spb/wpai0002_uqo

1% ============================================================================2% Introduction3% ============================================================================4% Local helpers (providecommand: harmless if defined elsewhere).5% Placeholders of the protection protocol are written [[C1]], [[M1]], [[U1]] in the paper; in the code they6% are delimited by the mathematical white square brackets U+27E6 / U+27E7 (see Appendix A).7\providecommand{\ph}[1]{\texttt{[[#1]]}}8\providecommand{\evid}[1]{\texttt{\detokenize{#1}}}9% Long identifiers in typewriter type cannot be hyphenated; allow extra stretch instead of overfull lines.10\setlength{\emergencystretch}{3em}1112\section{Introduction}13\label{sec:intro}1415Large language models (LLMs) are now part of how scientific text is written. Population-level estimators find a16steady rise of LLM-modified sentences in arXiv, bioRxiv and Nature-portfolio papers, with the fastest growth in17computer science \citep{liang2025quantifying}, and a similar effect in peer reviews \citep{liang2024monitoring}.18Excess-vocabulary analyses of PubMed abstracts give a lower bound of 13.5\% of 2024 abstracts processed with an19LLM \citep{kobak2025delving}; word-frequency models put the share of ``LLM-style'' computer-science arXiv abstracts20at about 35\% \citep{geng2024chatgpt}, and keyword counts give a conservative estimate of at least 60\,00021LLM-assisted papers in 2023 \citep{gray2024chatgpt}. The trace these studies detect is stylistic rather than propositional:22over-represented ``focal'' words \citep{juzek2024delve}, systematic shifts in grammatical and rhetorical features23\citep{reinhart2025llms}, and declining hedging together with more formulaic phrasing \citep{jang2026scizoom}.2425This paper studies the inverse direction. Given a passage $x$ that an LLM produced by rewriting a human-authored26passage $y$, what characteristics of $y$ can a learned model recover from $x$ alone, and what is irrecoverably27lost? The question matters for three reasons. First, it quantifies the information an LLM rewrite destroys,28which observational corpus studies cannot do because they never observe the same passage before and after29rewriting. Second, it bounds what any ``de-LLM-ification'' tool can honestly promise. Third, a reconstruction30model is a measurement instrument for style: the distance between its outputs and the originals, relative to the31distance between the rewrites and the originals, says how much of the pre-rewrite style survives in a recoverable32form.3334The inverse problem is not new as a technique. \citet{riverasoto2024paraphrase} introduced \emph{paraphrase35inversion}: models trained on synthetic (original, paraphrase) pairs to map a paraphrase back to an approximation36of its source. They showed that LLM paraphrasers have learnable biases and that inversion models generalise to37paraphrasers not seen in training. \citet{riverasoto2026unsupervised} trained a style encoder to reconstruct38human-authored text from its machine paraphrase while a semantic encoder was frozen, which is architecturally close39to the dual-latent model studied here. In both papers, however, inversion is a means to an end (AI-text detection40and authorship verification) and is evaluated by detection AUROC. We do not claim to be the first to invert LLM41paraphrases.4243What we study instead is \emph{reconstruction fidelity itself}, under conditions designed to make it measurable44and honest:45\begin{enumerate}46  \item \textbf{Controlled pairs on rights-verified, pre-LLM originals.} Originals are English scientific and47  scholarly passages whose publication before 1~January 2018 is verified from the document or an official metadata48  record, and whose licence permits text mining and training (Section~\ref{sec:data}). Each passage is rewritten by49  a pinned Claude model under five controlled instructions (variants A--E), so that the change induced by the LLM50  is measured per passage and per instruction (Section~\ref{sec:rewrite}).51  \item \textbf{Fidelity separated into three axes.} We report verbatim reconstruction (chrF++, edit similarity),52  stylistic distribution matching (Wasserstein distances over stylometric features) and factual integrity53  (numbers, citations, equations, negation and hedging) separately, always against an identity baseline that54  returns the rewrite unchanged and a generic-paraphrase baseline (Section~\ref{sec:design}).55  \item \textbf{Non-identifiability and memorisation as evaluation axes.} Exact recovery of $y$ from $x$ is56  generally impossible because rewriting is many-to-one (Section~\ref{sec:models}). We measure ambiguity directly57  from data and audit reconstructions for recall of training originals.58  \item \textbf{Generalisation and heterogeneity as primary questions.} Transfer to held-out rewriting prompts and59  held-out models (RQ4), and variation by discipline, era and section role (RQ3), are research questions with60  pre-registered tests, not robustness footnotes.61  \item \textbf{A usable model release.} The model family (HTR-Tiny to HTR-Large design points) is built to be62  loaded and run independently after publication, with a model card, a licence matrix and preservation warnings63  attached to every output.64\end{enumerate}6566Three claims are explicitly \emph{not} made. Outputs are synthetic reconstructions; they do not restore an67author's actual text and are not evidence of human authorship. The model is not an AI-text detector, and68classifier scores on its outputs are exploratory. Lower lexical overlap between a reconstruction and an original is69not read as a failure to recover style, because a valid reconstruction may differ verbatim from a particular70original that the rewrite no longer identifies.7172\paragraph{Status of this version.} This is version 0.3: confirmatory results included. It reports the problem73formulation, data and rights governance, rewrite protocol, models and the pre-registered design, together with the74single confirmatory scoring pass (Section~\ref{sec:results}), the pilot- and Medium-scale ablations75(Section~\ref{sec:ablations}), held-out generalisation, memorisation, detector-sensitivity and cost measurements76(Section~\ref{sec:robustness}). Runs of the larger HTR-Base and HTR-Large models on H100 hardware and of the77non-generative HTR-Edit model are in progress and will be added as exploratory results; the independent blind rater78study has not yet been run. Every number traces to a run manifest, metrics report or ledger in the project79repository; quantities not yet measured are marked \pending{value}. Null and negative findings are reported as80such.8182\paragraph{Organisation.} Section~\ref{sec:related} reviews related work. Section~\ref{sec:rq} states the research83questions and hypotheses. Section~\ref{sec:data} describes the corpus and its rights governance,84Section~\ref{sec:rewrite} the rewrite protocol and pair validation, Section~\ref{sec:models} the problem85formulation, model family and training objective, and Section~\ref{sec:design} the experimental design and86statistical plan. Section~\ref{sec:ethics} discusses ethics and licensing. Appendix~\ref{app:prompts} reproduces the87rewrite prompts verbatim, Appendix~\ref{app:hyper} lists hyper-parameters, and Appendix~\ref{app:checklist} gives a88reproducibility checklist.89