spb/wpai0002_uqo

1# WP-AI-0002 — Research gap and positioning23*Human Text Reconstruction: Learning the Inverse of LLM-Induced Transformations in Scientific Writing*45Literature snapshot: 2026-11-09. All bibkeys refer to `paper/references.bib`; each entry was checked against an arXiv API or OpenAlex record, logged in `docs/literature/verified_sources.jsonl` (85 entries). Claims below about a paper's findings come from the abstract returned by the API at verification time. Claims about full-text details are marked *to check*.67---89## 1. How LLM rewriting changes scientific prose1011**Prevalence is large and rising, and it is measured through style rather than content.** Population-level estimators show steady growth in LLM-modified text in arXiv, bioRxiv and Nature-portfolio papers, with the fastest growth in computer science [liang2025quantifying]. The same framework applied to peer reviews finds a similar effect [liang2024monitoring]. Excess-vocabulary analysis of more than 15 million PubMed abstracts gives a lower bound of 13.5% of 2024 abstracts processed with LLMs, reaching 40% in some subcorpora [kobak2025delving]. Word-frequency models of about one million arXiv abstracts put the share of "LLM-style" computer-science abstracts at about 35%, using GPT-3.5's response to "revise the following sentences" as the baseline [geng2024chatgpt]. Keyword counts give a conservative estimate of at least 60,000 LLM-assisted papers in 2023 [gray2024chatgpt].1213**The signature is lexical, grammatical and rhetorical.**14- **Lexical.** Specific "focal" words (e.g. *delve*, *intricate*, *underscore*) are over-represented. Their source is unresolved, and RLHF is a plausible contributor [juzek2024delve]. Authors appear to co-evolve with detectors: *delve* fell after it was publicised, while *significant* kept rising [geng2025humanllm].15- **Grammatical and rhetorical.** Biber-feature analysis shows systematic human–LLM differences. These differences are larger for instruction-tuned than for base models [reinhart2025llms]. Comparable contrasts appear in news text [munozortiz2023contrasting].16- **Long-run arXiv shifts.** Across 823,798 arXiv abstracts, lexical complexity rises and syntactic complexity falls [bao2025examining].17- **Recent ML venues (2020–2025).** Formulaic phrases rose by up to 10× and hedging declined by 23% [jang2026scizoom].1819**Homogenisation is plausible but contested.** Co-writing with an instruction-tuned model reduces lexical and content diversity across authors [padmakumar2023writing]. In contrast, a 44-year analysis of CHI papers finds that prose did *not* homogenise: LLMs accelerated a word-level shift that had already started [kosch2026chi]. These results matter for us because they imply an era confound. Pre-2018 originals differ from post-2022 human text for reasons unrelated to LLMs, such as long-run readability decline [plavensigray2017readability], rising informality [hyland2016informal] and changing stance marking [hyland2016stance]. That confound is the reason WP-AI-0002 fixes originals to pre-2018 rights-verified text and generates the LLM side under control. Human reviewers and detectors also struggle to tell ChatGPT-generated scientific abstracts from real ones [gao2023comparing].2021**Gap (i).** Almost all of this evidence is *observational and corpus-level*: word frequencies, aggregate shares and feature means. No study we found measures, for the *same* passage, which features an LLM rewrite changes, by how much, and conditionally on discipline, section and rewrite instruction. RQ1 needs exactly that. The paired, controlled design (one original, five rewrite variants) closes this gap.2223## 2. Detection, paraphrase attacks and inversion attempts2425**Detectors are brittle under rewriting.** Zero-shot and trained detectors are well studied: [mitchell2023detectgpt], [bao2023fastdetectgpt], [hans2024binoculars], with watermarking in [kirchenbauer2023watermark]. Their limits are also well documented:26- **Paraphrasing defeats detection.** DIPPER cut DetectGPT's detection accuracy from 70.3% to 4.6% at 1% FPR [krishna2023paraphrasing]. Recursive paraphrasing evades watermarking, neural and retrieval detectors [sadasivan2023reliably].27- **Detectors are biased against non-native writers** [liang2023detectors].28- **Robustness collapses across generators and attacks** on shared benchmarks [dugan2024raid]. "Humanising" attacks are evaluated systematically in [zheng2025thbench].2930**Polished and revised text is the hard case.