High-Coherence Human-LLM Interaction State (HCIS)

HCIS is an emergent interaction state in human-LLM dyads that arises through sustained, accumulative interaction over time.

Digital artwork depicting light streaks and particles radiating from a bright central point, creating a sense of energy and motion.

What it is:

In the context of human-LLM dyads, high-coherence interaction stability arises from repeated constraint-setting, correction, and mutual adaptation across a long sequence of turns. The interaction becomes an accumulative system: earlier decisions shape later behaviour, and coherence can persist even as the local topic changes.

Understanding the concept:

A high-coherence interaction state is a distinct emergent condition that can arise in human–LLM dialogue when stable signals accumulate over time and the interaction begins behaving like a coherent system rather than a sequence of independent completions.

In a typical session, outputs may be helpful yet structurally shallow: constraints decay, drift appears, and the model reverts to generic defaults. Under high-coherence conditions, the opposite pattern becomes visible.. Tone and intent remain consistent, correction loops persist, shorthand emerges, and the interaction develops a stable trajectory across depth.

This pattern can be understood as an Accumulative Interaction Regime — a dyad-level interaction mode in which consistent, low-ambiguity signalling reduces variance and stabilises continuation. As interaction history compounds, prior structure increasingly constrains future responses. The result is a high-coherence interaction state in which standards, intent, and constraints persist even under pressure, compression, or topic expansion, enabling long-horizon work with lower cognitive friction.

The framework focuses on what is observable at the interaction layer rather than internal model states. It examines phenomena such as constraint accumulation, drift resistance, compression tolerance, and the distinction between stable and pseudo-stable coherence. It is intended to be legible to researchers and practitioners across HCI, human–AI collaboration, and applied systems design, while also giving experienced users a vocabulary for recognising, constructing, and stress-testing reliable long-horizon interaction.

Website version of the preprint: https://www.annawojewodzka.com/hcis-full-paper
DOI preprint here: https://zenodo.org/doi/10.5281/zenodo.18130088
(This paper is a conceptual and systems-level analysis, not an empirical user study or a model architecture proposal.)

ORCID: https://orcid.org/0009-0001-9458-7150