AGENT-X: Adaptive Guideline-based Expert Network for Threshold-free AI-generated teXt detection

Cheng Peng; Jiatao Li; Mao Ye; Xiaojun Wan; Xunjian Yin

REVIEW 3 major objections 5 minor 12 references

AGENT-X: Adaptive Guideline-based Expert Network for Threshold-free AI-generated teXt detection

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AGENT-X claims that AI-generated text can be detected zero-shot and threshold-free by having specialized LLM agents evaluate hand-curated stylistic guidelines across semantic, stylistic, and structural dimensions, then confidence-weight…

desk verdict AGENT-X's accuracy edge over baselines is likely a threshold artifact; the appendix AUROC tables show the underlying detection signal is not better, so the 'substantially surpasses' claim needs a fair baseline comparison. read the letter →

arxiv 2505.15261 v1 pith:BTQ5SUMX submitted 2025-05-21 cs.CL

Jiatao Li , Mao Ye , Cheng Peng , Xunjian Yin , Xiaojun Wan This is my paper · ORCID

classification cs.CL

keywords AI-generatedtextdetectionzero-shotmulti-agentLLMsystemsstylisticssystemicfunctionallinguisticsconfidencecalibrationthreshold-freeclassificationauthorshipattribution

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that detecting AI-generated text does not require labeled training data or a tuned decision threshold. It proposes AGENT-X, a zero-shot pipeline in which a router picks relevant hand-written stylistic guidelines, specialized language-model agents evaluate the text against them in three dimensions (semantic, stylistic, structural), and a meta agent confidence-weights their votes into a final answer with an explicit rationale. The authors report that this pipeline beats supervised and zero-shot baselines in classification accuracy on news summaries, creative writing, and biomedical abstracts generated by ChatGPT, GPT-4, and Claude-3, with the largest margins on biomedical text where threshold-based detectors degrade. A reader should care because the method turns detection into an interpretable linguistic judgment rather than a probability cutoff, which is closer to what a deployed content-integrity tool needs.

What carries the argument

The machinery is a three-part pipeline. A router agent infers the text's domain and stylistic features, then activates only the most relevant hand-curated guidelines. Each base agent produces a decision, a rationale, and a verbalized confidence; the SteeringConf calibration repeats the judgment under five symmetric prompts (very cautious to very confident) and combines them into $c_{\mathrm{cal}} = \mu_c \cdot \kappa_{\mathrm{ans}} \cdot \kappa_{\mathrm{conf}}$, where $\mu_c$ is the mean verbalized confidence and the two consistency terms penalize disagreement across prompts. A meta agent then aggregates the calibrated reports by confidence-weighted reasoning. The entire design is threshold-free: the final classification is the meta agent's explicit decision, not a score compared against a tuned cutoff.

What would settle it

Run AGENT-X on a held-out corpus of concise formal human writing (legal memos, technical specifications, or academic abstracts) paired with current LLM generations on the same prompts, and compare its accuracy with the zero-shot likelihood baseline. If detection accuracy on that corpus drops to or below the baseline, or if the router activates 'AI-like' guidelines on clearly human formal prose, the generalization claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a small set of theoretically grounded stylistic guidelines, applied by independent LLM agents and reconciled by a meta agent, is enough to separate human from machine authorship in zero-shot settings. On the three datasets and four generators tested, it reports the highest average accuracy among all compared methods: 86.0% on ChatGPT, 85.9% on GPT-4, 83.8% on Claude-3 Opus, and 80.0% on Claude-3 Sonnet, against best zero-shot baselines of 82.6%, 68.0%, 81.8%, and 78.6% respectively. The framework also produces a per-case rationale, and its ablation study attributes most of the gain to the curated guidelines and to the adaptive router rather than to the multi-agent architecture alone.

Load-bearing premise

The load-bearing premise is that the hand-curated stylistic guidelines—that human writing is variable and spontaneous while AI writing is uniform, balanced, and formal—hold across new domains and generator models, since the paper tests only three English datasets and four commercial LLMs.