** AI-*polished* human text — the regime closest to our rewrite variants — is frequently flagged as AI-generated even when edits are minimal. Detectors also cannot separate degrees of AI involvement [saha2025polished]. Machine-*revised* text (rewriting, expansion, polishing) defeats standard detectors. Style-preference alignment (ImBD) improves AUC by 5–19 points [chen2024imitate]. Expert-edited machine outputs evade detection, while LLM-edited texts are rarely judged human [artemova2024beemo].3132**Inversion, the direct antecedent.** Rivera Soto, Chen and Andrews introduce **paraphrase inversion**: given a paraphrase, recover an approximation of the original. They treat it as translation from paraphrase back to source, trained on synthetic (original, paraphrase) pairs [riverasoto2024paraphrase]. They report three findings:331. GPT-4 and Llama-3 paraphrase with learnable biases.342. Inversion models generalise to paraphrasers unseen during training.353. Combining inversion with detectors gives an average +22% AUROC across seven detectors and three domains.3637A 2026 follow-up trains a style encoder to *reconstruct human-authored text from its machine paraphrase* while a semantic encoder is frozen. The frozen encoder pushes style into a separate channel. The learned style representations transfer to authorship verification [riverasoto2026unsupervised]. This is architecturally very close to our dual-latent content/style model. Both papers, however, treat inversion as a *means to detection* and report AUROC. They do not evaluate the reconstruction itself:38- no factual-integrity audit (numbers, citations, equations, negation, hedging);39- no stylometric distribution matching;40- no memorisation controls;41- no discipline or era stratification *(to check in the full text)*.4243**Related inversions.** Inversion of non-textual channels is now routine. Text can be reconstructed from embeddings [morris2023embeddings] and prompts from next-token distributions [morris2023lminversion]. Both show that "lossy" representations keep more recoverable information than expected, which supports the RQ2 prior that some pre-rewrite signal survives.4445**Adjacent style tasks.** Authorship obfuscation and style rewriting are the forward direction of our problem [fisher2024styleremix]. They rely on style representations that are separated from topic [riverasoto2021luar], [wegmann2022same], and on classical stylometry [burrows2002delta], [stamatatos2008survey], [koppel2008computational], [juola2008authorship]. Style-transfer research supplies:46- the paraphrase-based formulation we invert [krishna2020reformulating];47- disentangled and few-shot restyling [shen2017style], [fu2017style], [john2018disentangled], [riley2020textsettr];48- low-resource *authorship* style transfer, the closest analogue to "restore this author's voice" [patel2022lowresource];49- evaluation practice for style, content and fluency [mir2019evaluating], [jin2020survey];50- formality transfer data [rao2018gyafc].5152**Gap (ii).** Inversion exists as a detection auxiliary, but no published work, as far as we found:53- (a) evaluates **reconstruction fidelity as the target**, beyond identity and copy baselines;54- (b) uses **rights-verified, pre-LLM scientific originals** with controlled rewrite variants;55- (c) audits **factual drift** introduced or repaired by inversion;56- (d) tests **generalisation across rewrite prompts and LLM families** as an explicit research question (RQ4) rather than a robustness footnote.5758## 3. Non-identifiability: what cannot be recovered5960Inversion of a many-to-one map is ill-posed. Four strands of theory bound what any model can recover:61- **Unsupervised disentanglement is impossible without inductive biases on both model and data** [locatello2018challenging]. Nonlinear ICA is non-identifiable in general [hyvarinen1999nonlinear].62- **Identifiability returns with paired views.** If content is invariant across augmentations, the content block is identifiable up to an invertible map [vonkugelgen2021selfsupervised]. Contrastive objectives invert the generating process under similar assumptions [zimmermann2021contrastive]. Our (original, rewrite₁…₅) tuples are such paired views. In this setting *content* is the invariant part and *style* is what the LLM changes. The theory therefore suggests that content is identifiable, while the original style is identifiable only as far as the rewrite leaves traces of it.63- **Adversarial removal of an attribute is often incomplete.