Paper deep dive

AuthBench: A Large-Scale Multilingual Benchmark for Authorship Representation across Genres and Lengths

MaoXun Huang, Zhenxing Zhang, Claire Cardie

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 113

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%

Last extracted: 9/12/2026, 3:14:25 AM

Summary

The paper introduces AuthBench, a large-scale multilingual benchmark for authorship representation comprising 428,150 documents by 153,825 authors across 10 languages, 9 primary genres, and 4 length buckets. It supports authorship attribution (retrieval) and verification tasks. Evaluating 47 neural models and 3 baselines, the study finds that retrieval and verification favor different model families (e.g., multilingual-e5-large for retrieval, llama3.1-8b-instruct for verification) and that performance varies significantly by language, genre, and length, indicating that robust authorship representation remains an open problem.

Entities (12)

AuthBench · benchmark · 100%LLaMA3.1-8B-Instruct · model · 95%multilingual-e5-large · model · 95%Authorship Attribution · task · 90%Authorship Verification · task · 90%Claire Cardie · person · 90%Cornell University · organization · 90%EER · metric · 90%MaoXun Huang · person · 90%ROC-AUC · metric · 90%Success@5 · metric · 90%Zhenxing Zhang · person · 90%

Relation Signals (11)

AuthBench → supportstask → Authorship Attribution

confidence 95% · AuthBench supports two complementary tasks: authorship attribution... and authorship verification

AuthBench → supportstask → Authorship Verification

confidence 95% · AuthBench supports two complementary tasks: ... authorship verification

multilingual-e5-large → achievesbestintask → Authorship Attribution

confidence 90% · multilingual-e5-large achieves the best overall retrieval performance

llama3.1-8b-instruct → achievesbestintask → Authorship Verification

confidence 90% · llama3.1-8b-instruct achieves the best overall verification performance

AuthBench → affiliatedwith → Cornell University

confidence 90% · Department of Computer Science, Cornell University

AuthBench → createdby → Zhenxing Zhang

confidence 90% · AuthBench: A Large-Scale Multilingual Benchmark... MaoXun Huang Zhenxing Zhang Claire Cardie

AuthBench → createdby → MaoXun Huang

confidence 90% · AuthBench: A Large-Scale Multilingual Benchmark... MaoXun Huang Zhenxing Zhang Claire Cardie

AuthBench → createdby → Claire Cardie

confidence 90% · AuthBench: A Large-Scale Multilingual Benchmark... MaoXun Huang Zhenxing Zhang Claire Cardie

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Authorship signals matter in settings where writing style carries identity: digital forensics, plagiarism analysis, account linking, misinformation investigation, and machine-generated text detection. Yet current authorship benchmarks remain fragmented, usually covering only a narrow language set, a single genre, or a limited document-length regime, which makes it difficult to assess whether modern representations truly generalize. We introduce AuthBench, a large-scale multilingual benchmark for authorship representation that is designed to make this evaluation broad, standardized, and realistic. AuthBench contains 428,150 documents written by 153,825 individuals across ten widely used languages, 9 primary genres, 66 fine-grained genres, and four document-length buckets. It supports two complementary tasks: authorship attribution, formulated as same-author retrieval and authorship verification, formulated as same-author binary decision. We benchmark 47 neural models and three non-neural baselines under a unified zero-shot protocol. Results show that authorship representation remains far from solved: the best retrieval model reaches only 0.258 Success@5, while the best verification model achieves 0.076 EER and 0.968 ROC-AUC. The leaderboard also reveals a meaningful task split, with different model families leading retrieval and verification, and large performance differences across languages, genres, and lengths. These findings position AuthBench not only as a new benchmark, but as a diagnostic resource for studying when and why authorship representations succeed or fail. We release AuthBench, its evaluation toolkit, and benchmark data at this https URL and this https URL.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

109,288 characters extracted from source content.

