay B. Cohen
Technion – Israel Institute of Technology · Google Research · University of Edinburgh
Research Digest··2 min read
The authors study two LLMs discussing subjective binary topics over multiple turns, using a linear classifier on hidden states to measure internal agreement.
§