Statistical Analysis Reveals Kimi's Writing Style Closely Mirrors Claude

Summary

A technical analysis by researcher maxloh has used cross-entropy and linguistic pattern analysis to compare the writing styles of major large language models. The study found that Kimi's responses exhibit remarkable similarity to those of Claude in their statistical language patterns. Using character trigrams, word unigrams, and bigrams, the analysis creates a heatmap showing which LLM models write alike based solely on their response text. The methodology employs interpolated probability models backed off to pooled background distributions, allowing the research to distinguish between model-specific linguistic signatures and common patterns across language models.

The research demonstrates that despite being trained independently, different state-of-the-art language models converge on surprisingly similar linguistic patterns. This convergence likely stems from shared training data characteristics, similar optimization objectives, and architectural similarities in how transformer-based LLMs process and generate language.

• Modern LLMs trained independently converge toward similar writing styles, suggesting common training dynamics and architectural effects

Editorial Opinion

This research offers a fascinating window into the convergence of LLM outputs. It suggests that modern language models, despite different training approaches and datasets, develop remarkably similar linguistic fingerprints. This finding has important implications for understanding whether LLM capabilities are driven by fundamental architectural constraints or whether similarity reflects shared training paradigms. The work also raises questions about whether internet-scale LLMs inevitably converge toward similar distributions.