One person is
many voices.

Every point of light is a message I wrote — texts, emails, Slack, scholarship. A model trained to read how people write, not what they write about, sorts a single life into its registers.

scroll
The benchmark story

Writing benchmarks measure two things.

Ask any leaderboard how well a model writes and it answers with grammaticality — is the sentence well-formed? — and what we'd call truthiness: a coherent, readily intelligible utterance, optimized for frictionless comprehension and graded by factual Q&A. Both reward training pipelines rich in semantic data for knowledge domains — not the practice of reasoning in language, and not the shape of the utterance that delivers it.

Dimension · solved

Grammaticality

Well-formed, fluent, error-free. Models cleared this bar years ago; no frontier lab differentiates on it anymore.

Dimension · saturating

“Truthiness”

Coherent, intelligible, factually scoreable. Knowledge Q&A — biology, law, the academic canon — with one gold answer per question.

MMLUGPQATruthfulQA MT-BenchChatbot Arena knowledge & fluency leaderboards — none score voice6,7

The missing measurement isn't harder science. It's just not objective.

MMLU · knowledge Q&A (2020) GPQA Diamond · “Google-proof” (2023) expert baselines
100%75 5025 20202022 20232024 20252026 human experts · 89.8 (MMLU) PhD experts · ≈70 (GPQA) tone & style · intent-judgment no benchmark exists — no line to draw GPT-3 · 43.9 (2020) Flan-PaLM · 75.2 (2022) GPT-4 · 86.4 (2023) Gemini Ultra · 90.0 (2024) Claude Opus 4 · 92.0 (2025) GPT-4 · 38.8 (Nov 2023) o1 · 78 (2024) Gemini 2.5 Pro · 84 (2025) Gemini 3.1 Pro · 94 (2026) passes the PhD experts within a year of release saturated → labs move to MMLU-Pro 94 92
Fig. 1 — Best publicly reported frontier score by year. MMLU6 reached the human-expert ceiling in four years; GPQA Diamond7, written by PhDs to be un-googleable, passed its own expert baseline within one. Hugging Face archived the Open LLM Leaderboard in 2025 after saturation. Scores: llm-stats · Artificial Analysis.
The missing axes

Preference is a dimension of language.

Two things human writing does that no benchmark measures — both live under one umbrella: preference. And this isn't a new idea: corpus linguistics established forty years ago that register variation across situations rivals variation across people8.

Axis one

Tone & style

Register, rhythm, capitalization, punctuation, code-switching — how you sound, and how that changes between your group chat, your inbox, and your thesis. Content-independent, and measurable against a person1,2,9.

Axis two

Intent & judgment

Taste. What to say and what to leave out, when to hedge and when to be blunt, how much warmth a situation calls for. The reasoning inside the voice.

Neither has an objective gold answer. Both have a subjective one: the person. The eval anchor isn't an answer key — it's a corpus.

4.04.5 5.05.5 mean word length involved markers — pronouns & contractions per 100 words 2.55 7.510 Biber Dimension 1 informational pole involved pole academic · 1.5 involved/100w · 5.32 mean word length email · 9.2 involved/100w · 4.34 slack · 7.7 involved/100w · 4.51 text · 8.9 involved/100w · 4.00 academic email slack text 6× fewer involved markers
Persona Vectors · Anthropic '25 — traits as static directions5
IRiS · USTC '26 — persona neurons are situational; clusters 14.8 apart10
PsPLUG · TikTok '26 — persona = residual vs the base model11
PerFine · UMass/Adobe '25 — critique-refine against the user12
Fig. 2 — One author's four registers placed on Biber's8 Dimension 1 (involved ↔ informational), computed from this corpus using his feature loadings: pronouns + contractions vs word length. A 6× swing within one person — register variation rivaling individual variation, exactly as Multi-Dimensional Analysis predicts. Below: 2025–26 work converging on the same situational thesis.
Why not just train it in?

p ≫ n. You can't pretrain a voice.