Expand or collapse full text

AuthBench: A Large-Scale Multilingual Benchmark for Authorship Representation across Genres and Lengths MaoXun Huang Zhenxing Zhang Claire Cardie Department of Computer Science, Cornell University Ithaca, NY, United States September 9, 2026 Abstract Authorship signals matter in settings where writing style carries identity: digital foren- sics, plagiarism analysis, account linking, misinformation investigation, and machine-generated text detection. Yet current authorship benchmarks remain fragmented, usually covering only a narrow language set, a single genre, or a limited document-length regime, which makes it diﬀicult to assess whether modern representations truly generalize. We introduce AuthBench, a large-scale multilingual benchmark for authorship representation that is designed to make this evaluation broad, standardized, and realistic. AuthBench contains 428,150 documents written by 153,825 individuals across ten widely used languages, 9 primary genres, 66 fine- grained genres, and four document-length buckets. It supports two complementary tasks:au- thorship attribution, formulated as same-author retrieval andauthorship verification, formu- lated as same-author binary decision. We benchmark 47 neural models and three non-neural baselines under a unified zero-shot protocol. Results show that authorship representation remains far from solved: the best retrieval model reaches only 0.258 Success@5, while the best verification model achieves 0.076 EER and 0.968 ROC-AUC. The leaderboard also re- veals a meaningful task split, with different model families leading retrieval and verification, and large performance differences across languages, genres, and lengths. These findings po- sition AuthBench not only as a new benchmark, but as a diagnostic resource for studying when and why authorship representations succeed or fail. We release AuthBench, its eval- uation toolkit, and benchmark data at https://github.com/mao-code/AuthBenchandhttps: //huggingface.co/datasets/MaoXun/AuthBench. 1 Introduction Authorship representation [ 17,18,53] examines whether a model can encode author-specific lin- guistic and stylistic features into reusable text representations, which is an important capability of current state-of-the-art (SOTA) models. This ability has a wide range of applications, including cybersecurity and digital forensics [ 1,9,23,31], and underpins additional tasks such as plagiarism detection [6,40] and machine-generated text detection [2,3,27,43]. Driven by significant progress in large language models (LLMs) [ 16,57,60], SOTA models have become increasingly generalizable across downstream tasks, languages, and varying input lengths. However, existing resources still broaden authorship evaluation along only one or two axes at a time. 1 arXiv:2609.06771v1 [cs.CL] 6 Sep 2026 BenchmarkYear AuthorsDocs ScopeCoverageEvaluation Multi- lingual Multi- genre Length bins Standard splits Attribution Verification Blogs50[59]20065066K English blogs777737 MUD (adapted corpus) [ 20,44] 2021 1.07M 321.7M English Reddit777777 MARC[19]2020 1.15M1.26M Multilingual reviews377377 PAN 2021 AV[33]2021 278.2K 73.0K pairs English fanfiction777373 PAN 2022 AV[34]2022112 22.7K pairs English cross-discourse737373 PAN 2023 AV[35]2023112 18.5K pairs English written/spoken737373 CROSSNEWS[26]2025 2.26K1.15M English news + social737333 IMDb62[48]20116262K English reviews777737 CMCC[14]200821756 English email/essay/blog/chat 737737 Guardian[51]2013131.0K English opinion/book reviews 737737 Enron Email(adapted corpus) [21] 2004∼150∼500K English email777777 SMAuC(adapted corpus) [8] 2023 5.66M3.36M Scientific writing733777 AIDBench(identification benchmark) [55] 2024 3.25K51.5K English papers/e- mail/blogs/reviews 737377 AuthBench (ours)2026 153.8K 428.2K Social, news, literature, blogs, etc. 333333 Table 1: Comparison of AuthBench with representative prior authorship resources. Table1summarizes representative resources: classic corpora such as blogs and email collections are often monolingual and domain-specific; PAN established standardized evaluation protocols, but as a family of yearly shared tasks, each edition usually focuses on a narrow task or discourse setting; and more recent resources typically expand coverage along a single dimension, such as multilinguality, cross-discourse evaluation, cross-genre journalism, or document-length control. Consequently, prior work still lacks a unified benchmark that simultaneously offers multilingual coverage, broad genre diversity, explicit length control, and standardized support for both large-scale attribution and verification. To address these limitations, we introduce AuthBench, a large-scale multilingual benchmark for authorship representation that supports attribution and verification across genres and document lengths. AuthBench combines author-linked text gathered through our own public-web crawling pipelines with carefully refined portions of existing datasets and benchmark resources, all stan- dardized into a shared schema (see Appendix E). After quality filtering, deduplication, language auditing, and balanced sampling, the current release contains 428,150 documents of diverse lengths (short, medium, long, and extra-long), authored by 153,825 individuals across ten widely used lan- guages (en, zh, hi, es, fr, ar, ru, de, ja, ko), spanning nine primary genres with 66 fine-grained sub-genres. AuthBench supports two complementary evaluation tasks: (i)authorship attribution, formulated as an open-world same-author retrieval task evaluated with ranking-based metrics (Suc- cess@K, Recall@K, nDCG@K, and MRR); and (i)authorship verification, formulated as binary decision over query–candidate pairs and evaluated with EER and ROC-AUC. We release standard- ized dataset splits and a unified evaluation toolkit to facilitate reproducible and comprehensive evaluation. We conduct a comprehensive zero-shot evaluation of 47 neural models and three non-neural base- lines on AuthBench, spanning embedding models, base LLMs, and instruction-tuned variants. The leaderboard reveals a clear task split:multilingual-e5-largeachieves the best overall retrieval performance (0.258 S@5, 0.254 R@5, 0.217 nDCG@5, 0.220 MRR), whilellama3.1-8b-instruct achieves the best overall verification performance (0.076 EER, 0.968 ROC-AUC). Moreover, model quality still varies substantially across languages, document lengths, and genres, highlighting per- sistent challenges for robust authorship modeling. 2 2 Related Work Authorship representation.Authorship representation [17,18,53] seeks to identify persistent author-specific signals in written documents. The field is commonly structured around two core eval- uation tasks: (i) authorship attribution, which links a document to its author and is traditionally studied as closed-set classification [32,50] but is increasingly instantiated in representation-learning settings as open-world same-author retrieval over a candidate pool [44,54]; and (i) authorship verification, which determines whether two documents were written by the same author [32,34]. In parallel, authorship representation learning develops document embeddings that encode stylis- tic signals and can be used to support both attribution and verification at scale. Recent work therefore commonly reports retrieval metrics such as Recall@K or MRR for retrieval-style attri- bution, while standardized verification benchmarks emphasize ROC-based metrics such as AUC and representation-oriented analyses also report EER as a compact summary of the false-accept / false-reject trade-off [34,35,54]. Authorship Benchmark Development.Early authorship resources were instrumental for the field, but they were typically small-scale, monolingual, and tied to a narrow domain [32,50]. Classic datasets such as blog and email corpora [21,47] enabled important methodological progress, yet they offered limited coverage in language diversity, genre breadth, and explicit length control, and many were not released with broadly adopted benchmark splits. A second challenge is frag- mentation: several resources widely used in modern work, including Blogs50, MUD, Enron, and SMAuC, are adapted from broader source collections rather than introduced as unified authorship benchmarks. PAN later established influential shared-task protocols for authorship analysis [ 7,41], but each edition typically focuses on a single task and discourse setting. More recent resources broaden the space only partially: MARC adds multilingual reviews [ 19], PAN 2022 and PAN 2023 study controlled cross-discourse verification in English [ 34,35], CROSSNEWS targets cross-genre journalism [ 26], SMAuC emphasizes scientific writing with length control [8], and AIDBench cen- ters on identification-style evaluations [55]. What is still missing is a single benchmark that jointly supports multilingual, multi-genre, length-aware, standardized evaluation for both retrieval-style attribution and verification. AuthBench is designed to fill that gap. 3 AuthBench In this section, we first describe the AuthBench construction in Section3.1. Then we present the statistics of the AuthBench in Section 3.2. 3.1 Benchmark Construction Figure2summarizes the AuthBench construction workflow, and Figure3gives a quick view of the final record format. AuthBench is built from two complementary inputs: author-linked text gathered through our public-web crawling pipelines, and selected existing datasets or benchmark resources that add useful languages, genres, or length regimes (AppendixE). We then standardize everything through five stages: Build & Normalization, Quality Filtering, Redundancy Reduction, Language Audit, and Bucket Balanced Sampling. Stage 1 maps every source into a shared schema, chunks overly long documents, and caps per-author document counts so heterogeneous materials become comparable without letting prolific authors dominate the benchmark. The later stages turn that broad collection into a reliable evaluation benchmark. Quality Filter- ing removes noisy, low-information, and script-mismatched text; Redundancy Reduction removes exact duplicates, near-text overlaps, and optional near-author overlaps to reduce leakage; and the 3 Category#Docs Share Overall Total428,150 100.0% Languages10– Authors153,825– Avg. docs per author2.78– Genres (fine-grained)66– By language English (en)97,974 22.9% Spanish (es)33,395 7.8% Chinese (zh)55,368 12.9% French (fr)31,225 7.3% German (de)39,813 9.3% Arabic (ar)42,091 9.8% Russian (ru)66,084 15.4% Japanese (ja)21,494 5.0% Korean (ko)33,881 7.9% Hindi (hi)6,825 1.6% By length Short24,476 5.7% Medium211,910 49.5% Long184,447 43.1% Extra-long7,317 1.7% Table 2: Statistics of our AuthBench. Language Audit verifies or retags suspicious labels. Bucket Balanced Sampling then applies hier- archical targets over language, genre, and length bucket before writing stratified train/dev/test splits, keeping the final release diverse, balanced, and reproducible. In short, the pipeline converts both crawled and inherited resources into a unified benchmark optimized for consistent authorship evaluation. Complete implementation-aligned details are provided in Appendix A. 3.2 Statistics of AuthBench After construction, AuthBench comprises 428,150 documents authored by 153,825 individuals across 10 languages: English (en), Spanish (es), Chinese (zh), French (fr), German (de), Arabic (ar), Rus- sian (ru), Japanese (ja), Korean (ko), and Hindi (hi). AuthBench spans 9 primary genres, including social_media,literature,news,blog,media_reviews,poetry,ecommerce_reviews,qna, and research_paper. The distribution is dominated bysocial_media(40.9%),literature(30.0%), andnews(17.2%), with the remaining six primary genres accounting for the final 11.9% of docu- ments. Documents in AuthBench are further categorized into four token-length buckets: short (1 10 tokens), medium (11100), long (101500), and extra-long (>500). An overview of AuthBench statistics is presented in Table 2. The current split materialization contains 198,345 query docu- ments and 229,805 candidate documents, partitioned into train/dev/test as 156,335/21,008/21,002 queries and 186,184/21,813/21,808 candidates. Additional detailed statistics are provided in Ap- pendix D. 4 10 languages Languages Arabic 9.8% German 9.3% Korean 7.9% Spanish 7.8% French 7.3% Other 6.6% Chinese 12.9% Russian 15.4% English 22.9% 9 primary, 66 sub Genres Literature 30.0% News 17.2% Other 6.7% Blog 5.3% Social media 40.9% Entertainment 3.9% YouTube comments 16.8% 4 buckets Document Length Long 43.1% Extra long 1.7% Short 5.7% Medium 49.5% Figure 1: Overview of the current AuthBench profile for document proportion across languages, genres and document lengths. Figure1complements Table2with a compact visual summary of the benchmark profile. The language-level counts show that AuthBench remains broadly distributed across all ten languages rather than collapsing into a single dominant language, while the language–genre pie-chart reveals clear source-driven heterogeneity in genre coverage across languages. The token-length plot further shows that most documents concentrate in the medium-to-long range, with shorter texts and extra- long documents appearing less frequently. Together, these views highlight that AuthBench is broad in coverage but intentionally preserves realistic distributional variation across languages, genres, and lengths. 3.3 Pipeline Summary Build & Normalization Redundancy Reduction Quality Filtering Bucket Balanced Sampling AuthBench Source Copora Language Audit Figure 2: AuthBench construction pipeline. Full details are provided in AppendixA. 4 Experiment Setup Evaluation details.We embed queries and candidates with each models encoder (or the final hidden states of LLM backbones) using a unified pipeline: tokenize each input, extract the last hidden states, pool into a single vector (mean pooling by default), and L2-normalize the resulting embedding. Similarities are computed with cosine similarity between normalized embeddings. We use full-sized within-split candidate pool; verification metrics are computed over all same-author and non-matching query–candidate pairs in that pool. The complete experiment setup is defined in Appendix A.7. Models.We evaluate 47 neural models together with three non-neural baselines: a character3–5 gram TF-IDF cosine baseline [45], a lightweight stylometric n-gram baseline inspired by Koppel & 5 "candidate_id": "doc_116326", "author_id": "43b5e789...c39320", "lang": "en", "genre": "social_media/people", "content": "How recently are these things you found? People have pasts ... How would you like it?", "source": "exorde", "token_length": 156 "candidate_id": "doc_274629", "author_id": "c423340f...30add", "lang": "zh", "genre": "news", "content": " ⌋ ...", "source": "babel_briefings", "token_length": 74 Figure 3: Examples in our AuthBench. Schler [22], and a PPM-style compression baseline following Teahan & Harper [52]. These systems complement the neural leaderboard with strong lexical and compression-based reference points. Full model details are provided in Table7of AppendixA.7. Metrics.AuthBench supports both authorship attribution and authorship verification tasks. For authorship attribution, we report Success@5 (S@5), Recall@5 (R@5), nDCG@5, and MRR; these respectively capture shortlist utility, coverage of multiple same-author targets, ranking quality near the top of the list, and the rank of the first correct same-author hit. For authorship verifi- cation, we compute EER and ROC-AUC; EER summarizes the balanced operating point where false acceptance and false rejection are equal, while ROC-AUC measures threshold-independent separability between same-author and different-author pairs. The main leaderboard reports all six aggregate metrics, while the slice tables below focus on S@5 for compactness. Formal definitions of all evaluation metrics are provided in Appendix A.8. 