Transfer of loop operators to instruction-tuned dialogue models

Determine whether the habituation and interruption operators studied for forced open-ended continuation transfer to instruction-tuned language models when used in dialogue.

Background

The core experiments use base LLMs in a deliberately constrained forced-continuation setting, rather than instruction-tuned models in ordinary dialogue. The paper notes that instruction tuning introduces dialogue formatting and preference training, both of which may alter open-ended continuation behavior. Although one post-trained Qwen3-8B model is examined in raw continuation, the study does not test whether the operators work when such models are used as intended in interactive dialogue.

References

An instruction-tuned model brings a dialogue format and preference training that change what an open-ended continuation is; whether the operators studied here transfer to it is left open (Section~\ref{sec:limitations}).

— Interrupting the Loop: Periodic Subject Changes Raise Judged Surprise and Connection in Base Language Models  (2608.19893 - Filho, 20 Aug 2026) in Section 2, subsection “Measurement stack” (Method)