Beyond Content: Surface Matching, Temporal Divergence, and the Limits of Synthetic Baselines

Read full text Summarise with VIRAT
Authors: Kristina Baldovino
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This preprint examines whether the temporal structure observed in long-form Human-LLM interaction can be reproduced by a stateless synthetic baseline matched on broad surface properties. Building on Beyond Content: Temporal Dynamics of Long-Form Human-LLM Interaction (Baldovino, 2026), the study compares a 25-thread empirical archive against five synthetic runs generated with a surface-matched null designed to preserve broad interactional constraints while excluding persistent identity, biography, long-horizon memory, explicit goals, and user-specific adaptation. The synthetic baseline closely approximated assistant-side bulk behavior, but did not recover the broader user-side variability, amplification structure, or recovery dynamics observed in the empirical archive. Coupled metrics showed the same pattern: rolling amplification remained comparatively flat in the synthetic condition, while the empirical archive displayed stronger positional variation and broader excursions. Lagged coupling analyses provided secondary, directionally supportive evidence of differential local interaction structure. The signficance of this result is methodological. Most current LLM evaluation focuses on single-turn outputs or short conversational windows. By contrast, this study asks whether a surface matched synthetic baseline can reproduce the temporal structure of interaction that unfolds over hundreds of turns. Within this archive and this baseline family, substantial surface matching alone was insufficient to reproduce the observed temporal morphology of long-form Human-LLM interaction. The results also suggest that constructing a strong synthetic null for long-form interaction is itself a nontrivial methodological problem.

Fields of Study & Subjects

Speech and dialogue systems Language Discourse Communication Strategies Cybernetics and Technology in Society

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
21 March 2026
Language
EN
Document type
Preprint