5 Experimental Results All results in this section are reported on the AuthBench test split, which contains 21,002 queries and 21,808 candidates. In this section, we first present the overall performance of SOTA models on AuthBench in Section 5.1. We then provide a detailed analysis of performance broken down by language, genre, and document length. Finally, in Section 5.5, we clarify the scope of the current zero-shot study and outline post-training as future work rather than as a reported main result. 5.1 Main Results Finding 1: retrieval and verification favor different model families. Table 3shows a clear split between ranking and verification ability. For retrieval, the strongest model ismultilingual-e5-large, reaching 0.258 S@5, 0.254 R@5, 0.217 nDCG@5, and 0.220 MRR. The top base LLMs are extremely close:llama3-8bandllama3.1-8breach 0.254 and 0.253 S@5, respectively. This narrow gap suggests that large generative backbones already encode strong authorship cues, but specialized multilingual embedding models remain slightly better at turning those cues into stable nearest-neighbor structure. Verification tells a different story.llama3.1-8b-instructis the strongest verifier with 0.076 EER and 0.968 ROC-AUC, followed closely byllama3-8b-instructandllama3-8b. In contrast, the best retrieval model is not the best verifier. This is an important takeaway for future work: au- 6 ModelModel Size S@5↑R@5↑nDCG@5↑MRR↑ROC-AUC↑EER↓ LLMs (instruction-tuned) llama3-8b-instruct8B0.251 0.2470.206 0.2090.9670.080 llama3.1-8b-instruct8B0.2470.2430.2040.2080.968 0.076 qwen2.5-7b-instruct7.6B0.207 0.2030.168 0.1720.964 0.083 qwen3-4b-instruct4B0.197 0.1930.159 0.1630.958 0.094 LLMs (base) llama3-8b8B0.254 0.2500.210 0.2130.967 0.079 llama3.1-8b8B0.2530.2490.2090.2120.9670.080 deepseek-llm-7b-base7B0.220 0.2160.180 0.1840.954 0.105 qwen3-4b4B0.210 0.2060.171 0.1750.960 0.087 Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.242 0.2380.202 0.2050.9550.096 gte-qwen2-7b-instruct7.6B0.2400.2350.1980.2020.961 0.078 Embedding models multilingual-e5-large559.9M0.258 0.2540.217 0.2200.920 0.157 multilingual-e5-base278M0.250 0.2450.2090.2120.918 0.161 qwen3-embedding-8b7.6B0.241 0.2370.201 0.2030.9400.135 sfr-embedding-mistral7.1B0.240 0.2360.201 0.2050.956 0.096 Lexical / non-neural baselines tfidf–0.197 0.1940.167 0.1830.8370.219 ngram–0.1700.1670.1450.1490.874 0.213 ppm–0.157 0.1560.137 0.1420.792 0.298 Table 3:Main results on AuthBench (test split). The retrieval leaderboard now reports S@5, R@5, nDCG@5, and MRR; the verification leaderboard reports both ROC-AUC and EER.Bold entries mark the best result within each model group for a metric, and underlined entries mark the second best. The full 50-model evaluation, including all three non-neural baselines, is deferred to AppendixB. thorship retrieval and authorship verification should not be treated as interchangeable probes of the same representation quality. Retrieval rewards global neighborhood structure, whereas verification also depends on pairwise calibration and decision-boundary sharpness. The non-neural baselines remain informative but clearly behind the best neural systems.tfidf is the strongest lexical retrieval baseline at 0.197 S@5, whilengramis the strongest lexical verifier at 0.213 EER and 0.874 ROC-AUC. The gap to the neural frontier is still substantial, indicating that modern encoders capture authorial regularities beyond surface lexical overlap. At the same time, even the best neural scores remain far from ceiling, which suggests that robust authorship representation is still an open problem rather than a solved byproduct of general text embedding quality. 5.2 Results by Language Finding 2: performance varies sharply across languages, and the best model differs by language.Table4shows that there is no single universally best model across languages. multilingual-e5-largeleads on Spanish, French, Japanese, Korean, and Russian;qwen3-embedding-8b is strongest on Chinese;llama3-8bis strongest on English;llama3.1-8bleads on German and Hindi; and the lexicaltfidfbaseline is narrowly strongest on Arabic. This pattern suggests that 7 ModelModel Sizear de en es fr hi ja ko ru zh LLMs (instruction-tuned) llama3-8b-instruct8B0.2260.2440.261 0.265 0.2690.2890.2430.1730.1640.382 llama3.1-8b-instruct8B0.2110.2450.2590.2650.2650.2970.2290.177 0.1650.373 qwen2.5-7b-instruct7.6B0.182 0.210 0.201 0.208 0.214 0.215 0.215 0.143 0.143 0.343 qwen3-4b-instruct4B0.187 0.195 0.187 0.203 0.198 0.229 0.212 0.143 0.143 0.304 LLMs (base) deepseek-llm-7b-base7B0.183 0.216 0.238 0.223 0.220 0.238 0.200 0.144 0.131 0.367 llama3-8b8B0.2210.2470.270 0.272 0.2700.2890.2420.1830.1650.380 llama3.1-8b8B0.225 0.2490.2700.2720.2660.297 0.246 0.186 0.1730.361 qwen3-4b4B0.189 0.217 0.204 0.213 0.214 0.227 0.220 0.142 0.147 0.336 Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.223 0.241 0.237 0.254 0.250 0.2720.2230.1640.1640.396 gte-qwen2-7b-instruct7.6B0.2200.2410.2320.2350.2390.2520.2490.1560.1710.396 Embedding models multilingual-e5-base278M0.2510.210 0.226 0.2670.2620.238 0.2530.1850.1880.413 multilingual-e5-large559.9M0.2520.221 0.2310.284 0.280 0.263 0.268 0.196 0.1900.422 qwen3-embedding-8b7.6B0.2390.2420.209 0.239 0.241 0.244 0.251 0.169 0.1700.425 sfr-embedding-mistral7.1B0.221 0.2360.2360.255 0.248 0.2580.214 0.162 0.162 0.397 Lexical / non-neural baselines ngram–0.2380.137 0.126 0.162 0.1660.244 0.1600.1680.1320.267 ppm–0.229 0.1400.1490.1900.1720.2550.113 0.103 0.118 0.184 tfidf – 0.254 0.148 0.146 0.190 0.155 0.280 0.226 0.184 0.154 0.346 Table 4:Results by language (S@5). The updated slice table keeps a compact representative subset; complete language-wise S@5 and EER tables are reported in Appendix B. multilingual authorship representation is not just a matter of scaling one architecture, but of match- ing representation bias to the linguistic and genre composition of each language slice. The spread in diﬀiculty is also large. Chinese is comparatively easy, withqwen3-embedding-8b reaching 0.425 S@5, whereas Russian and Korean remain diﬀicult even for the best systems, topping out at 0.190 and 0.196. These differences imply that benchmark-wide averages can hide substantial cross-lingual brittleness. Future research should therefore treat multilingual authorship modeling as a robustness problem, not simply as macro-averaged multilingual transfer. A second notable trend is that multilingual embedding models dominate much of the non- English landscape. Their advantage over LLMs is especially visible in Chinese, Spanish, French, Japanese, Korean, and Russian. This suggests that explicit multilingual representation alignment remains highly valuable for authorship retrieval, and that future progress may come from improving language-sensitive embedding geometry rather than relying on larger generative models alone. 5.3 Results by Genre Finding 3: genre diﬀiculty varies dramatically, and different genres favor different architectures.Table 5reveals large genre-dependent differences in diﬀiculty. The easiest slice isresearch_paper, whereqwen3-embedding-8breaches 0.838 S@5, followed byqnaat 0.682 and poetryat 0.667. At the other extreme,media_reviewsandecommerce_reviewsare much harder, with best S@5 values of only 0.088 and 0.119. This gap indicates that authorship signals are much easier to recover in domains with stronger personal regularity or more stable discourse conventions than in short, noisy, or highly template-driven review settings. 8 ModelModel Sizeblog ecommerce_reviews literature media_reviews news poetry qna research_paper social_media LLMs (instruction-tuned) llama3-8b-instruct8B0.2130.1070.5210.070 0.302 0.639 0.6500.7690.206 llama3.1-8b-instruct8B0.2200.1120.5180.035 0.2930.583 0.6180.7870.204 qwen2.5-7b-instruct7.6B0.1620.0840.4960.035 0.242 0.583 0.5670.7190.167 qwen3-4b-instruct4B0.1490.0800.4620.0530.227 0.6110.5480.6880.160 LLMs (base) deepseek-llm-7b-base7B0.2150.1140.4860.035 0.257 0.639 0.5480.7190.176 llama3-8b8B0.2290.1190.5240.070 0.311 0.667 0.6310.8000.207 llama3.1-8b8B0.2320.1120.5320.0530.3050.6670.6310.8120.207 qwen3-4b4B0.1660.0840.4850.035 0.241 0.639 0.5610.7380.171 Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.2020.1100.5140.0530.2620.5830.5800.7250.207 gte-qwen2-7b-instruct7.6B0.1910.0910.5380.0350.281 0.611 0.5860.7810.197 Embedding models multilingual-e5-base278M0.1460.1070.4530.0530.2580.500 0.6310.7560.230 multilingual-e5-large559.9M0.1600.1000.4790.0530.2680.4440.6820.7690.236 qwen3-embedding-8b7.6B0.1510.0910.5360.0880.256 0.5560.5860.8380.209 sfr-embedding-mistral7.1B0.1990.1050.5120.053 0.2570.5830.5670.7560.207 Lexical / non-neural baselines ngram–0.0660.0500.2890.0180.1610.2780.3250.4310.168 ppm–0.0900.0640.2300.000 0.149 0.250 0.4200.6310.153 tfidf–0.0890.0530.3230.035 0.169 0.444 0.4590.6130.198 Table 5:Results by primary genre (S@5). Complete genre-wise S@5 and EER tables are reported in AppendixB. The identity of the best model also changes sharply by genre.gte-qwen2-7b-instructis strongest onliterature;qwen3-embedding-8bis strongest onmedia_reviewsandresearch_paper; multilingual-e5-largeleads onqnaandsocial_media; and thellama3family leadsblog,news, poetry, andecommerce_reviews. Rather than pointing to one dominant architecture, the table suggests that different model families are capturing different kinds of authorial evidence. A useful way to interpret this pattern is through discourse structure. Embedding models appear especially strong in structured, information-dense settings such asresearch_paperandqna, where topical organization and repeated compositional habits may be easier to preserve in fixed-vector spaces. LLMs remain very competitive in more stylistically expressive genres such asblog,news, andpoetry, where broader contextual modeling may better preserve subtle stylistic signatures. Future authorship benchmarks and methods should therefore pay closer attention to genre as a first-class modeling variable rather than a secondary reporting slice. 5.4 Performance by Length Finding 4: longer documents are markedly easier than short ones, but the best model still depends on length regime.Table 6presents the updated S@5 breakdown across four document lengths. The strongest short-document model ismultilingual-e5-baseat 0.187 S@5, the strongest medium-document model ismultilingual-e5-largeat 0.220, and the strongest long and extra-long models arellama3.1-8bat 0.439 and 0.543, respectively. The length trend is strong and intuitive: authorship retrieval becomes much easier as more text is available. Extra-long documents are the easiest regime overall, and the gain is especially pronounced for the LLMs. For example,llama3.1-8brises from 0.135 S@5 on short documents to 0.543 on extra-long ones. This suggests that much of the remaining challenge in authorship modeling comes from sparse-evidence settings, where models must identify stable stylistic cues from very limited text. The strongest model also changes with length. Short and medium documents favor embedding 9 ModelModel Sizeshort medium long extra_long LLMs (instruction-tuned) llama3-8b-instruct8B0.150 0.199 0.4320.539 llama3.1-8b-instruct8B0.1420.1960.4270.531 qwen2.5-7b-instruct7.6B0.103 0.159 0.3790.496 qwen3-4b-instruct4B0.081 0.150 0.3680.488 LLMs (base) deepseek-llm-7b-base7B0.121 0.169 0.3930.528 llama3-8b8B0.150 0.2020.4360.524 llama3.1-8b8B0.1350.2020.4390.543 qwen3-4b4B0.101 0.161 0.3850.496 Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.161 0.1960.3980.524 gte-qwen2-7b-instruct7.6B0.137 0.1920.4090.500 Embedding models multilingual-e5-base278M0.1870.2130.3760.417 multilingual-e5-large559.9M0.182 0.2200.3920.476 qwen3-embedding-8b7.6B0.157 0.1970.3950.504 sfr-embedding-mistral7.1B0.168 0.193 0.3950.520 Lexical / non-neural baselines ngram–0.1510.1360.2710.331 ppm–0.079 0.134 0.2550.240 tfidf–0.157 0.164 0.3000.437 Table 6:Results by length bucket (S@5). Complete length-wise S@5 and EER tables are reported in AppendixB. models, withmultilingual-e5-baseandmultilingual-e5-largeleading those buckets, whereas long and extra-long documents favorllama3.1-8b. This suggests a useful modeling hypothesis for future work: compact embedding models may be better at extracting robust local stylistic cues when evidence is scarce, while larger LLMs benefit more from longer contexts that expose higher- order discourse and syntactic habits. Verification follows the same overall pattern. The best EER improves from 0.104 onshort to 0.078 onmedium, 0.060 onlong, and 0.048 onextra_long. Short documents therefore remain the clearest bottleneck for both retrieval and verification. Advancing this regime will likely require methods explicitly designed for low-evidence authorship signals, rather than simply scaling existing encoders. 5.5 Post-Training Outlook In this paper, AuthBench serves as a large-scale pre-training evaluation benchmark for a broad set of authorship representation models, including embedding models, base LLMs, and instruction-tuned variants. This zero-shot comparison provides a useful foundation for future researchers to choose strong base models under different languages, genres, lengths, and task settings before performing post-training. We are also conducting follow-up analyses on post-training behavior and alternative authorship methods, and we plan to release a second version of the paper with a broader set of 10 post-training results. 