Fluid Interaction Protocols for Real-Time Knowledge Co-Construction Between Humans and Language Models

Abstract

This study proposes and evaluates a fluid interaction protocol for real-time knowledge co-construction between human users and language models. The protocol was designed for five interaction layers: intention capture, contextual alignment, generative response, reflective negotiation, and validation closure. Empirical testing across literature synthesis, policy interpretation, technical explanation, and decision-support scenarios showed that the proposed protocol outperformed baseline prompting and structured turn-taking across all evaluation dimensions. Interaction fluidity increased from 0.59 in baseline prompting to 0.81 under the fluid protocol, while the redundant turn ratio decreased from 0.31 to 0.12. Task completion improved from 76.4% to 93.2%, indicating that shorter interaction did not reduce artifact quality. Scenario-level results showed the highest coherence in technical explanation tasks at 0.89, the highest validation score in decision support at 0.90, and the strongest ambiguity reduction in policy interpretation at 0.88. Expert validation confirmed strong relevance, usability, and accuracy, with mean scores of 4.63, 4.56, and 4.37, respectively. Traceability received the lowest mean score at 4.02, showing that transparent claim history remains a key design challenge. The integrated performance profile further showed that the fluid protocol achieved higher scores in coherence, adjustment efficiency, validation reliability, and artifact quality than comparison conditions. These findings demonstrate that language model collaboration requires more than prompt optimization. Reliable knowledge co-construction depends on adaptive interaction structures that regulate feedback, correction, validation, and shared epistemic control.