The diversity of human tone — every person × every register × every situation demanding judgment — can never be a training set. Any one voice is radically underdetermined: billions of parameters against one person's words. So personalization is a post-training problem.

So we measured. Here is what one person's writing actually looks like.

The geometry

What you say is not
how you say it.

The same messages, embedded two ways. Semantic space — where search, RAG, and every retrieval system lives — organizes by topic. Style space1 organizes by voice. Flip the question and watch the same points reorganize.

Register separation
silhouette, full vectors

Three people, one medium

Voice is a fingerprint.

Hold the medium constant — iMessages from three of us — and ask each space: who wrote this? Semantic embeddings see one crowd talking. Style embeddings pick Ben's voice out instantly — and notice something true: Sam and Ahmet actually do text alike. The map measures resemblance; it doesn't invent difference.

Ben's voice separation
silhouette, full vectors
Style vs semantic
cleaner on Ben's voice
Stylometrics

The signature, made legible.

The same separation the embedding sees, in features a person can read. Bars are scaled to the loudest register per feature.

Inside the map

fly the embedding space live
Dataset
Space  ·  HOW (style) ⟷ WHAT (semantic)
Color by
Layers
Points
Selected message
click a point to inspect it
drag · orbit    shift-drag · pan    scroll · zoom
Grounding

References

  1. Wegmann, Schraagen & Nguyen (2022). Same Author or Just Same Topic? Towards Content-Independent Style Representations. ACL RepL4NLP. AnnaWegmann/Style-Embedding — the style stream in every figure here.
  2. Patel et al. (2024). StyleDistance: Stronger Content-Independent Style Embeddings with Synthetic Parallel Examples. StyleDistance/styledistance — second style model we triangulate with in eval.
  3. Dettmers, Pagnoni, Holtzman & Zettlemoyer (2023). QLoRA: Efficient Finetuning of Quantized LLMs. arXiv:2305.14314.
  4. Schulman et al., Thinking Machines Lab (2025). LoRA Without Regret. thinkingmachines.ai/blog/lora — LoRA matches full fine-tuning sample efficiency when applied to all layers; per-user adapter training is no longer the bottleneck.
  5. Chen, Arditi, Sleight, Evans & Lindsey (2025). Persona Vectors: Monitoring and Controlling Character Traits in Language Models. arXiv:2507.21509 — character traits as directions in activation space; the representation-level cousin of what these figures show in embedding space.
  6. Hendrycks et al. (2020). Measuring Massive Multitask Language Understanding. arXiv:2009.03300 — MMLU; estimated human-expert accuracy 89.8%.
  7. Rein et al. (2023). GPQA: A Graduate-Level Google-Proof Q&A Benchmark. arXiv:2311.12022 — PhD experts score ≈70% on the Diamond set frontier models now clear.
  8. Biber, D. (1988). Variation across Speech and Writing. Cambridge University Press — Multi-Dimensional Analysis: factor analysis of 67 linguistic features; register variation across situations rivals variation across individuals.
  9. Yang & Carpuat (2025). Steering Large Language Models with Register Analysis for Arbitrary Style Transfer. arXiv:2505.00679 — Biber's dimensions as LLM rewriting instructions.
  10. Wei, Li, Wang & Deng (2026). Beyond Static Personas: Situational Personality Steering for Large Language Models. arXiv:2604.13846 — IRiS: situational vs global persona neurons cluster 14.8 apart; no static intervention steers all contexts.
  11. Styles + Persona-plug = Customized LLMs (2026). arXiv:2601.06362 — personalization as the residual between the user's distribution and the base model's.
  12. Iterative Critique-Refine Framework for Enhancing LLM Personalization (2025). arXiv:2510.24469 — PerFine: profile-grounded critic loops.

All figures: one user's real corpus, extracted by the production self-study pipeline, embedded locally (style) and via Gemini (semantic), PCA-projected for display with scores computed on full vectors. Built with the same clustering code that ships in the product.