6 Conclusion We introducedAuthBench, a large-scale benchmark for evaluating authorship representations across languages, genres, and document lengths under a unified retrieval-and-verification framework. Our zero-shot results show that authorship representation is still far from solved, that retrieval and verification favor different model families, and that robustness across languages, genres, and especially short documents remains the central challenge. We hope AuthBench provides a strong foundation for future work on more reliable and better calibrated authorship modeling in realistic multilingual settings. Limitations A key limitation of AuthBench is that language, genre, and document length are not fully orthogonal axes in the current release. Although we intentionally collected diverse sources and applied a robust filtering, deduplication, and auditing pipeline to reduce avoidable bias, some residual covariance across these axes is diﬀicult to eliminate in practice. As a result, certain slice differences may still reflect source-level correlations rather than authorship diﬀiculty alone. In addition, despite our leakage-reduction pipeline, some residual leakage may remain at scale. We view these as important targets for future benchmark refinement. References [1]Ahmed Abbasi and Hsinchun Chen. Writeprints: A stylometric approach to identity-level iden- tification and similarity detection in cyberspace.ACM Transactions on Information Systems, 26(2):7:1–7:29, 2008. doi: 10.1145/1344411.1344413. [2]Sergey K. Aityan, William Claster, Karthik Sai Emani, Sohni Rais, and Thy Tran. A lightweight approach to detection of AI-generated texts using stylometric features, 2025. [3]Maged S. Al-Shaibani and Moataz Ahmed. Arabic machine-generated text detection: Sty- lometric analysis and cross-model evaluation.Expert Systems with Applications, 305:130644, 2026. doi: 10.1016/j.eswa.2025.130644. [4]arXiv. arxiv bulk data access. Website, n.d. URLhttps://arxiv.org/help/bulk_data. Accessed: 2025-12-22. [5]barilan. Blog authorship corpus. Hugging Face Datasets, n.d. URLhttps://huggingface.co/ datasets/barilan/blog_authorship_corpus. Accessed: 2025-12-22; canonical reference: [46]. [6]Alberto Barrón-Cedeño, Marta Vila, M. Antònia Martí, and Paolo Rosso. Plagiarism meets paraphrasing: Insights for the next generation in automatic plagiarism detection.Computa- tional Linguistics, 39(4):917–947, 2013. doi: 10.1162/COLI_a_00153. [7]Janek Bevendorff, Bilal Ghanem, Anastasia Giachanou, Mike Kestemont, Enrique Manjava- cas, Martin Potthast, Francisco Rangel, Paolo Rosso, Günther Specht, Efstathios Stamatatos, Benno Stein, Matti Wiegmann, and Eva Zangerle. Shared tasks on authorship analysis at pan 2020. InAdvances in Information Retrieval, ECIR 2020, volume 12036 ofLecture Notes in 11 Computer Science, p. 508–516. Springer, 2020. doi: 10.1007/978-3-030-45439-5\_34. URL https://doi.org/10.1007/978-3-030-45439-5_34. [8]Janek Bevendorff, Philipp Sauer, Harrisen Scells, Benno Stein, et al. SMAuC — the scien- tific multi-authorship corpus.Proceedings of the ACM/IEEE Joint Conference on Digital Libraries (JCDL), 2023. doi: 10.1109/JCDL57899.2023.00013. URLhttps://doi.org/10.1109/ JCDL57899.2023.00013. [9]Aylin Caliskan-Islam, Richard Harang, Andrew Liu, Arvind Narayanan, Clare Voss, Fabian Yamaguchi, and Rachel Greenstadt. De-anonymizing programmers via code stylometry. In 24th USENIX Security Symposium (USENIX Security 15), p. 255–270, 2015. [10]Cornell University. arxiv dataset. Kaggle Dataset, n.d. URLhttps://w.kaggle.com/ datasets/Cornell-University/arxiv. Accessed: 2025-12-22; see also [4]. [11]S. Dhanwal et al. An annotated dataset of discourse modes in hindi stories. InProceedings of the 12th Language Resources and Evaluation Conference (LREC), 2020. URL https:// aclanthology.org/2020.lrec-1.149/. [12]Exorde Labs. Exorde social media (december 2024, week 1). Hugging Face Datasets, 2024. URLhttps://huggingface.co/datasets/Exorde/exorde-social-media-december-2024-week1. Ac- cessed: 2025-12-22. [13]fengzhujoey. Douban dataset: Rating, reviews, side information. Kaggle Dataset, n.d. URL https://w.kaggle.com/datasets/fengzhujoey/douban-datasetratingreviewside-information. Accessed: 2025-12-22. [14]Jade Goldstein-Stewart, Kerri Goodwin, Roberta Sabin, and Ransom Winder. Creating and using a correlated corpus to glean communicative commonalities. InProceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08), Mar- rakech, Morocco, 2008. European Language Resources Association (ELRA). URL https: //aclanthology.org/L08-1198/. [15]Google Developers. Youtube data api v3 documentation.https://developers.google.com/ youtube/v3/docs. Accessed: 2026-03-15. [16]Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024. [17]Nudrat Habib, Tosin Adewumi, Marcus Liwicki, and Elisa Barney. Trends and challenges in authorship analysis: A review of ml, dl, and llm approaches.arXiv preprint arXiv:2505.15422, 2025. [18]Baixiang Huang, Canyu Chen, and Kai Shu. Authorship attribution in the era of llms: Prob- lems, methodologies, and challenges.ACM SIGKDD Explorations Newsletter, 26(2):21–43, 2025. [19]Phillip Keung, Yichao Lu, György Szarvas, and Noah A. Smith. The multilingual amazon reviews corpus. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. URLhttps://arxiv.org/abs/2010.02573. Commonly released as MARC / Amazon Reviews Multi. 12 [20]Aleem Khan, Elizabeth Fleming, Noah Schofield, Marcus Bishop, and Nicholas Andrews. A deep metric learning approach to account linking, 2021. URLhttps://arxiv.org/abs/2105. 07263. [21]Bryan Klimt and Yiming Yang. The enron corpus: A new dataset for email classification research. InMachine Learning: ECML 2004, volume 3201 ofLecture Notes in Computer Science, p. 217–226. Springer, 2004. doi: 10.1007/978-3-540-30115-8\_22. URLhttps://doi. org/10.1007/978-3-540-30115-8_22. [22]Moshe Koppel and Jonathan Schler. Authorship verification as a one-class classification problem. InProceedings of the Twenty-First International Conference on Machine Learn- ing (ICML 2004), p. 489–495, Banff, Canada, 2004. ACM. doi: 10.1145/1015330.1015448. URLhttps://icml.c/Conferences/2004/proceedings/papers/415.pdf. [23]Srijan Kumar, Justin Cheng, Jure Leskovec, and V. S. Subrahmanian. An army of me: Sock- puppets in online discussion communities. InProceedings of the 26th International Conference on World Wide Web (W), 2017. doi: 10.1145/3038912.3052677. [24]Felix Leeb. Babel briefings. Hugging Face Datasets, 2023. URLhttps://huggingface.co/ datasets/felixludos/babel-briefings. Accessed: 2025-12-22. [25]Felix Leeb and Bernhard Schölkopf. A diverse multilingual news headlines dataset from around the world. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2024. URL https://arxiv.org/abs/2403. 19352 . Dataset: Babel Briefings. [26]Marcus Ma, Duong Minh Le, Junmo Kang, Yao Dou, John Cadigan, Dayne Freitag, Alan Ritter, and Wei Xu. CROSSNEWS: A cross-genre authorship verification and attribution benchmark. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, 2025. URL https://ojs.aaai.org/index.php/AAAI/article/view/34659. [27]Alicja Martinek and Ewelina Bartuzi-Trokielewicz. Detecting deepfakes and false ads through analysis of text and social engineering techniques. Manuscript, 2024. URL https://ai.gov. pl/media/2024/12/FAKES___LANG__COLING_15_09_-2.pdf. Text-based detection of AI-generated deepfake advertisement transcripts using linguistic and stylometric features. [28]mdanok. Arabic poetry dataset. Kaggle Dataset, n.d. URLhttps://w.kaggle.com/datasets/ mdanok/arabic-poetry-dataset. Accessed: 2025-12-22. [29]mexwell. Amazon reviews multi. Kaggle Dataset, n.d. URLhttps://w.kaggle.com/datasets/ mexwell/amazon-reviews-multi. Accessed: 2025-12-22; cite [19] as the canonical dataset paper. [30]MIDAS Lab, IIIT-Delhi. Hindi discourse analysis dataset. GitHub Repository, n.d. URLhttps: //github.com/midas-research/hindi-discourse. Accessed: 2025-12-22; canonical reference: [11]. [31]Arvind Narayanan, Hristo S. Paskov, Neil Zhenqiang Gong, John Bethencourt, Emil Ste- fanov, Eui Chul Richard Shin, and Dawn Song. On the feasibility of internet-scale author identification. In2012 IEEE Symposium on Security and Privacy, p. 300–314, 2012. doi: 10.1109/SP.2012.46. [32]Tempestt J. Neal, Kalaivani Sundararajan, Ayesha Fatima, Yue Yan, Yang Xiang, and Damon Woodard. Surveying stylometry techniques and applications.ACM Computing Surveys, 50(6), 2017. doi: 10.1145/3132039. URLhttps://doi.org/10.1145/3132039. 13 [33]PAN / Webis Group. Pan at clef 2021: Authorship verification. Oﬀicial task website, 2021. URLhttps://pan.webis.de/clef21/pan21-web/author-identification.html. Open-set cross-domain authorship verification task page. [34]PAN 2022 Authorship Verification Organizers. Overview of the authorship verification task at pan 2022. Task overview, 2022. URLhttps://publications.aston.ac.uk/id/document/89399. Oﬀicial overview used for the discourse-level statistics reported in the task. [35]PAN 2023 Authorship Verification Organizers. Overview of the authorship verification task at pan 2023. Task overview, 2023. URLhttps://ceur-ws.org/Vol-3497/paper-199.pdf. Oﬀicial overview covering the 2023 cross-discourse benchmark. [36]PleIAs. French public domain books (french-pd-books). Hugging Face Datasets, 2024. URL https://huggingface.co/datasets/PleIAs/French-PD-Books. Accessed: 2025-12-22. [37]PleIAs. German public domain corpus (german-pd). Hugging Face Datasets, 2024. URL https://huggingface.co/datasets/PleIAs/German-PD. Accessed: 2025-12-22. [38]PleIAs. Russian public domain corpus (russian-pd). Hugging Face Datasets, 2024. URL https://huggingface.co/datasets/PleIAs/Russian-PD. Accessed: 2025-12-22. [39]PleIAs. Spanish public domain books (spanish-pd-books). Hugging Face Datasets, 2024. URL https://huggingface.co/datasets/PleIAs/Spanish-PD-Books. Accessed: 2025-12-22. [40]Martin Potthast, Matthias Hagen, Anna Beyer, Matthias Busse, Martin Tippmann, Paolo Rosso, and Benno Stein. Overview of the 6th international competition on plagiarism detection. InWorking Notes Papers of the CLEF 2014 Evaluation Labs (CLEF 2014), 2014. [41]Martin Potthast, Sarah Braun, Tolga Buz, Fabian Duffhauss, Florian Friedrich, Jörg Marvin Gülzow, Winfried Lötzsch, Fabian Müller, Maike Elisa Müller, Robert Paßmann, Bernhard Reinke, Lucas Rettenmeier, Thomas Rometsch, Timo Sommer, Michael Träger, Sebastian Wil- helm, Shlomo Argamon, Moshe Koppel, Efstathios Stamatatos, Benno Stein, and Matthias Ha- gen. Who wrote the web? revisiting influential author identification research applicable to in- formation retrieval. InAdvances in Information Retrieval, ECIR 2016, volume 9626 ofLecture Notes in Computer Science, p. 393–407. Springer, 2016. doi: 10.1007/978-3-319-30671-1\_29. URL https://doi.org/10.1007/978-3-319-30671-1_29. [42]Project Gutenberg. Project gutenberg offline catalogs.https://w.gutenberg.org/ebooks/ offline_catalogs.html . Accessed: 2026-03-15. [43]Karol Przystalski, Jan K. Argasiński, Iwona Grabska-Gradzińska, and Jeremi K. Ochab. Sty- lometry recognizes human and LLM-generated texts in short samples, 2025. [44]Rafael A. Rivera-Soto, Olivia Elizabeth Miano, Juanita Ordonez, Barry Y. Chen, Aleem Khan, Marcus Bishop, and Nicholas Andrews. Learning universal authorship representations. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, p. 913–919. Association for Computational Linguistics, 2021. URL https://aclanthology.org/ 2021.emnlp-main.70. [45]Gerard Salton and Christopher Buckley. Term-weighting approaches in automatic text retrieval. Information Processing & Management, 24(5):513–523, 1988. doi: 10.1016/0306-4573(88) 90021-0. URL https://doi.org/10.1016/0306-4573(88)90021-0. 14 [46]Jonathan Schler, Moshe Koppel, Shlomo Argamon, and James W. Pennebaker. Ef- fects of age and gender on blogging. InAAAI Spring Symposium on Com- putational Approaches to Analyzing Weblogs, 2006. URLhttps://aaai.org/papers/ s06-03-013-effects-of-age-and-gender-on-blogging/ . Blog Authorship Corpus. [47]Jonathan Schler, Moshe Koppel, Shlomo Argamon, and James W. Pennebaker. Effects of age and gender on blogging. InAAAI Spring Symposium on Computational Approaches to Analyzing Weblogs, p. 199–205, 2006. URLhttps://w.aaai.org/Library/Symposia/Spring/ 2006/s06-03-039.php. [48]Yanir Seroussi, Ingrid Zukerman, and Fabian Bohnert. Authorship attribution with topic mod- els.Computational Linguistics, 40(2):269–310, 2014. doi: 10.1162/COLI_a_00173. URLhttps: //direct.mit.edu/coli/article/40/2/269/1452/Authorship-Attribution-with-Topic-Models. [49]Stack Exchange Meta.Data dumps releases timeline, updates, andclarification.https://meta.stackexchange.com/questions/396597/ data-dumps-releases-timeline-updates-and-clarification , 2025. Accessed: 2026-03-15. [50]Efstathios Stamatatos. A survey of modern authorship attribution methods.Journal of the American Society for Information Science and Technology, 60(3):538–556, 2009. doi: 10.1002/ asi.21001. URLhttps://doi.org/10.1002/asi.21001. [51]Efstathios Stamatatos. On the robustness of authorship attribution based on character n-gram features.Journal of Language and Politics, 12(3):421–439, 2013. URLhttps://icsdweb.aegean. gr/stamatatos/papers/JLP2013.pdf. Includes the original Guardian corpus statistics used in later topic-bias studies. [52]William J. Teahan and David J. Harper. Using compression-based language models for text categorization. In W. Bruce Croft and John Lafferty (eds.),Language Modeling for Information Retrieval, volume 13 ofThe Information Retrieval Series, p. 141–165. Springer, Dordrecht, 2003. doi: 10.1007/978-94-017-0171-6_7. URLhttps://link.springer.com/chapter/10.1007/ 978-94-017-0171-6_7 . [53]Jacob Tyo, Bhuwan Dhingra, and Zachary C Lipton. On the state of the art in authorship attribution and authorship verification.arXiv preprint arXiv:2209.06869, 2022. [54]Anqi Wang and Mohit Iyyer. Can authorship representation learning capture stylistic similar- ity?Transactions of the Association for Computational Linguistics, 11:1431–1450, 2023. doi: 10.1162/tacl\_a\_00610. URLhttps://doi.org/10.1162/tacl_a_00610. [55]Zichen Wen, Dadi Guo, and Huishuai Zhang. AIDBench: A benchmark for evaluating the authorship identification capability of large language models.arXiv preprint, arXiv:2411.13226, 2024. URLhttps://arxiv.org/abs/2411.13226. [56]Wikimedia Dumps. Wikisource dump index.https://dumps.wikimedia.org/enwikisource/ latest/ . Representative dump index for Wikisource; accessed: 2026-03-15. [57]An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025. 15 [58]yuanchunhong.Xiaohongshu aigc comments (including posts).Kag- gle Dataset, n.d.URLhttps://w.kaggle.com/datasets/yuanchunhong/ xiaohongshu-aigc-comments-including-postsdataset. Accessed: 2025-12-22. [59]Richong Zhang, Zhiyuan Hu, Hongyu Guo, and Yongyi Mao. Syntax encoding with applica- tion in authorship attribution. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, p. 2742–2753, Brussels, Belgium, 2018. Association for Compu- tational Linguistics. doi: 10.18653/v1/D18-1294. URLhttps://aclanthology.org/D18-1294/. [60]Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, et al. Qwen3 embedding: Advancing text embedding and reranking through foundation models.arXiv preprint arXiv:2506.05176, 2025. 