** Gradient reversal reaches chance-level adversary accuracy while a post-hoc probe still recovers the attribute [elazar2018adversarial]. This warns against reading "disentangled" off the adversary loss alone in our gradient-reversal / domain-separation objectives [ganin2014unsupervised], [bousmalis2016domain], [hu2017controlled].64- **Recursive training on synthetic text erodes the tails of the distribution** [shumailov2024collapse], [alemohammad2023mad], [dohmatob2024tails]. If a rewrite destroys tail stylistic variation, no inverse can restore it beyond the prior. A reconstruction model will then regress toward a "mean human style". Our stylometric distribution-matching metrics, rather than point metrics, are designed to detect this.6566**Gap (iii).** Prior inversion work does not separate *recoverable* from *non-identifiable* components, and it does not report where reconstruction amounts to hallucinating a plausible human style. WP-AI-0002 makes this an explicit evaluation axis through:67- the identity baseline;68- the multiple-rewrite consistency test;69- distributional (not only per-sample) stylometry;70- memorisation probes, because a pretrained decoder could "reconstruct" by recalling a memorised original [carlini2020extracting], [carlini2022quantifying], [lee2021dedup], [magar2022contamination], [sainz2023contamination].7172## 4. What WP-AI-0002 adds73741. **Inverse-mapping benchmark with rights-verified originals.** Pre-2018 human prose is drawn from open-licence scientific sources, with licence provenance logged per passage. Our corpus differs from general LLM corpora on two counts: licensing audits [longpre2023provenance] and documentation of large text corpora [gao2020pile], [soldaini2024dolma]. Scientific full-text corpora provide the parsing precedent [lo2019s2orc], [saier2023unarxive]. Five controlled Claude rewrite variants per passage support paired, per-passage measurement of LLM-induced change (RQ1, RQ3).752. **Dual-latent reconstruction model.** An encoder–decoder [vaswani2017attention], [raffel2019t5], [lewis2019bart] with copy and preservation mechanisms [see2017pointer], [gu2016copying]. Content and style latents are separated by gradient reversal, orthogonality [bousmalis2016domain] and contrastive [oord2018cpc], [chen2020simclr] objectives. Ablations go against from-scratch and style-conditioned baselines (RQ2, RQ5). Implementation choices follow [kudo2018sentencepiece], [xiong2020layernorm], [su2021roformer].763. **Factual-integrity evaluation.** We evaluate numbers, citations, equations, negation and hedging alongside lexical metrics ([popovic2015chrf], [papineni2002bleu], [lin2004rouge]) and semantic metrics ([zhang2019bertscore], [reimers2019sentencebert]). This answers RQ6 on factual distortions, a dimension none of the inversion papers report. Hedging is singled out because LLM-era prose shows reduced hedging [jang2026scizoom] and because stance marking is itself era-dependent [hyland2016stance].774. **Explicit non-identifiability and memorisation controls** (Section 3), and **cross-prompt / cross-family generalisation** as a primary RQ (RQ4). The generalisation result reported by [riverasoto2024paraphrase] is the hypothesis to replicate on scientific text.7879## 5. Closest prior work8081| bibkey | What they did | What they reported | Difference from WP-AI-0002 |82|---|---|---|---|83| riverasoto2024paraphrase | Paraphrase inversion as paraphrase→original translation on synthetic pairs | LLM paraphrase biases are learnable; generalises to unseen paraphrasers; +22% AUROC for detectors | Goal is detection, not fidelity; not scientific-specific; no factual / stylometric / memorisation audit |84| riverasoto2026unsupervised | Style encoder trained to reconstruct human text from machine paraphrase with frozen semantic encoder | Strong few/zero-shot AI-text detection; transfers to authorship verification | Near-identical content/style split, but reconstruction is a pretext task; no rights-verified scientific corpus or fidelity evaluation |85| krishna2020reformulating | Style transfer as paraphrase-then-restyle (STRAP) | Strong unsupervised style transfer via inverse paraphrasing per style | Target style is a corpus style, not the specific original passage |86| krishna2023paraphrasing | DIPPER paraphraser vs. detectors; retrieval defence | DetectGPT 70.3%→4.6% at 1% FPR | Forward attack; no inversion |87| saha2025polished | APT-Eval: 14.7k AI-polished