Everyday Style Imitation (Enron Email) - Authorship Attribution Top-5: leaderboard
Tests whether a language model can imitate an ordinary person's writing style from a few of that person's real texts. Wang et al. (EMNLP 2025 Findings) give each model five random writing samples by one author plus a GPT-4.1 summary of a held-out text, and ask it to write that text in the author's voice at temperature 0. Enron: one-to-one internal emails by 150 employees, 3,884 test texts. Two authorship-attribution classifiers (Longformer-base-4096 and ModernBERT-base) trained on the corpus's human texts rank every candidate author; a generation counts as correct when its true author is among the top five, averaged over both classifiers. On human test texts the classifiers reach 79.8% top-5 accuracy. Top-5 attribution accuracy (%), five-shot; higher is better.
Metric: AA Top-5 Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 27B | 62.38 | #606 |
| 2 | GPT-4o | 59.65 | #357 |
| 3 | Gemini 2.0 Flash | 59.56 | #351 |
| 4 | GPT-4o Mini | 56.45 | #598 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=everyday-style-imitation-enron-email-authorship-attribution-top-5 · How It Works · Data refreshed daily, snapshot 2026-11-05.