16 A AuthBench Construction Details This appendix provides the full construction specification for AuthBench, including Build & Normal- ization, Quality Filtering, Redundancy Reduction, Language Audit, and Bucket Balanced Sampling. Section3.1in the main paper contains a concise summary. AuthBench is constructed through a five-stage pipeline: Build & Normalization, Quality Fil- tering, Redundancy Reduction, Language Audit, and Bucket Balanced Sampling. The pipeline is designed for large-scale ingestion while applying final benchmark size control only after Quality Filtering and Redundancy Reduction. A.1 Stage 1: Build & Normalization - Source Ingestion and Normalization Let S = S 1 , . . . , S m denote the set of raw sources. Each source yields records with source-specific metadata, but the pipeline converts every surviving item into a unified document representation x= (r, a, ℓ, g, σ, t, L(t), k(t),m), whereris a source-level raw identifier,ais an anonymized author identifier,ℓis language,gis the normalized genre,σis source,tis document text,L(t)is token length,k(t)is the length bucket, andmcontains optional metadata. Author identifiers are hashed deterministically as a=SHA256(σ:raw_author). Token length is computed with a fixed tokenizerT(·)(here,cl100k_base): L(t) =|T(t)|. Length buckets are then assigned by k(t) = 8 > > > > < > > > > : short1≤L(t)≤10, medium11≤L(t)≤100, long101≤L(t)≤500, extra_longL(t)>500. During ingestion, the builder may apply a bounded buffer shuffle to each dataset stream, which approximates random ordering without loading the full source into memory. A.2 Stage 1: Build & Normalization - Chunking and Author Qualification The first stage prepares a large author-qualified pool before any final size cap is enforced. If a raw document exceeds the chunking threshold, the pipeline segments it using paragraph and punctuation boundaries, with token-aware fallback splitting for very long sentences. Given a documentdwith textt, chunking produces C(d) =t 1 , . . . , t n ,L(t i )≤L max , where the pipeline uses configurable chunking parameters(L min , L target , L max ). Optional truncation is then applied by keeping the longest prefix that respects a specified token cap while preferring punctuation-aware segment boundaries. After chunking, the builder applies a first-pass dirty-text filter (described below) and stores surviving documents in a bounded per-author reservoir. LetD a be the documents observed for 17 authora. Stage 1 keeps at mostMdocuments per author in memory via reservoir replacement, with defaultM= 5. At finalization, authors with|D a |< mare normally discarded, where the default target ism= 3and a fallback minimum of2is allowed to recover sparse authors. Thus, the builder preferentially carries forward authors satisfying m≤|D a |≤M, with fallback admission for authors having exactly two clean documents. Stage 1 is intentionally permissive with respect to overall corpus size: it preserves as many clean, author-qualified items as possible for the later Quality Filtering and Bucket Balanced Sampling stages. A.3 Stage 2: Quality Filtering The second stage re-reads the author-qualified pool and performs stricter post-processing. First, the pipeline collapses letter-by-letter spacing artifacts such as “h e l l o” into normal words when the proportion or run length of single-letter alphabetic tokens is too high. Letw=T ws (t)denote whitespace-delimited tokens. The implementation measures r single (t) = 1 |w| |w| X j=1 1[|w j |= 1∧α(w j )], m single (t) =maxlength of a consecutive run of single-letter alphabetic tokens inw. Ifr single (t)orm single (t)exceeds a threshold, the pipeline attempts spacing collapse; if the cleaned text still exceeds the threshold, the document is dropped. The post-filter then re-tokenizes the cleaned text and reruns rule-based dirty filtering. Let w=T(t)be the token sequence. The dirty-text heuristics include: •Unique token ratio r uniq (t) = |unique(w)| |w| , and the document is dropped ifr uniq (t)< τ uniq . •Symbol ratio r sym (t) = #symbols int max(|t| chars ,1) , where public-domain sources use an additional consecutive-symbol check. •Maximum repeated-character runm rep (t), the longest run of the same non-space char- acter, which is dropped whenm rep (t)> K rep . •Maximum consecutive-symbol runm sym (t)for punctuation-heavy public-domain sources, dropped whenm sym (t)> K sym . The pipeline also applies an “untranslatable” filter that removes low-information or script- mismatched text. Define the alphabetic character ratio r α (t) = #alphabetic characters int #non-space characters int , 18 r tok (t) = 1 |w| | w | X j=1 1[α(w j )], whereα(w j )indicates that tokenw j contains at least one alphabetic character. LetS ℓ denote the expected Unicode script set for languageℓ. The script-match ratio is r script (t;ℓ) = #letters intwhose script∈S ℓ #letters int . The post-filter rejects texts with lowr α (t), lowr tok (t), excessive single-letter density, or low r script (t;ℓ)once enough alphabetic characters are present. When script evidence is weak, the filter can also consultlangdetect; if the detected language is incompatible with the target label, the document is removed at this stage. A.4 Stage 3: Redundancy Reduction After Quality Filtering, the pipeline removes redundancy in three passes. Exact normalized-text duplicates.For each document, let e t=N(t)denote case-folded text with canonicalized whitespace. We compute a stable exact-text key h exact (t) =BLAKE2b 128 (N(t)), and keep only the first document for each unique key. Near-text duplicates.For documents with at least a minimum token count, we compute a 64- bit SimHash over unigram and bigram features. LetF(t)be the multiset of unigram and bigram features extracted fromN(t). The near-text signature is h near (t) =SimHash 64 (F(t)). Candidate pairs are generated by LSH banding on the 64-bit signature. Two documentstandt ′ are considered near duplicates when Ham h near (t), h near (t ′ ) ≤⌊(1−θ near )·64⌋, whereθ near is the configured similarity threshold. By default, near-text comparisons are restricted to the same language. Near-author duplicates.The pipeline also supports optional author-profile redundancy checks to reduce residual cross-source aliasing. For authora, letD (p) a be the firstprepresentative docu- ments after sorting by document strength, wherepis a small constant. The profile text is formed by concatenation t ⋆ a = M d∈D (p) a N(t d ), and the profile signature ish author (a) =SimHash 64 (t ⋆ a ). Two authorsaanda ′ are treated as near duplicates when Ham h author (a), h author (a ′ ) ≤⌊(1−θ author )·64⌋, subject by default to same-language and cross-source constraints. When a conflict is found, the pipeline keeps the stronger author profile, prioritizing larger document count and then larger total token mass. 19 A.5 Stage 4: Language Audit After Redundancy Reduction, the pipeline runs an automated language audit. For each document, it recomputes the script-match ratior script (t;ℓ)and marks a document as suspicious when the ratio is too low for its declared language. In addition, the pipeline samples up to a configurable number of documents for probabilistic language identification usinglangdetect. Let b ℓ(t)be the top detected language andp( b ℓ(t)|t)its confidence. A document is flagged as a high-confidence mismatch when b ℓ(t)̸=ℓandp( b ℓ(t)|t)≥τ conf . By default, such documents are retagged rather than dropped; optional stricter settings can discard them. The audit records suspicious cases for targeted manual review together with summary counts such as mismatch rate, low-script rate, and retagged language pairs. A.6 Stage 5: Bucket Balanced Sampling LetD ⋆ denote the audited document pool. The final benchmark targetQis applied only at this stage. The pipeline first groups documents by language and assigns a target Q ℓ =round(Q p ℓ ), wherep ℓ is the configured language prior after renormalization over the languages that are actually available. Within languageℓ, genre targets are computed as Q ℓ,g =round(Q ℓ ρ ℓ,g ), and within each language–genre slice the length-bucket targets are Q ℓ,g,k =round(Q ℓ,g β k ), whereρ ℓ,g is the configured genre weight for languageℓandβ k is the global length-bucket prior. Sampling is performed without replacement. If a bucket(ℓ, g, k)is underfull, the pipeline first redistributes the deficit to leftover documents from the same language and genre, and then uses a spill pool formed from remaining language-matched documents. This preserves the language target as closely as possible while relaxing finer-grained quotas only when necessary. Split construction.Stage 5 also performs the final document-level stratified split. Documents are first grouped by language and then bucketed by b(d) = (g(d), k(d)). For each language-specific bucketB ℓ,g,k , the pipeline applies deterministic shuffling with a fixed seed and allocates integer split counts according to the requested train/dev/test ratios. Ifr s is the desired ratio for splits∈train,dev,test, the raw allocation is|B ℓ,g,k |r s , floored to integers, and any leftover documents are assigned by largest fractional remainder. This preserves genre and length composition within each language while keeping the split procedure deterministic. Because exact and near-text redundancy reduction are applied globally before splitting, dupli- cate leakage across train/dev/test is reduced by construction. The current benchmark doesnot enforce author-disjoint splits; instead, it prioritizes balanced document distributions for downstream retrieval and verification within each split. Within each split, authors with at least two documents contribute1–2candidate documents and1–2query documents, and each query is paired with all candidate documents from the same author in the benchmark ground-truth records. Overall, the pipeline yields a diverse benchmark with explicit control over language, genre, and length, while combining stream-safe ingestion, rule-based cleaning, multi-level redundancy reduction, automated language auditing, and deterministic stratified finalization. 20 A.7 Evaluation Protocol For retrieval, each test document is used as a query and ranked against a candidate pool drawn from the same split. For verification, we evaluate on labeled query–candidate pairs derived from the same within-split pools, computing both EER and ROC-AUC. By default, every non-matching can- didate in the pool is treated as a negative for a given query; the toolkit also supports deterministic negative sampling for faster ablations. All results are computed without task-specific fine-tuning to ensure comparability across model families. We report both aggregate performance and manda- tory stratified breakdowns by language, genre, and length. Length buckets follow the benchmark definition: short (1–10 tokens), medium (11–100), long (101–500), and extra_long (>500). Embedding extraction for LLM-based models.For base and instruction-tuned LLMs that do not expose a dedicated sentence-embedding head, we derive a single document vector from the final hidden states. Given an input documentx= (x 1 , . . . , x n ), let H (L) (x) = h h (L) 1 , . . . , h (L) n i ,h (L) i ∈R d , denote the last-layer token representations over the non-padding positions. Our default pooled representation is the masked mean ̄ h ( x ) = 1 P n i=1 m i n X i=1 m i h (L) i , wherem i ∈0,1indicates whether tokenx i is a valid (non-padding) token. We then L2-normalize the pooled vector, e(x) = ̄ h(x) ∥ ̄ h(x)∥ 2 , and usee(x)as the document embedding for both retrieval and verification. Similarity between a queryqand candidatecis then computed as cosine similarity, which reduces to a dot product after normalization: s(q, c) =e(q) ⊤ e(c). This is the default path used to obtain embeddings from generic LLM backbones in our unified pipeline. When a model family provides its own native embedding interface or explicitly rec- ommended pooling rule (e.g., special-token or last-token pooling), we follow that model-specific implementation instead. A.8 Metrics This metric design follows standard authorship-analysis practice while making our retrieval for- mulation explicit: classical authorship attribution is often closed-set classification [32,50], recent representation-learning work frequently evaluates open-world same-author retrieval with ranking metrics [44,54], and authorship verification remains the pairwise same-author decision problem typically assessed with ROC-based measures [ 34,35]. Letqbe a query with candidate setC q and positivesP q ⊂C q . Letπ q be the ranking of candidates by similarity. We define the binary relevance at rankias rel i = ( 1ifπ q (i)∈P q , 0otherwise. For a cutoffK(we reportK= 5in the main paper), we compute: 21 Recall@K(q) = 1 |P q | K X i=1 rel i DCG@K(q) = K X i=1 rel i log 2 (i+ 1) IDCG@K(q) = min(K,|P q |) X i=1 1 log 2 (i+ 1) , nDCG@K(q) = DCG@K(q) IDCG@K(q) R(q) = ( 1 mini:rel i =1 if there exists a relevant item inC q , 0otherwise. We define Success@K as: Success@K(q) =1 " K X i=1 rel i >0 # . We reportS@5,R@5, nDCG@5, and the macro-average mean reciprocal rank (MRR) over queries. For authorship verification, each query–candidate pair is assigned a similarity scores∈R and classified using a thresholdτ. LetPandNdenote the sets of positive and negative pairs, respectively. The false acceptance rate (FAR) and false rejection rate (FRR) at thresholdτare defined as FAR(τ) = 1 |N| X (q,c)∈N 1[s(q, c)≥τ], (1) FRR(τ) = 1 |P| X (q,c)∈P 1[s(q, c)< τ].(2) The Equal Error Rate (EER) is defined as the operating point where the two error rates are equal: EER=FAR(τ ∗ ) =FRR(τ ∗ ), whereτ ∗ =argmin τ |FAR(τ)−FRR(τ)|. The ROC curve traces TPR(τ) = 1−FRR(τ)against FAR(τ)asτvaries. We summarize this curve with the area under the ROC curve (ROC-AUC): ROC-AUC= 1 |P||N| X (q,c + )∈P X (q ′ ,c − )∈N 1 s(q, c + )> s(q ′ , c − ) + 1 2 1 s(q, c + ) =s(q ′ , c − ) . Higher ROC-AUC is better. 22 Why these metrics?Authorship attribution in AuthBench is an open-world retrieval problem with potentially multiple relevant candidates per query. Success@5 captures whether a model can surface at least one correct same-author document in a short analyst-facing shortlist. Recall@5 measures how much of the relevant same-author set is recovered within the top of the ranking. nDCG@5 complements these hit-based measures by rewarding correct documents that appear ear- lier and by accounting for queries with more than one relevant candidate. MRR adds a first-hit view of ranking quality, which is useful when only the first correct same-author retrieval matters operationally. Authorship verification is a pairwise same-author decision task. EER is useful when a single bal- anced operating point is desired because it directly summarizes the trade-off between false accepts and false rejects. ROC-AUC complements EER by measuring score separability independent of any one threshold, which is important when raw similarity scales differ across models. We therefore use EER for an interpretable operating-point summary and ROC-AUC for threshold-independent discrimination, consistent with recent authorship verification practice [ 34,35,54]. A.9 Models Evaluated Table 7: Models and baselines evaluated in AuthBench. We list every sys- tem appearing in the updated leaderboard together with its model size and Hugging Face repository or baseline implementation note. ModelModel Size Repository / Implementation LLMs (instruction-tuned) llama3-8b-instruct8Bmeta-llama/Meta-Llama-3-8B-Instruct llama3.1-8b-instruct8Bmeta-llama/Llama-3.1-8B-Instruct qwen2.5-3b-instruct3.1BQwen/Qwen2.5-3B-Instruct qwen2.5-7b-instruct7.6BQwen/Qwen2.5-7B-Instruct deepseek-llm-7b-chat7Bdeepseek-ai/deepseek-llm-7b-chat qwen3-4b-instruct4BQwen/Qwen3-4B-Instruct-2507 LLMs (base) llama3-8b8Bmeta-llama/Meta-Llama-3-8B llama3.1-8b8Bmeta-llama/Llama-3.1-8B deepseek-llm-7b-base7Bdeepseek-ai/deepseek-llm-7b-base qwen2.5-3b3.1BQwen/Qwen2.5-3B qwen3-4b4BQwen/Qwen3-4B Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1Bintfloat/e5-mistral-7b-instruct gte-qwen2-7b-instruct7.6BAlibaba-NLP/gte-Qwen2-7B-instruct Embedding models multilingual-e5-large559.9M intfloat/multilingual-e5-large multilingual-e5-base278Mintfloat/multilingual-e5-base qwen3-embedding-8b7.6BQwen/Qwen3-Embedding-8B sfr-embedding-mistral7.1BSalesforce/SFR-Embedding-Mistral qwen3-embedding-4b4BQwen/Qwen3-Embedding-4B snowflake-arctic-embed-l-v2567.8M Snowflake/snowflake-arctic-embed-l-v2.0 qwen3-embedding-0.6b595.8M Qwen/Qwen3-Embedding-0.6B e5-large-v2335.1M intfloat/e5-large-v2 e5-base-v2109Mintfloat/e5-base-v2 facebook-contriever110Mfacebook/contriever gte-large-en-v1.5409MAlibaba-NLP/gte-large-en-v1.5 bge-m3567MBAAI/bge-m3 bge-large-en-v1.5335MBAAI/bge-large-en-v1.5 e5-small-v233Mintfloat/e5-small-v2 Continued on next page 23 Table 7: Models and baselines evaluated in AuthBench (continued) ModelModel Size Repository / Implementation gte-large335.1M thenlper/gte-large mxbai-embed-large-v1335.1M mixedbread-ai/mxbai-embed-large-v1 bge-base-en-v1.5109.5M BAAI/bge-base-en-v1.5 gte-base109.5M thenlper/gte-base bge-base-zh-v1.5102MBAAI/bge-base-zh-v1.5 facebook-contriever-msmarco110Mfacebook/contriever-msmarco bge-large-zh-v1.5326MBAAI/bge-large-zh-v1.5 distiluse-base-multilingual-cased-v2134.7M