human texts at graded involvement | Detectors over-flag minimal polishing and miss degrees | Measures detectability, not recoverability |88| chen2024imitate | ImBD: style-preference-aligned detection of machine-revised text | +13% AUC (open models), +5/+19% (GPT-3.5/4o) | Detects revision; does not undo it |89| artemova2024beemo | Expert-edited vs LLM-edited machine outputs benchmark | Expert editing evades detectors | Edit direction is human-over-machine; no inversion |90| patel2022lowresource | Low-resource authorship style transfer | Non-famous authors can be partially imitated | Imitates an author's style, not a specific original text |91| morris2023embeddings | Iterative inversion of text from embeddings | Embeddings reveal near-exact text | Inverts a representation, not an LLM rewrite |92| geng2024chatgpt / kobak2025delving | Corpus-level excess-word estimation of LLM-style in abstracts | ~35% LLM-style CS abstracts; ≥13.5% PubMed 2024 | Observational aggregate; no per-passage paired measurement or inversion |93| vonkugelgen2021selfsupervised | Identifiability of content vs style from paired views | Content block identifiable up to invertible map | Theory our design instantiates; no text inversion |9495**Assessment.** Rivera Soto et al. (2024, 2026) already show that LLM paraphrases can be partly inverted and that a reconstruction objective with a frozen semantic channel separates style. WP-AI-0002 must cite both prominently. Its novelty claim has to rest on four points:96- reconstruction fidelity as the object of study;97- rights-verified pre-2018 *scientific* originals with controlled variants;98- factual-integrity and non-identifiability evaluation;99- the discipline, era and section analysis.100101It should not claim to be first to invert LLM paraphrases.102103## 6. Unverified or rejected candidates (not in `references.bib`)104105- **Koehn (2004), "Statistical significance tests for machine translation evaluation"** (EMNLP; bootstrap resampling). Neither the OpenAlex nor the Crossref search returned the record. Verify via ACL Anthology W04-3250 before citing.106- **"Lu et al. (2019), 'Scientific writing is becoming harder to read'".** No matching record found. This probably conflates Plavén-Sigray et al. (2017), which *is* verified as `plavensigray2017readability`.107- **Hyland & Jiang, DOI 10.1016/j.esp.2017.06.001.** The DOI resolves to Larsson (2017), a different paper. It was replaced by the verified `hyland2016informal` (10.1016/j.esp.2016.09.001).108- **arXiv 2105.07263.** It resolved to Khan et al., "A Deep Metric Learning Approach to Account Linking", not LUAR. LUAR was verified instead via its EMNLP 2021 DOI (`riverasoto2021luar`).109- **GPTZero-type evaluations.** No peer-reviewed primary record was searched. GPTZero is covered indirectly in [krishna2023paraphrasing] and [chen2024imitate].110- **Non-bibliographic resources (no API record possible), to cite as URLs/footnotes after manual check:**111  - PMC Open Access Subset documentation;112  - arXiv licence documentation;113  - Creative Commons legal code (CC BY 4.0, CC0);114  - Claude model documentation.115- **Verified but left out for scope** (re-add if needed; records available from the verification run):116  - text-generation and dataset papers: Bender et al. 2021; Herbold et al. 2023 (Sci. Rep.); Ghostbuster (2305.15047); M4 (2305.14902); GLTR (1906.04043); Wu et al. survey (2310.14724);117  - paraphrase and style resources: ParaNMT (1711.05732); SCPN (1804.06059); INLP (2004.07667); Subramanian et al. (1811.00552); Prabhumoye et al. (1804.09000);118  - theory and evaluation: Khemakhem et al. iVAE (1907.04809); Mosteller & Wallace 1963; sacreBLEU (1804.08771);119  - homogenisation, collapse and attacks: Guo et al. linguistic-diversity collapse (2311.09807); Agarwal et al. homogenisation (2409.11360); JAMDEC (2402.08761); Lu et al. evasion (2305.10847); Kirchenbauer et al. watermark reliability (2306.04634); PADBen (2511.00416); Zhou et al. humanising attack (2404.01907);120  - recent prevalence work: Gray 2025 prevalence (2512.01560); Lee 2026 Korean excess vocabulary (2609.07447).121122**Metadata caveats.**123- arXiv-only entries carry the arXiv year and no venue. Several are published (e.g. T5 in JMLR, BART at ACL 2020, LUAR at EMNLP), so upgrade venues after a Crossref check before submission.124- OpenAlex years for Stamatatos (2008) and Juola (2008) are online-first or volume years. They differ from the commonly cited 2009 and 2006.125