Sarin Nhek — CNR-INO #
Coherence in natural and artificial texts #
Long-range coherence is a defining property of written language, persisting across entire books through stable statistical scaling and structured information flow. We show that this property is organized along two nearly orthogonal directions of variability. In this multivariate space, the mean representation of natural books follows a smooth trajectory parameterized by text length.

Using a contemporary LLM with an iterative self-conditioning protocol, we generate long texts that converge toward stable coherence regimes and occupy regions of descriptor space comparable to those of human-written books at similar lengths. Under highly stochastic generation, statistical scaling descriptors remain within natural-text ranges, whereas informational descriptors systematically deviate, indicating a differential robustness across coherence measures.

These results show that contemporary autoregressive language models can reproduce key large-scale statistical structures of written language, while revealing distinct stability regimes governing different aspects of coherence under chaotic generation.