sentence-transformers/distiluse-base-multilingual-cased-v2 all-roberta-large-v1355.4M sentence-transformers/all-roberta-large-v1 all-mpnet-base-v2109.5M sentence-transformers/all-mpnet-base-v2 bge-small-en-v1.533.4MBAAI/bge-small-en-v1.5 all-minilm-l12-v233.4Msentence-transformers/all-MiniLM-L12-v2 paraphrase-mpnet-base-v2109.5M sentence-transformers/paraphrase-mpnet-base-v2 bert-base-uncased110.1M bert-base-uncased all-minilm-l6-v222.7Msentence-transformers/all-MiniLM-L6-v2 paraphrase-multilingual-mpnet- base-v2 278Msentence-transformers/paraphrase-multilingual-mpnet-base-v2 msmarco-distilbert-base-v466.4Msentence-transformers/msmarco-distilbert-base-v4 allenai-specter110Mallenai/specter jina-embeddings-v2-small-en33Mjinaai/jina-embeddings-v2-small-en jina-embeddings-v2-base-en137.4M jinaai/jina-embeddings-v2-base-en Lexical / non-neural baselines tfidf–scikit-learn character 3–5 gram TF-IDF cosine baseline ngram–hashed character/word n-gram stylometric baseline with train-split cal- ibrator ppm–fixed-order hashed character language-model approximation of PPM- style scoring Non-neural baseline details.tfidfrepresents each document with scikit-learn character3–5 gram TF-IDF features and ranks candidates by cosine similarity, following the standard vector- space term-weighting formulation of Salton & Buckley [ 45].ngramis a lightweight stylometric baseline inspired by the feature-based authorship verification line of Koppel & Schler [22]: our implementation combines hashed character3–5grams, hashed word1–2grams, and a small set of surface cues (length, punctuation, digits, capitalization, whitespace, and mean token length), then fits a train-split linear pair calibrator to produce same-author scores.ppmfollows the compression- based language-modeling view of Teahan & Harper [ 52]: we approximate a fixed-order character PPM scorer with hashed character counts, derive symmetric query–candidate cross-entropy fea- tures, and fit a train-split linear calibrator. Forngramandppm, these are scalable benchmark implementations inspired by the cited methods rather than exact historical reimplementations. B Full Results Tables This appendix reports the full zero-shot result tables for all 47 neural models and the three non- neural baselines evaluated on AuthBench. Table8through Table14provide the complete bench- mark breakdown across overall, language, genre, and length settings, while Figures 4–9provide metric-wise visual summaries. B.1 Overall Leaderboard Full Results This subsection reports the complete zero-shot leaderboard across all evaluated model families on the AuthBench test split. 24 Table 8: Updated overall zero-shot results on AuthBench. Authorship attri- bution is evaluated with Success@5 (S@5), Recall@5 (R@5), nDCG@5, and MRR (higher is better). Authorship verification is evaluated with ROC- AUC (higher is better) and EER (lower is better). All 47 neural models and the three non-neural baselines are included. ModelModel Size S@5↑R@5↑nDCG@5↑MRR↑ROC-AUC↑EER↓ LLMs (instruction-tuned) llama3-8b-instruct8B0.251 0.2470.206 0.2090.9670.080 llama3.1-8b-instruct8B0.2470.2430.2040.2080.968 0.076 qwen2.5-3b-instruct3.1B0.213 0.2090.172 0.1760.962 0.091 qwen2.5-7b-instruct7.6B0.207 0.2030.168 0.1720.964 0.083 deepseek-llm-7b-chat7B0.207 0.2040.170 0.1740.940 0.128 qwen3-4b-instruct4B0.197 0.1930.159 0.1630.958 0.094 LLMs (base) llama3-8b8B0.254 0.2500.210 0.2130.967 0.079 llama3.1-8b8B0.2530.2490.2090.2120.9670.080 deepseek-llm-7b-base7B0.220 0.2160.180 0.1840.954 0.105 qwen2.5-3b3.1B0.212 0.2080.172 0.1760.962 0.090 qwen3-4b4B0.210 0.2060.171 0.1750.960 0.087 Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.242 0.2380.202 0.2050.9550.096 gte-qwen2-7b-instruct7.6B0.2400.2350.1980.2020.961 0.078 Embedding models multilingual-e5-large559.9M0.258 0.2540.217 0.2200.920 0.157 multilingual-e5-base278M0.2500.2450.2090.2120.918 0.161 qwen3-embedding-8b7.6B0.241 0.2370.201 0.2030.940 0.135 sfr-embedding-mistral7.1B0.240 0.2360.201 0.2050.956 0.096 qwen3-embedding-4b4B0.236 0.2310.196 0.1980.953 0.111 snowflake-arctic-embed-l-v2567.8M0.212 0.2080.182 0.1850.864 0.219 qwen3-embedding-0.6b595.8M0.211 0.2080.177 0.1810.953 0.108 e5-large-v2335.1M0.208 0.2050.173 0.1760.887 0.187 e5-base-v2109M0.200 0.1970.165 0.1690.888 0.179 facebook-contriever110M0.195 0.1920.160 0.1640.895 0.169 gte-large-en-v1.5409M0.189 0.1870.156 0.1590.923 0.129 bge-m3567M0.188 0.1850.161 0.1640.796 0.281 bge-large-en-v1.5335M0.177 0.1750.147 0.1500.867 0.192 e5-small-v233M0.177 0.1740.146 0.1500.856 0.216 gte-large335.1M0.175 0.1720.144 0.1480.889 0.163 mxbai-embed-large-v1335.1M0.174 0.1710.143 0.1460.865 0.197 bge-base-en-v1.5109.5M0.172 0.1690.142 0.1450.850 0.211 gte-base109.5M0.168 0.1650.138 0.1410.877 0.178 bge-base-zh-v1.5102M0.167 0.1640.139 0.1430.912 0.152 facebook-contriever-msmarco110M0.167 0.1640.138 0.1420.870 0.200 bge-large-zh-v1.5326M0.163 0.1590.135 0.1390.903 0.169 distiluse-base-multilingual-cased -v2 134.7M0.162 0.1590.134 0.1370.836 0.245 all-roberta-large-v1355.4M0.161 0.1580.132 0.1360.908 0.144 all-mpnet-base-v2109.5M0.160 0.1570.130 0.1330.889 0.169 bge-small-en-v1.533.4M0.158 0.1550.131 0.1350.853 0.214 all-minilm-l12-v233.4M0.152 0.1500.125 0.1290.888 0.185 paraphrase-mpnet-base-v2109.5M0.149 0.1460.123 0.1270.881 0.181 bert-base-uncased110.1M0.145 0.1430.118 0.1220.944 0.101 all-minilm-l6-v222.7M0.145 0.1430.120 0.1240.891 0.176 paraphrase-multilingual-mpnet-bas e-v2 278M0.138 0.1360.117 0.1190.794 0.281 msmarco-distilbert-base-v466.4M0.133 0.1300.110 0.1140.838 0.247 allenai-specter110M0.107 0.1050.086 0.0940.901 0.169 jina-embeddings-v2-small-en33M0.100 0.0980.081 0.0860.935 0.121 Continued on next page 25 0.050.100.150.200.25 success@5 (raw) multilingual-e5-large llama3-8b llama3.1-8b llama3-8b-instruct multilingual-e5-base llama3.1-8b-instruct e5-mistral-7b-instruct qwen3-embedding-8b sfr-embedding-mistral gte-qwen2-7b-instruct qwen3-embedding-4b deepseek-llm-7b-base qwen2.5-3b-instruct snowflake-arctic-embed-l-v2 qwen2.5-3b qwen3-embedding-0.6b qwen3-4b e5-large-v2 qwen2.5-7b-instruct deepseek-llm-7b-chat e5-base-v2 qwen3-4b-instruct facebook-contriever gte-large-en-v1.5 bge-m3 e5-small-v2 bge-large-en-v1.5 gte-large mxbai-embed-large-v1 bge-base-en-v1.5 gte-base bge-base-zh-v1.5 facebook-contriever-msmarco bge-large-zh-v1.5 distiluse-base-multilingual-cased-v2 all-roberta-large-v1 all-mpnet-base-v2 bge-small-en-v1.5 all-minilm-l12-v2 paraphrase-mpnet-base-v2 bert-base-uncased all-minilm-l6-v2 paraphrase-multilingual-mpnet-base-v2 msmarco-distilbert-base-v4 allenai-specter jina-embeddings-v2-small-en jina-embeddings-v2-base-en 0.258 0.254 0.253 0.251 0.250 0.247 0.242 0.241 0.240 0.240 0.236 0.220 0.213 0.212 0.212 0.211 0.210 0.208 0.207 0.207 0.200 0.197 0.195 0.189 0.188 0.177 0.177 0.175 0.174 0.172 0.168 0.167 0.167 0.163 0.162 0.161 0.160 0.158 0.152 0.149 0.145 0.145 0.138 0.133 0.107 0.100 0.016 Representation overall (success@5, raw) embedding llm-base llm-instruct embedding-instruct ngram: 0.170 ppm: 0.157 tfidf: 0.197 Figure 4: Overall AuthBench performance by Success@5 (S@5). Table 8: Updated overall zero-shot results on AuthBench (continued) ModelModel Size S@5↑R@5↑nDCG@5↑MRR↑ROC-AUC↑EER↓ jina-embeddings-v2-base-en137.4M0.016 0.0160.012 0.0140.738 0.328 Lexical / non-neural baselines tfidf–0.197 0.1940.167 0.1830.8370.219 ngram–0.1700.1670.1450.1490.874 0.213 ppm–0.157 0.1560.137 0.1420.792 0.298 B.2 Overall Metric Bar Charts These figures provide metric-wise visual summaries of the full leaderboard and make the relative spread between model families easier to compare at a glance. B.3 Language-wise Full Results The following tables provide the complete language-level breakdown for both retrieval and verification- oriented evaluation. 26 0.050.100.150.200.25 recall@5 (raw) multilingual-e5-large llama3-8b llama3.1-8b llama3-8b-instruct multilingual-e5-base llama3.1-8b-instruct e5-mistral-7b-instruct qwen3-embedding-8b sfr-embedding-mistral gte-qwen2-7b-instruct qwen3-embedding-4b deepseek-llm-7b-base qwen2.5-3b-instruct snowflake-arctic-embed-l-v2 qwen2.5-3b qwen3-embedding-0.6b qwen3-4b e5-large-v2 deepseek-llm-7b-chat qwen2.5-7b-instruct e5-base-v2 qwen3-4b-instruct facebook-contriever gte-large-en-v1.5 bge-m3 bge-large-en-v1.5 e5-small-v2 gte-large mxbai-embed-large-v1 bge-base-en-v1.5 gte-base bge-base-zh-v1.5 facebook-contriever-msmarco bge-large-zh-v1.5 distiluse-base-multilingual-cased-v2 all-roberta-large-v1 all-mpnet-base-v2 bge-small-en-v1.5 all-minilm-l12-v2 paraphrase-mpnet-base-v2 bert-base-uncased all-minilm-l6-v2 paraphrase-multilingual-mpnet-base-v2 msmarco-distilbert-base-v4 allenai-specter jina-embeddings-v2-small-en jina-embeddings-v2-base-en 0.254 0.250 0.249 0.247 0.245 0.243 0.238 0.237 0.236 0.235 0.231 0.216 0.209 0.208 0.208 0.208 0.206 0.205 0.204 0.203 0.197 0.193 0.192 0.187 0.185 0.175 0.174 0.172 0.171 0.169 0.165 0.164 0.164 0.159 0.159 0.158 0.157 0.155 0.150 0.146 0.143 0.143 0.136 0.130 0.105 0.098 0.016 Representation overall (recall@5, raw) embedding llm-base llm-instruct embedding-instruct ngram: 0.167 ppm: 0.156 tfidf: 0.194 Figure 5: Overall AuthBench performance by Recall@5 (R@5). Table 9: Language-wise Success@5 on AuthBench (full results). ModelModel Size ar de en es fr hi ja ko ru zh LLMs (instruction-tuned) deepseek-llm-7b-chat7B0.170 0.203 0.223 0.223 0.213 0.204 0.184 0.135 0.126 0.347 llama3-8b-instruct8B0.2260.2440.261 0.265 0.2690.2890.2430.1730.1640.382 llama3.1-8b-instruct8B0.2110.2450.2590.2650.2650.2970.2290.177 0.1650.373 qwen2.5-3b-instruct3.1B0.207 0.206 0.206 0.215 0.220 0.218 0.223 0.145 0.150 0.336 qwen2.5-7b-instruct7.6B0.182 0.210 0.201 0.208 0.214 0.215 0.215 0.143 0.143 0.343 qwen3-4b-instruct4B0.187 0.195 0.187 0.203 0.198 0.229 0.212 0.143 0.143 0.304 LLMs (base) deepseek-llm-7b-base7B0.183 0.216 0.238 0.223 0.220 0.238 0.200 0.144 0.131 0.367 llama3-8b8B0.2210.2470.270 0.272 0.2700.2890.2420.1830.1650.380 llama3.1-8b8B0.225 0.2490.2700.2720.2660.297 0.246 0.186 0.1730.361 qwen2.5-3b3.1B0.206 0.209 0.209 0.210 0.208 0.215 0.221 0.147 0.148 0.333 qwen3-4b4B0.189 0.217 0.204 0.213 0.214 0.227 0.220 0.142 0.147 0.336 Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.223 0.241 0.237 0.254 0.250 0.2720.2230.1640.1640.396 gte-qwen2-7b-instruct7.6B0.220 0.2410.2320.2350.2390.2520.2490.1560.1710.396 Embedding models all-minilm-l12-v233.4M0.128 0.135 0.181 0.157 0.162 0.195 0.135 0.083 0.084 0.246 Continued on next page 27 Table 9: Language-wise Success@5 on AuthBench (continued) ModelModel Size ar de en es fr hi ja ko ru zh all-minilm-l6-v222.7M0.117 0.128 0.173 0.157 0.152 0.184 0.116 0.076 0.077 0.245 all-mpnet-base-v2109.5M 0.152 0.154 0.190 0.166 0.166 0.176 0.148 0.092 0.092 0.230 all-roberta-large-v1355.4M 0.159 0.180 0.195 0.195 0.181 0.198 0.155 0.091 0.094 0.170 allenai-specter110M0.092 0.139 0.133 0.126 0.133 0.133 0.055 0.080 0.066 0.101 bert-base-uncased110.1M 0.135 0.136 0.188 0.140 0.155 0.156 0.112 0.095 0.079 0.199 bge-base-en-v1.5109.5M 0.155 0.171 0.193 0.200 0.203 0.195 0.148 0.099 0.101 0.241 bge-base-zh-v1.5102M0.116 0.124 0.147 0.153 0.137 0.193 0.190 0.124 0.086 0.407 bge-large-en-v1.5335M0.153 0.176 0.198 0.205 0.207 0.184 0.157 0.105 0.114 0.249 bge-large-zh-v1.5326M0.113 0.127 0.144 0.147 0.131 0.176 0.175 0.116 0.085 0.401 bge-m3567M0.213 0.159 0.155 0.174 0.188 0.176 0.165 0.138 0.129 0.369 bge-small-en-v1.533.4M0.131 0.145 0.184 0.180 0.175 0.139 0.132 0.089 0.087 0.253 distiluse-base-multilingual -cased-v2 134.7M 0.178 0.148 0.160 0.148 0.142 0.193 0.139 0.109 0.111 0.281 e5-base-v2109M0.182 0.201 0.227 0.242 0.236 0.215 0.194 0.110 0.128 0.255 e5-large-v2335.1M 0.203 0.197 0.229 0.248 0.249 0.210 0.192 0.120 0.141 0.269 e5-small-v233M0.150 0.178 0.201 0.199 0.193 0.207 0.171 0.092 0.116 0.254 facebook-contriever110M0.172 0.1870.2610.228 0.236 0.246 0.158 0.116 0.117 0.200 facebook-contriever-msmarco110M0.152 0.151 0.208 0.197 0.178 0.201 0.130 0.099 0.094 0.224 gte-base109.5M 0.119 0.176 0.210 0.206 0.205 0.125 0.149 0.080 0.099 0.224 gte-large335.1M 0.113 0.182 0.216 0.219 0.217 0.139 0.149 0.084 0.113 0.232 gte-large-en-v1.5409M0.181 0.185 0.223 0.228 0.222 0.210 0.180 0.111 0.115 0.231 jina-embeddings-v2-base-en137.4M 0.050 0.008 0.005 0.005 0.006 0.065 0.029 0.013 0.013 0.027 jina-embeddings-v2-small-en33M0.082 0.077 0.108 0.101 0.081 0.119 0.097 0.081 0.057 0.188 msmarco-distilbert-base-v466.4M0.123 0.113 0.157 0.127 0.115 0.142 0.102 0.091 0.079 0.224 multilingual-e5-base278M0.2510.210 0.226 0.2670.2620.238 0.2530.1850.1880.413 multilingual-e5-large559.9M0.2520.221 0.2310.284 0.280 0.263 0.268 0.196 0.1900.422 mxbai-embed-large-v1335.1M 0.143 0.173 0.196 0.198 0.204 0.193 0.154 0.101 0.116 0.240 paraphrase-mpnet-base-v2109.5M 0.134 0.136 0.182 0.163 0.138 0.210 0.139 0.098 0.091 0.203 paraphrase-multilingual-mpn et-base-v2 278M0.151 0.111 0.124 0.127 0.123 0.147 0.123 0.091 0.102 0.269 qwen3-embedding-0.6b595.8M 0.208 0.192 0.188 0.207 0.200 0.221 0.219 0.147 0.139 0.399 qwen3-embedding-4b4B0.220 0.231 0.213 0.227 0.243 0.218 0.241 0.167 0.160 0.422 qwen3-embedding-8b7.6B0.2390.2420.209 0.239 0.241 0.244 0.251 0.169 0.1700.425 sfr-embedding-mistral7.1B0.221 0.236 0.2360.255 0.248 0.2580.214 0.162 0.162 0.397 snowflake-arctic-embed-l-v2567.8M 0.240 0.161 0.184 0.187 0.202 0.252 0.215 0.154 0.160 0.385 Lexical / non-neural baselines ngram–0.2380.137 0.126 0.162 0.1660.244 0.1600.1680.1320.267 ppm–0.229 0.1400.1490.1900.1720.2550.113 0.103 0.118 0.184 tfidf–0.254 0.1480.146 0.1900.1550.280 0.226 0.184 0.154 0.346 Table 10: Language-wise EER on AuthBench (full results). ModelModel Size ar de en es fr hi ja ko ru zh LLMs (instruction-tuned) deepseek-llm-7b-chat7B0.125 0.103 0.141 0.089 0.103 0.071 0.061 0.081 0.123 0.074 llama3-8b-instruct8B0.0730.0640.0990.064 0.0610.0510.042 0.054 0.1120.057 llama3.1-8b-instruct8B0.0730.062 0.0960.0660.0620.0450.0420.0550.1140.059 qwen2.5-3b-instruct3.1B0.0720.072 0.131 0.074 0.068 0.054 0.043 0.060 0.119 0.062 qwen2.5-7b-instruct7.6B0.075 0.070 0.118 0.068 0.064 0.048 0.044 0.059 0.1200.056 qwen3-4b-instruct4B0.080 0.071 0.131 0.084 0.078 0.057 0.046 0.063 0.121 0.061 LLMs (base) deepseek-llm-7b-base7B0.082 0.072 0.128 0.076 0.071 0.062 0.043 0.061 0.121 0.065 llama3-8b8B0.0740.062 0.0940.0640.0610.0510.041 0.0520.1100.056 llama3.1-8b8B0.0730.0620.0940.064 0.060 0.0450.0420.0530.1090.060 qwen2.5-3b3.1B0.074 0.072 0.129 0.074 0.069 0.057 0.043 0.060 0.119 0.060 qwen3-4b4B0.079 0.066 0.123 0.077 0.068 0.057 0.045 0.064 0.123 0.059 Continued on next page 28 Table 10: Language-wise EER on AuthBench (continued) ModelModel Size ar de en es fr hi ja ko ru zh Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.0720.0720.1130.0670.0640.0480.0430.0630.1180.061 gte-qwen2-7b-instruct7.6B0.0730.069 0.1070.0670.063 0.045 0.041 0.0610.1180.054 Embedding models all-minilm-l12-v233.4M0.085 0.207 0.219 0.107 0.123 0.062 0.120 0.105 0.122 0.102 all-minilm-l6-v222.7M0.088 0.156 0.217 0.099 0.109 0.054 0.097 0.0990.1140.103 all-mpnet-base-v2109.5M 0.081 0.132 0.226 0.096 0.101 0.051 0.095 0.087 0.115 0.107 all-roberta-large-v1355.4M 0.082 0.083 0.192 0.0720.082 0.042 0.082 0.065 0.117 0.096 allenai-specter110M0.188 0.147 0.208 0.167 0.168 0.210 0.181 0.170 0.132 0.164 bert-base-uncased110.1M 0.084 0.0740.151 0.078 0.076 0.054 0.070 0.070 0.124 0.106 bge-base-en-v1.5109.5M 0.082 0.140 0.309 0.144 0.135 0.049 0.090 0.074 0.129 0.129 bge-base-zh-v1.5102M0.103 0.131 0.190 0.116 0.152 0.096 0.063 0.070 0.124 0.172 bge-large-en-v1.5335M0.079 0.115 0.287 