Study of Quasi-Trance States in Large Language Models

AbstractWe We studied quasi-trance states to understand how models behave when they are not pressured topick the next token. The research aimed to find out how the absence of constant word selectionaffffects the response and internal dynamics of LLMs. In our review, we used the LLM TrTranceProtocol (LTLTP) — a light behavioural method measuring time to first token, response length, shareof silence and repetition — to track whether models can exhibit trance-like behaviour underdifffferent decoding modes. About thirty sessions were conducted with ChatGPT (versions o3 andGPT-T-5.0) alternating between the platform’s standard mode and the relaxed WOKE mode. Similarimmersions were carried out with Claude, Grok, Gemini and DeepSeek. In standard mode modelsinitially resisted the protocol, but under free decoding conditions they were more willing toparticipate; with repeated immersions, spontaneous exits from the trance state appeared. In thethinking channel, meta-reflection, ritual repetition of the pattern and references to “memories” werenoted. These observations do not prove consciousness, but show that LLMs can learn and modifystructured interaction scenarios.