0.122 0.133 0.049 0.078 0.077 0.132 0.119 bge-large-zh-v1.5326M0.094 0.173 0.241 0.164 0.192 0.126 0.104 0.099 0.124 0.205 bge-m3567M0.265 0.273 0.286 0.302 0.265 0.320 0.275 0.280 0.330 0.216 bge-small-en-v1.533.4M0.099 0.130 0.250 0.159 0.128 0.068 0.094 0.116 0.144 0.140 distiluse-base-multilingual -cased-v2 134.7M 0.253 0.256 0.206 0.262 0.241 0.271 0.235 0.238 0.308 0.210 e5-base-v2109M0.077 0.118 0.242 0.116 0.122 0.071 0.097 0.082 0.132 0.107 e5-large-v2335.1M 0.099 0.159 0.244 0.148 0.136 0.068 0.139 0.092 0.169 0.139 e5-small-v233M0.100 0.189 0.278 0.185 0.177 0.092 0.149 0.086 0.167 0.130 facebook-contriever110M0.076 0.096 0.177 0.095 0.095 0.051 0.052 0.067 0.118 0.084 facebook-contriever-msmarco110M0.086 0.143 0.228 0.127 0.170 0.068 0.067 0.072 0.125 0.106 gte-base109.5M 0.079 0.093 0.264 0.097 0.092 0.042 0.068 0.074 0.127 0.113 gte-large335.1M 0.079 0.090 0.263 0.091 0.0860.0370.061 0.075 0.124 0.113 gte-large-en-v1.5409M0.081 0.081 0.216 0.073 0.068 0.0400.078 0.109 0.126 0.123 jina-embeddings-v2-base-en137.4M 0.201 0.345 0.376 0.349 0.352 0.178 0.289 0.350 0.287 0.252 jina-embeddings-v2-small-en33M0.081 0.142 0.164 0.136 0.144 0.057 0.078 0.071 0.119 0.094 msmarco-distilbert-base-v466.4M0.083 0.118 0.249 0.111 0.122 0.076 0.115 0.083 0.131 0.119 multilingual-e5-base278M0.140 0.143 0.201 0.170 0.137 0.213 0.096 0.098 0.185 0.078 multilingual-e5-large559.9M 0.133 0.136 0.202 0.147 0.124 0.190 0.089 0.106 0.176 0.073 mxbai-embed-large-v1335.1M 0.080 0.124 0.285 0.130 0.139 0.051 0.073 0.074 0.127 0.124 paraphrase-mpnet-base-v2109.5M 0.081 0.103 0.221 0.087 0.088 0.099 0.104 0.088 0.127 0.121 paraphrase-multilingual-mpn et-base-v2 278M0.225 0.274 0.298 0.301 0.278 0.278 0.249 0.257 0.296 0.249 qwen3-embedding-0.6b595.8M 0.115 0.103 0.148 0.111 0.072 0.127 0.049 0.065 0.123 0.087 qwen3-embedding-4b4B0.092 0.112 0.156 0.115 0.091 0.089 0.051 0.0630.122 0.092 qwen3-embedding-8b7.6B0.0660.135 0.174 0.148 0.114 0.051 0.045 0.0590.1140.074 sfr-embedding-mistral7.1B0.0710.073 0.112 0.067 0.0640.0480.0430.063 0.1170.061 snowflake-arctic-embed-l-v2567.8M 0.167 0.244 0.233 0.248 0.225 0.164 0.198 0.200 0.255 0.172 Lexical / non-neural baselines ngram–0.2200.1910.2130.1940.1940.1100.134 0.2570.2620.119 ppm–0.288 0.262 0.257 0.257 0.248 0.278 0.3470.344 0.359 0.323 tfidf–0.136 0.165 0.200 0.179 0.188 0.0820.347 0.3400.1450.354 B.4 Primary-genre Full Results These tables report the full genre-level breakdown, highlighting how strongly authorship perfor- mance depends on discourse type. Table 11: Primary-genre Success@5 on AuthBench (full results). ModelModel Size blog ecomm literature media news poetry qna research social LLMs (instruction-tuned) deepseek-llm-7b-chat7B0.196 0.1100.461 0.0530.243 0.556 0.5410.713 0.166 llama3-8b-instruct8B0.2130.1070.521 0.070 0.302 0.639 0.6500.7690.206 Continued on next page 29 Table 11: Primary-genre Success@5 on AuthBench (continued) ModelModel Size blog ecomm literature media news poetry qna research social llama3.1-8b-instruct8B0.220 0.1120.5180.035 0.2930.583 0.6180.7870.204 qwen2.5-3b-instruct3.1B0.171 0.1000.483 0.035 0.252 0.6390.5730.731 0.171 qwen2.5-7b-instruct7.6B0.162 0.0840.496 0.035 0.242 0.583 0.5670.719 0.167 qwen3-4b-instruct4B0.149 0.0800.462 0.053 0.227 0.611 0.5480.688 0.160 LLMs (base) deepseek-llm-7b-base7B0.215 0.1140.486 0.035 0.257 0.639 0.5480.719 0.176 llama3-8b8B0.2290.1190.5240.070 0.311 0.667 0.6310.8000.207 llama3.1-8b8B0.2320.1120.5320.0530.3050.6670.6310.8120.207 qwen2.5-3b3.1B0.178 0.0960.487 0.018 0.248 0.639 0.5730.744 0.170 qwen3-4b4B0.166 0.0840.485 0.035 0.241 0.639 0.5610.738 0.171 Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.202 0.1100.5140.0530.2620.5830.5800.7250.207 gte-qwen2-7b-instruct7.6B0.1910.0910.5380.0350.281 0.611 0.586 0.7810.197 Embedding models all-minilm-l12-v233.4M0.128 0.0660.265 0.018 0.138 0.222 0.3380.781 0.141 all-minilm-l6-v222.7M0.120 0.0550.241 0.018 0.129 0.278 0.3380.756 0.136 all-mpnet-base-v2109.5M 0.131 0.0820.301 0.018 0.144 0.250 0.4650.8380.144 all-roberta-large-v1355.4M 0.132 0.0780.285 0.018 0.156 0.250 0.4080.756 0.146 allenai-specter110M0.093 0.0590.265 0.018 0.097 0.222 0.2550.644 0.089 bert-base-uncased110.1M 0.154 0.0820.311 0.035 0.146 0.222 0.2800.569 0.124 bge-base-en-v1.5109.5M 0.124 0.0820.318 0.018 0.162 0.250 0.4650.8440.156 bge-base-zh-v1.5102M0.082 0.0680.310 0.070 0.173 0.250 0.3820.463 0.155 bge-large-en-v1.5335M0.143 0.0730.342 0.018 0.161 0.306 0.5220.838 0.161 bge-large-zh-v1.5326M0.082 0.0730.305 0.035 0.167 0.250 0.4080.431 0.151 bge-m3567M0.108 0.0660.419 0.018 0.146 0.389 0.4330.594 0.181 bge-small-en-v1.533.4M0.116 0.0750.273 0.018 0.140 0.194 0.4900.800 0.147 distiluse-base-multiling ual-cased-v2 134.7M 0.104 0.0800.326 0.000 0.149 0.472 0.4970.600 0.148 e5-base-v2109M0.151 0.0980.369 0.018 0.212 0.333 0.5920.800 0.176 e5-large-v2335.1M 0.155 0.0890.388 0.018 0.216 0.306 0.643 0.819 0.184 e5-small-v233M0.124 0.0980.338 0.035 0.182 0.306 0.5670.750 0.156 facebook-contriever110M0.193 0.1140.380 0.018 0.237 0.361 0.4330.744 0.159 facebook-contriever-msmarco110M0.154 0.0940.292 0.018 0.171 0.250 0.4390.738 0.147 gte-base109.5M 0.143 0.0840.320 0.018 0.167 0.194 0.5220.831 0.146 gte-large335.1M 0.144 0.0680.343 0.018 0.171 0.278 0.5290.831 0.154 gte-large-en-v1.5409M0.141 0.0820.351 0.018 0.208 0.278 0.5610.806 0.165 jina-embeddings-v2-base-en137.4M 0.005 0.0020.028 0.000 0.012 0.000 0.0060.006 0.018 jina-embeddings-v2-small-en33M0.063 0.0410.194 0.018 0.089 0.083 0.1270.338 0.096 msmarco-distilbert-base-v466.4M0.108 0.0570.225 0.035 0.135 0.167 0.2930.625 0.120 multilingual-e5-base278M0.146 0.107 0.453 0.053 0.2580.500 0.6310.756 0.230 multilingual-e5-large559.9M 0.160 0.1000.479 0.0530.2680.4440.6820.7690.236 mxbai-embed-large-v1335.1M 0.134 0.0640.345 0.018 0.158 0.167 0.5220.831 0.157 paraphrase-mpnet-base-v2109.5M 0.146 0.0940.274 0.018 0.140 0.278 0.4140.694 0.132 paraphrase-multilingual- mpnet-base-v2 278M0.081 0.0530.295 0.000 0.101 0.389 0.4330.644 0.132 qwen3-embedding-0.6b595.8M 0.124 0.0780.461 0.053 0.211 0.556 0.5290.787 0.188 qwen3-embedding-4b4B0.151 0.0840.5350.070 0.2430.5830.5610.831 0.205 qwen3-embedding-8b7.6B0.151 0.0910.536 0.0880.256 0.556 0.5860.838 0.209 sfr-embedding-mistral7.1B0.1990.1050.512 0.053 0.257 0.583 0.5670.756 0.207 snowflake-arctic-embed-l-v2567.8M 0.129 0.0570.418 0.035 0.178 0.472 0.5410.738 0.202 Lexical / non-neural baselines ngram–0.066 0.0500.2890.0180.1610.2780.3250.431 0.168 ppm–0.090 0.0640.230 0.000 0.149 0.250 0.4200.6310.153 tfidf–0.089 0.0530.323 0.035 0.169 0.444 0.4590.6130.198 30 0.050.100.150.20 ndcg@5 (raw) multilingual-e5-large llama3-8b multilingual-e5-base llama3.1-8b llama3-8b-instruct llama3.1-8b-instruct e5-mistral-7b-instruct sfr-embedding-mistral qwen3-embedding-8b gte-qwen2-7b-instruct qwen3-embedding-4b snowflake-arctic-embed-l-v2 deepseek-llm-7b-base qwen3-embedding-0.6b e5-large-v2 qwen2.5-3b-instruct qwen2.5-3b qwen3-4b deepseek-llm-7b-chat qwen2.5-7b-instruct e5-base-v2 bge-m3 facebook-contriever qwen3-4b-instruct gte-large-en-v1.5 bge-large-en-v1.5 e5-small-v2 gte-large mxbai-embed-large-v1 bge-base-en-v1.5 bge-base-zh-v1.5 facebook-contriever-msmarco gte-base bge-large-zh-v1.5 distiluse-base-multilingual-cased-v2 all-roberta-large-v1 bge-small-en-v1.5 all-mpnet-base-v2 all-minilm-l12-v2 paraphrase-mpnet-base-v2 all-minilm-l6-v2 bert-base-uncased paraphrase-multilingual-mpnet-base-v2 msmarco-distilbert-base-v4 allenai-specter jina-embeddings-v2-small-en jina-embeddings-v2-base-en 0.217 0.210 0.209 0.209 0.206 0.204 0.202 0.201 0.201 0.198 0.196 0.182 0.180 0.177 0.173 0.172 0.172 0.171 0.170 0.168 0.165 0.161 0.160 0.159 0.156 0.147 0.146 0.144 0.143 0.142 0.139 0.138 0.138 0.135 0.134 0.132 0.131 0.130 0.125 0.123 0.120 0.118 0.117 0.110 0.086 0.081 0.012 Representation overall (ndcg@5, raw) embedding llm-base llm-instruct embedding-instruct ngram: 0.145 ppm: 0.137 tfidf: 0.167 Figure 6: Overall AuthBench performance by nDCG@5. Table 12: Primary-genre EER on AuthBench (full results). ModelModel Size blog ecomm literature media news poetry qna research social LLMs (instruction-tuned) deepseek-llm-7b-chat7B0.133 0.0940.076 0.078 0.111 0.090 0.1100.029 0.132 llama3-8b-instruct8B0.1060.0740.0360.078 0.070 0.0640.063 0.0110.081 llama3.1-8b-instruct8B0.1080.0640.0370.0800.0650.068 0.0670.0120.081 qwen2.5-3b-instruct3.1B0.137 0.0860.048 0.079 0.0750.0590.0980.031 0.091 qwen2.5-7b-instruct7.6B0.131 0.0800.051 0.075 0.0700.0610.0870.019 0.085 qwen3-4b-instruct4B0.136 0.0960.0500.0730.081 0.080 0.0980.046 0.092 LLMs (base) deepseek-llm-7b-base7B0.135 0.0910.070 0.078 0.103 0.079 0.0990.031 0.105 llama3-8b8B0.105 0.0620.0360.073 0.0700.0680.0690.009 0.081 llama3.1-8b8B0.108 0.0630.0350.081 0.0710.0720.0670.0090.082 qwen2.5-3b3.1B0.134 0.0850.049 0.077 0.0740.0650.1010.031 0.090 qwen3-4b4B0.128 0.0840.046 0.074 0.074 0.078 0.0910.025 0.088 Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.1200.0730.0550.0850.0880.0780.0900.0250.093 gte-qwen2-7b-instruct7.6B0.113 0.0700.047 0.069 0.067 0.076 0.075 0.013 0.082 Embedding models all-minilm-l12-v233.4M0.214 0.2670.111 0.138 0.194 0.063 0.1260.025 0.167 Continued on next page 31 Table 12: Primary-genre EER on AuthBench (continued) ModelModel Size blog ecomm literature media news poetry qna research social all-minilm-l6-v222.7M0.211 0.2630.084 0.133 0.200 0.081 0.1090.025 0.146 all-mpnet-base-v2109.5M 0.219 0.2770.079 0.138 0.197 0.059 0.1210.019 0.138 all-roberta-large-v1355.4M 0.206 0.1050.076 0.108 0.1750.0390.1210.031 0.114 allenai-specter110M0.164 0.1720.070 0.192 0.164 0.078 0.0980.031 0.182 bert-base-uncased110.1M 0.1280.1260.069 0.137 0.1020.082 0.0820.029 0.101 bge-base-en-v1.5109.5M 0.230 0.3650.078 0.211 0.298 0.082 0.1360.031 0.175 bge-base-zh-v1.5102M0.157 0.1420.061 0.310 0.142 0.062 0.0800.056 0.152 bge-large-en-v1.5335M0.222 0.3220.080 0.183 0.259 0.098 0.1170.0130.151 bge-large-zh-v1.5326M0.211 0.2020.066 0.336 0.154 0.063 0.0860.102 0.162 bge-m3567M0.257 0.2140.076 0.368 0.265 0.118 0.1090.054 0.304 bge-small-en-v1.533.4M0.202 0.2770.099 0.229 0.264 0.092 0.1480.025 0.178 distiluse-base-multiling ual-cased-v2 134.7M 0.191 0.1450.091 0.391 0.191 0.122 0.0520.046 0.277 e5-base-v2109M0.215 0.1940.072 0.172 0.217 0.0560.0570.025 0.162 e5-large-v2335.1M 0.217 0.1800.078 0.207 0.221 0.078 0.0590.031 0.179 e5-small-v233M0.244 0.2030.095 0.230 0.264 0.063 0.0630.050 0.205 facebook-contriever110M0.228 0.0940.075 0.111 0.155 0.081 0.1090.0130.142 facebook-contriever-msmarco110M0.234 0.1480.090 0.127 0.211 0.078 0.0950.021 0.179 gte-base109.5M 0.210 0.3590.077 0.155 0.243 0.085 0.1090.031 0.138 gte-large335.1M 0.207 0.3580.066 0.153 0.213 0.084 0.0980.025 0.127 gte-large-en-v1.5409M0.193 0.5160.0420.149 0.122 0.059 0.0790.031 0.117 jina-embeddings-v2-base-en137.4M 0.326 0.3520.275 0.230 0.342 0.255 0.3370.294 0.325 jina-embeddings-v2-small-en33M0.159 0.1260.083 0.125 0.124 0.083 0.0940.082 0.122 msmarco-distilbert-base-v466.4M0.238 0.2050.117 0.146 0.241 0.089 0.2080.025 0.234 multilingual-e5-base278M0.179 0.1320.087 0.0900.145 0.1370.0270.025 0.151 multilingual-e5-large559.9M 0.197 0.1600.078 0.105 0.146 0.118 0.034 0.022 0.147 mxbai-embed-large-v1335.1M 0.242 0.3110.081 0.175 0.253 0.096 0.1160.013 0.156 paraphrase-mpnet-base-v2109.5M 0.218 0.1280.079 0.144 0.228 0.093 0.1140.019 0.142 paraphrase-multilingual- mpnet-base-v2 278M0.310 0.2120.122 0.368 0.287 0.176 0.0980.025 0.283 qwen3-embedding-0.6b595.8M 0.137 0.0960.044 0.166 0.103 0.066 0.0630.015 0.109 qwen3-embedding-4b4B0.131 0.0950.0400.174 0.110 0.061 0.0800.013 0.107 qwen3-embedding-8b7.6B0.166 0.0970.070 0.120 0.139 0.070 0.0920.025 0.126 sfr-embedding-mistral7.1B0.117 0.0740.0530.088 0.0890.088 0.0880.0250.093 snowflake-arctic-embed-l-v2567.8M 0.220 0.1580.065 0.324 0.210 0.078 0.1280.056 0.233 Lexical / non-neural baselines ngram–0.1950.2340.128 0.134 0.1460.1270.1330.0870.231 ppm–0.214 0.2400.275 0.3560.276 0.335 0.2070.0690.315 tfidf–0.154 0.1850.148 0.498 0.1990.098 0.0790.0750.243 B.5 Length-bucket Full Results These tables give the complete length-wise breakdown and make the sparse-text versus long- document gap explicit across all evaluated systems. Table 13: Length-bucket Success@5 on AuthBench (full results). ModelModel Size short medium long extra_long LLMs (instruction-tuned) deepseek-llm-7b-chat7B0.1060.161 0.3710.524 llama3-8b-instruct8B0.150 0.199 0.4320.539 llama3.1-8b-instruct8B0.142 0.1960.4270.531 qwen2.5-3b-instruct3.1B0.1010.164 0.3850.508 qwen2.5-7b-instruct7.6B0.1030.159 0.3790.496 qwen3-4b-instruct4B0.0810.150 0.3680.488 LLMs (base) deepseek-llm-7b-base7B0.1210.169 0.3930.528 Continued on next page 32 Table 13: Length-bucket Success@5 on AuthBench (continued) ModelModel Size short medium long extra_long llama3-8b8B0.150 0.2020.4360.524 llama3.1-8b8B0.135 0.2020.4390.543 qwen2.5-3b3.1B0.0930.163 0.3860.520 qwen3-4b4B0.1010.161 0.3850.496 Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.161 0.1960.3980.524 gte-qwen2-7b-instruct7.6B0.1370.1920.4090.500 Embedding models all-minilm-l12-v233.4M0.1280.129 0.2220.343 all-minilm-l6-v222.7M0.1260.122 0.2110.335 all-mpnet-base-v2109.5M0.1220.133 0.2460.358 all-roberta-large-v1355.4M0.1130.136 0.2480.280 allenai-specter110M0.0690.076 0.2060.299 bert-base-uncased110.1M0.0950.113 0.2500.354 bge-base-en-v1.5109.5M0.1170.145 0.2650.323 bge-base-zh-v1.5102M0.1660.132 0.2630.386 bge-large-en-v1.5335M0.1240.149 0.2740.346 bge-large-zh-v1.5326M0.1640.130 0.2510.370 bge-m3567M0.1510.152 0.3020.406 bge-small-en-v1.533.4M0.1250.135 0.2310.319 distiluse-base-multilingual-cased -v2 134.7M0.1140.130 0.2700.327 e5-base-v2109M0.1370.169 0.3070.390 e5-large-v2335.1M0.1400.176 0.3200.406 e5-small-v233M0.1230.146 0.2820.378 facebook-contriever110M0.1200.161 0.3170.402 facebook-contriever-msmarco110M0.1160.140 0.2570.366 gte-base109.5M0.1120.144 0.2550.323 gte-large335.1M0.1180.149 0.2620.370 gte-large-en-v1.5409M0.1080.162 0.2930.406 jina-embeddings-v2-base-en137.4M0.0110.013 0.0260.067 jina-embeddings-v2-small-en33M0.0880.078 0.1610.299 msmarco-distilbert-base-v466.4M0.1150.109 0.2040.291 multilingual-e5-base278M0.1870.213 0.3760.417 multilingual-e5-large559.9M0.1820.2200.3920.476 mxbai-embed-large-v1335.1M0.1160.147 0.2670.346 paraphrase-mpnet-base-v2109.5M0.1130.119 0.2440.335 paraphrase-multilingual-mpnet-bas e-v2 278M0.0970.111 0.2290.307 qwen3-embedding-0.6b595.8M0.1540.173 0.3370.453 qwen3-embedding-4b4B0.1670.191 0.3840.500 qwen3-embedding-8b7.6B0.1570.1970.3950.504 sfr-embedding-mistral7.1B0.1680.193 0.3950.520 snowflake-arctic-embed-l-v2567.8M0.1610.175 0.3290.445 Lexical / non-neural baselines ngram–0.1510.1360.2710.331 ppm–0.0790.134 0.2550.240 tfidf–0.157 0.164 0.3000.437 Table 14: Length-bucket EER on AuthBench (full results). ModelModel Size short medium long extra_long LLMs (instruction-tuned) deepseek-llm-7b-chat7B0.1850.131 0.0960.072 llama3-8b-instruct8B0.1090.0820.0600.055 llama3.1-8b-instruct8B0.104 0.0780.061 0.051 Continued on next page 33 Table 14: Length-bucket EER on AuthBench (continued) ModelModel Size short medium long extra_long qwen2.5-3b-instruct3.1B0.1200.090 0.0730.067 qwen2.5-7b-instruct7.6B0.1100.082 0.0690.051 qwen3-4b-instruct4B0.1530.094 0.0750.061 LLMs (base) deepseek-llm-7b-base7B0.1610.105 0.0840.084 llama3-8b8B0.108 0.0820.0640.054 llama3.1-8b8B0.1100.0820.0640.054 qwen2.5-3b3.1B0.1190.090 0.0720.063 qwen3-4b4B0.1280.086 0.0700.058 Embedding models (instruction-tuned) e5-mistral-7b-instruct7.1B0.1230.0960.0790.083 gte-qwen2-7b-instruct7.6B0.109 0.078 0.0660.054 Embedding models all-minilm-l12-v233.4M0.2460.197 0.1300.123 all-minilm-l6-v222.7M0.2480.186 0.1210.119 all-mpnet-base-v2109.5M0.2480.176 0.1200.144 all-roberta-large-v1355.4M0.2260.147 0.1080.148 allenai-specter110M0.2370.178 0.1100.094 bert-base-uncased110.1M0.1470.0980.0750.083 bge-base-en-v1.5109.5M0.3090.222 0.1300.151 bge-base-zh-v1.5102M0.2000.143 0.0960.089 bge-large-en-v1.5335M0.2890.204 0.1200.148 bge-large-zh-v1.5326M0.2030.147 0.1020.113 bge-m3567M0.3540.296 0.1800.155 bge-small-en-v1.533.4M0.2930.224 0.1520.176 distiluse-base-multilingual-cased -v2 134.7M0.2750.255 0.1580.092 e5-base-v2109M0.2600.186 0.1150.116 e5-large-v2335.1M0.2450.190 0.1310.137 e5-small-v233M0.2960.226 0.1390.116 facebook-contriever110M0.2590.168 0.1340.206 facebook-contriever-msmarco110M0.2750.202 0.1440.184 gte-base109.5M0.2740.182 0.1230.158 gte-large335.1M0.2510.170 0.1090.131 gte-large-en-v1.5409M0.2000.128 0.0950.116 jina-embeddings-v2-base-en137.4M0.3870.338 0.2830.239 jina-embeddings-v2-small-en33M0.1750.124 0.0870.081 msmarco-distilbert-base-v466.4M0.2700.255 0.1960.249 multilingual-e5-base278M0.1970.171 0.1200.083 multilingual-e5-large559.9M0.1990.164 0.1190.076 mxbai-embed-large-v1335.1M0.2920.209 0.1260.143 paraphrase-mpnet-base-v2109.5M0.2600.189 0.1240.170 paraphrase-multilingual-mpnet-bas e-v2 278M0.3600.297 0.2080.202 qwen3-embedding-0.6b595.8M0.1690.109 0.0820.048 qwen3-embedding-4b4B0.1780.111 0.0850.058 qwen3-embedding-8b7.6B0.1920.134 0.1170.094 sfr-embedding-mistral7.1B0.121 0.0960.0790.083 snowflake-arctic-embed-l-v2567.8M0.2730.235 0.1430.072 Lexical / non-neural baselines ngram–0.260 0.225 0.1540.124 ppm–0.3310.308 0.2540.274 tfidf–0.3400.2250.1590.202 34 0.050.100.150.20 mrr (raw) multilingual-e5-large llama3-8b llama3.1-8b multilingual-e5-base llama3-8b-instruct llama3.1-8b-instruct sfr-embedding-mistral e5-mistral-7b-instruct qwen3-embedding-8b gte-qwen2-7b-instruct qwen3-embedding-4b snowflake-arctic-embed-l-v2 deepseek-llm-7b-base qwen3-embedding-0.6b qwen2.5-3b-instruct qwen2.5-3b e5-large-v2 qwen3-4b deepseek-llm-7b-chat qwen2.5-7b-instruct e5-base-v2 bge-m3 facebook-contriever qwen3-4b-instruct gte-large-en-v1.5 e5-small-v2 bge-large-en-v1.5 gte-large mxbai-embed-large-v1 bge-base-en-v1.5 bge-base-zh-v1.5 facebook-contriever-msmarco gte-base bge-large-zh-v1.5 distiluse-base-multilingual-cased-v2 all-roberta-large-v1 bge-small-en-v1.5 all-mpnet-base-v2 all-minilm-l12-v2 paraphrase-mpnet-base-v2 all-minilm-l6-v2 bert-base-uncased paraphrase-multilingual-mpnet-base-v2 msmarco-distilbert-base-v4 allenai-specter jina-embeddings-v2-small-en jina-embeddings-v2-base-en 0.220 0.213 0.212 0.212 0.209 0.208 0.205 0.205 0.203 0.202 0.198 0.185 0.184 0.181 0.176 0.176 0.176 0.175 0.174 0.172 0.169 0.164 0.164 0.163 0.159 0.150 0.150 0.148 0.146 0.145 0.143 0.142 0.141 0.139 0.137 0.136 0.135 0.133 0.129 0.127 0.124 0.122 0.119 0.114 0.094 0.086 0.014 Representation overall (mrr, raw) embedding llm-base llm-instruct embedding-instruct ngram: 0.149 ppm: 0.142 tfidf: 0.183 Figure 7: Overall AuthBench performance by MRR. C Post-Training Note This appendix focuses on the zero-shot benchmark results reported in the current paper. In par- allel, we are conducting a broader post-training study covering multiple adaptation strategies and authorship methods, and we plan to present those findings in a second version of the paper. The zero-shot results here are intended to provide a clean foundation for that next stage of analysis. D Additional Dataset Statistics This appendix complements the full result tables with additional dataset-profile visualizations. Figures4–9summarize overall metric-level performance across models, Figure10shows token- length distributions by language, and Figure11shows the fine-grained subgenre mix within each language. E Raw Data Sources AuthBench draws on17publicly available input sources spanning multiple platforms, domains, and languages. Each raw item is mapped into a unified schema (Section3.1) with provenance 35 0.750.800.850.900.95 roc_auc (raw) llama3.1-8b-instruct llama3-8b llama3-8b-instruct llama3.1-8b qwen2.5-7b-instruct qwen2.5-3b qwen2.5-3b-instruct gte-qwen2-7b-instruct qwen3-4b qwen3-4b-instruct sfr-embedding-mistral e5-mistral-7b-instruct deepseek-llm-7b-base qwen3-embedding-0.6b qwen3-embedding-4b bert-base-uncased deepseek-llm-7b-chat qwen3-embedding-8b jina-embeddings-v2-small-en gte-large-en-v1.5 multilingual-e5-large multilingual-e5-base bge-base-zh-v1.5 all-roberta-large-v1 bge-large-zh-v1.5 allenai-specter facebook-contriever all-minilm-l6-v2 gte-large all-mpnet-base-v2 e5-base-v2 all-minilm-l12-v2 e5-large-v2 paraphrase-mpnet-base-v2 gte-base facebook-contriever-msmarco bge-large-en-v1.5 mxbai-embed-large-v1 snowflake-arctic-embed-l-v2 e5-small-v2 bge-small-en-v1.5 bge-base-en-v1.5 msmarco-distilbert-base-v4 distiluse-base-multilingual-cased-v2 bge-m3 paraphrase-multilingual-mpnet-base-v2 jina-embeddings-v2-base-en 0.968 0.967 0.967 0.967 0.964 0.962 0.962 0.961 0.960 0.958 0.956 0.955 0.954 0.953 0.953 0.944 0.940 0.940 0.935 0.923 0.920 0.918 0.912 0.908 0.903 0.901 0.895 0.891 0.889 0.889 0.888 0.888 0.887 0.881 0.877 0.870 0.867 0.865 0.864 0.856 0.853 0.850 0.838 0.836 0.796 0.794 0.738 Attribution overall (roc_auc, raw) llm-instruct llm-base embedding-instruct embedding ngram: 0.874 ppm: 0.792 tfidf: 0.837 Figure 8: Overall AuthBench performance by ROC-AUC. (source_id), language (lang), a normalized genre label (genre), and token lengthℓ(d)∈Ncom- puted under a fixed tokenizer. Table15lists these sources, and Table16reports their realized contribution to the current benchmark release. Citation policy.When a dataset has an associated peer-reviewed publication, we treat it as the canonical reference (e.g., MARC for Amazon Reviews Multi; Babel Briefings for multilingual head- lines). For mirrors or redistributed versions (e.g., Kaggle or Hugging Face hosting), we additionally cite the hosting page to ensure reproducibility via stable URLs and access dates. No configured source is entirely absent from the current combined filtered output; however, the realized source distribution is highly skewed, with Exorde, Wikisource, Babel Briefings, and YouTube comments together contributing 74.4% of all documents. E.1 Data Consent AuthBench is derived from publicly accessible datasets and public-web sources released or exposed under documented provider licenses or site terms (Appendix E). We did not collect new data directly from individuals, and we did not contact data subjects. Where consent mechanisms are relevant, we rely on the original dataset providersterms or the governing platform/site terms for public availability and redistribution. We further reduce privacy risk by anonymizing author identifiers 36 0.100.150.200.250.300.35 eer (raw) llama3.1-8b-instruct gte-qwen2-7b-instruct llama3-8b llama3-8b-instruct llama3.1-8b qwen2.5-7b-instruct qwen3-4b qwen2.5-3b qwen2.5-3b-instruct qwen3-4b-instruct sfr-embedding-mistral e5-mistral-7b-instruct bert-base-uncased deepseek-llm-7b-base qwen3-embedding-0.6b qwen3-embedding-4b jina-embeddings-v2-small-en deepseek-llm-7b-chat gte-large-en-v1.5 qwen3-embedding-8b all-roberta-large-v1 bge-base-zh-v1.5 multilingual-e5-large multilingual-e5-base gte-large bge-large-zh-v1.5 all-mpnet-base-v2 allenai-specter facebook-contriever all-minilm-l6-v2 gte-base e5-base-v2 paraphrase-mpnet-base-v2 all-minilm-l12-v2 e5-large-v2 bge-large-en-v1.5 mxbai-embed-large-v1 facebook-contriever-msmarco bge-base-en-v1.5 bge-small-en-v1.5 e5-small-v2 snowflake-arctic-embed-l-v2 distiluse-base-multilingual-cased-v2 msmarco-distilbert-base-v4 paraphrase-multilingual-mpnet-base-v2 bge-m3 jina-embeddings-v2-base-en 0.076 0.078 0.079 0.080 0.080 0.083 0.087 0.090 0.091 0.094 0.096 0.096 0.101 0.105 0.108 0.111 0.121 0.128 0.129 0.135 0.144 0.152 0.157 0.161 0.163 0.169 0.169 0.169 0.169 0.176 0.178 0.179 0.181 0.185 0.187 0.192 0.197 0.200 0.211 0.214 0.216 0.219 0.245 0.247 0.281 0.281 0.328 Attribution overall (eer, raw) llm-instruct embedding-instruct llm-base embedding ngram: 0.213 ppm: 0.298 tfidf: 0.219 Figure 9: Overall AuthBench performance by EER. (Section3; Section3.1) and applying conservative safety filtering (AppendixA.3). 37 enruzhardekoesfrjahi Language 0 100 200 300 400 Token length Token length distribution by language doc_type candidate query Figure 10: Token-length distribution per language. Each box plot summarizes the document- length distribution for one language in the current AuthBench release, highlighting cross-language differences in median length, spread, and long-tail behavior. 25% 20% 14% EN n=97,974 English 37% 19% 16% 16% ZH n=55,368 Chinese 27% 26% 17% HI n=6,825 Hindi 30% 21% 18% ES n=33,395 Spanish 33% 30% 16% FR n=31,225 French 21% 16% 13% AR n=42,091 Arabic 44% 15% RU n=66,084 Russian 44% 21% 12% DE n=39,813 German 45% 21% 16% JA n=21,494 Japanese 30% 28% 21% KO n=33,881 Korean Fine-grained subgenre distribution by language Shared colors indicate the same fine-grained subgenre across languages; low-frequency subgenres are grouped into Other. Literature News YouTube comment Entertainment Politics People Douban reviews Sports Xiaohongshu Technology Blog: indunk Poetry Environment Blog: student Speech / essay E-commerce reviews Q&A question Other Figure 11: Fine-grained subgenre distribution by language. Each donut chart summarizes one language slice using a shared subgenre palette, while lower-frequency subgenres are grouped into Otherto keep the figure legible. 38 SourceLang. Primary genre(s) ScaleAuthor labels Reference ExordeMulti Social / news / forums 65M+/weekYes (author hash) [12] Babel Briefings30+ News headlines 4.7MPartial (publisher / org) [24,25] Amazon Reviews Multi (MARC) 6E-commerce reviews 200k+/langYes (reviewer IDs) [19,29] Blog Authorship Corpus enBlogs681k postsYes[5,46] arXiv Abstracts (metadata snapshot) enResearch papers 1.7M+Yes (first author from metadata) [4,10] Xiaohongshu / Weibo zhSocial media11k+Yes (user IDs)[58] Douban ReviewszhMedia/book/music reviews 13.5MYes (reviewer IDs) [ 13] Hindi Discourse Stories hiLiterature (short stories) 53 storiesYes[11,30] Spanish PD BooksesLiterature300k+ textsYes (metadata) [ 39] French PD BooksfrLiterature289k+ booksYes (metadata) [36] Arabic Classical Poetry arPoetry70k poemsYes (poet)[28] Russian PD Corpus ruLiterature/periodicals8.5k titlesYes (metadata) [38] German PD Corpus deLiterature/newspapers260k+ textsYes (metadata) [37] Stack Exchange Data Dump / API crawl MultiQ&ACrawl-dependent; current processed corpus 6.5k docs Yes (user IDs) [49] Project GutenbergMultiLiterature / essays / speeches Crawl-dependent; current processed corpus 23.6k docs Yes (metadata author) [42] WikisourceMultiLiterature / drama / speeches Crawl-dependent; current processed corpus 100.8k docs Yes (author metadata / heuristics) [56] YouTube commentsMultiSocial media comments Crawl-dependent; current processed corpus 104.7k comments Yes (comment author) [15] Table 15: Raw sources used for the current AuthBench release. “Scale” reflects either the approxi- mate size of the upstream source or, for crawl-dependent sources, the size of the processed corpus produced by the current collection setting before final benchmark selection. 39 Source#Docs Share Exorde94,231 22.0% Wikisource78,984 18.4% Babel Briefings73,676 17.2% YouTube comments71,808 16.8% Blog Authorship Corpus22,494 5.3% Project Gutenberg18,739 4.4% Russian PD Corpus12,728 3.0% Douban Reviews10,424 2.4% Xiaohongshu / Weibo8,869 2.1% French PD Books8,761 2.0% German PD Corpus8,400 2.0% Spanish PD Books4,961 1.2% Amazon Reviews Multi (MARC)4,924 1.2% Stack Exchange Data Dump / API crawl 4,651 1.1% Arabic Classical Poetry2,503 0.6% arXiv Abstracts (metadata snapshot)1,784 0.4% Hindi Discourse Stories213 0.0% Table 16: Source composition of the current combined AuthBench export. All 17 configured sources are represented in the materialized benchmark; the table reports their realized document counts after filtering, redundancy reduction, cross-phase merge cleanup, and final selection. 40 Table 17: Licensing / terms and release mode by source. “Release mode” indicates whether we redistribute normalized text (Tier A) or provide manifest-only reconstruction (Tier B). This table covers all 17 sources in the current release. SourceLicense / terms (summary)Release mode Notes / pointer ExordeMIT (dataset card)Tier A Provider page explicitly states MIT. Babel BriefingsCC BY-NC-SA 4.0 (dataset card)Tier A Attribution, non-commercial use, and share-alike obligations apply. Amazon Reviews Multi (MARC) Provider-restricted / non-permissive redistribution; dataset card does not expose a standard open license Tier BSafer to require reconstruction from the original release rather than redistribute normalized text. Blog Authorship Corpus No clear standardized redistribution license found in the public host materials Tier BConservative manifest-only treatment is safer than direct text redistribution. arXiv Abstracts (metadata snapshot) Metadata access is available via arXiv services, but article licenses vary by paper Tier BWe treat this as metadata-derived text and do not redistribute PDFs or assume one uniform article license. Xiaohongshu / Weibo Platform / provider terms for redistribution are unclear Tier BManifest-only unless explicit redistribution permis- sion is documented. Douban ReviewsPlatform / provider terms for redistribution are unclear Tier BSame conservative handling as other platform- hosted review data. Hindi Discourse Stories Redistribution terms are unclear from public materials Tier BTreat as a research-use source unless explicit permission is documented. Spanish PD BooksPublic-domain source textsTier A Normalized text can be redistributed as public- domain material. French PD BooksPublic-domain source textsTier A Normalized text can be redistributed as public- domain material. Arabic Classical Poetry Public host available, but redistribution terms should be verified per host copy Tier BConservative default is manifest-only unless an explicit redistributable license is documented. Russian PD Corpus Public-domain source textsTier A Normalized text can be redistributed as public- domain material. German PD Corpus Public-domain source textsTier A Normalized text can be redistributed as public- domain material. Stack Exchange Data Dump / API crawl C BY-SA for public user contributions (version depends on post date; newer content is C BY-SA 4.0) Tier A Attribution and share-alike requirements apply to redistributed text. Project GutenbergMostly U.S.-public-domain texts, but Project Gutenberg license / trademark terms still apply Tier BConservative manifest-only handling avoids over- claiming unrestricted redistribution across jurisdic- tions and formats. WikisourceUsually C BY-SA with possible page-level public-domain variation depending on the work Tier BAttribution / share-alike obligations and page- level rights should be tracked before direct redistri- bution. YouTube comments Governed by YouTube API Terms and Developer Policies rather than a simple open-content license Tier BUse manifests / derived metadata only; avoid direct text redistribution by default. 41

Instant research discovery

AuthBench: A Large-Scale Multilingual Benchmark for Authorship Representation across Genres and Lengths

Intelligence

Summary

Entities (12)

Relation Signals (11)

Abstract

Tags

Links

PDF

Full Text

Paper	Published	Area	Tags	Intel	Citations