Editorial extensions

If this is right

Deployed detectors can classify AI-generated text without a held-out calibration set to choose a threshold, which removes a major practical barrier to zero-shot use.
Each decision comes with a rationale that cites concrete stylistic evidence, so users can audit why a text was flagged.
The accuracy advantage over probability-curvature baselines is largest on PubMedQA-style biomedical text, suggesting the guideline approach transfers better to unfamiliar domains than likelihood-based detectors.
Because the router suppresses irrelevant guidelines, the framework remains interpretable even when the full guideline set is large.
Swapping in a single comprehensive agent or random guidelines drops accuracy substantially, so the structured division of labor is load-bearing for the reported results.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

My inference: the hand-curated guidelines are tuned to English expository and narrative prose, so non-English or highly dialectal text likely needs revised guidelines before the reported gains transfer.
My inference: the five steering prompts calibrate confidence relative to the specific base model, so switching to a weaker open-weight model may require re-tuning the steering wording; the paper's limitation section concedes a dependency on a capable base LLM.
My inference: a natural extension the paper does not test is bootstrapping guidelines from a few labeled examples per domain, which would let the router adapt beyond the fixed stylistics literature.
My inference: because the meta agent is also a prompted LLM, the aggregation step could inherit the same miscalibration the steering method corrects at the base-agent level; comparing meta-agent expected calibration error with and without steer calibration would settle this.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The short version: the paper builds something genuinely interpretable and the writing is clean, but the headline accuracy margin is likely an artifact of the baseline threshold protocol. Baselines get thresholds tuned on SQuAD and then applied to XSum, WritingPrompts, and PubMedQA, while AGENT-X is threshold-free. The AUROC tables in Appendix F undercut the accuracy story: Fast-DetectGPT and Likelihood usually rank better (e.g., 0.9615 vs 0.9206 on ChatGPT). That means the detection signal is not better; the practical advantage is being threshold-free. That is a real advantage, but it is not what the abstract claims. The reader's stress-test holds up.

What is genuinely new: the guideline taxonomy (semantic/stylistic/structural) and the router-plus-specialized-agents architecture. Decomposing detection into interpretable dimensions, with a meta agent aggregating calibrated confidences, is clean, and the ablation shows the components matter — removing guidelines drops accuracy to 61%. The full prompt set is in the appendix, which is good for reproducibility.

Soft spots beyond the threshold issue: no error bars, no significance tests, no code release yet. The guidelines are hand-curated generalizations ('human text is variable,' 'AI text is balanced') that are broad enough to raise doubts about domains outside news, stories, and biomedical QA. Three domains and four commercial LLMs is a thin base for the generalization claim. The core idea is nonetheless coherent, and the limitation section is honest about needing a capable base model.

This paper deserves a serious referee. The threshold comparison is fixable: rerun baselines with per-dataset or oracle thresholds, show the AUROC gap in the main paper, and add variance estimates. With those changes, the claim can be recalibrated to 'comparable detection signal, no threshold needed, better interpretability,' which would be credible. As it stands, I would not cite it as a 'substantially surpasses' result. Worth a reading group discussion on evaluation protocol design.

Referee Report

3 major / 5 minor

Summary. The paper proposes AGENT-X, a zero-shot, threshold-free multi-agent framework for detecting AI-generated text. Detection guidelines are organized into semantic, stylistic, and structural dimensions; individual LLM agents judge each dimension using five steering prompts for confidence calibration; a router agent selects guidelines; and a meta agent aggregates decisions. Experiments on XSum, WritingPrompts, and PubMedQA with ChatGPT, GPT-4, and Claude-3 outputs report higher average accuracy than supervised and zero-shot baselines, and an ablation study attributes most of the gain to the guidelines, routing, and confidence calibration.

Significance. If the headline accuracy results held, AGENT-X would be a practical, interpretable alternative to threshold-dependent detectors, and the paper's explicit guideline catalog and prompt suite would be a useful resource. The ablation is carefully designed, and the guideline inventory in Appendix A is a concrete, reusable artifact. However, the central comparison is undermined by the baseline threshold protocol: baselines are tied to a single threshold tuned on SQuAD, while AGENT-X is threshold-free. The threshold-independent AUROC results in Appendix F show several baselines ranking texts better than AGENT-X, so the claimed "substantially surpasses" is not yet established. The significance is contingent on a fairer, threshold-matched evaluation.

major comments (3)

[Tables 1-2, Appendix C] The accuracy comparison is not methodologically fair to the baselines. As stated in Appendix C and the table captions, baseline thresholds are optimized on a separate SQuAD-gpt-neo-2.7B validation set and then frozen for XSum, WritingPrompts, and PubMedQA, whereas AGENT-X produces its binary decision without any threshold. A fixed threshold chosen for one domain can be badly miscalibrated on another, which is visible in the large column-to-column swings of Fast-DetectGPT (e.g., 0.9467 on XSum vs 0.5433 on PubMed for ChatGPT in Table 1). The claim in Section 6 that AGENT-X "substantially surpasses" zero-shot baselines in accuracy therefore conflates detection quality with threshold calibration. Please re-run the baselines under a per-dataset oracle threshold or an explicitly matched operating-point protocol, and restrict the accuracy claim to the protocol actually used.
[Appendix F, Tables 5-6] The AUROC results contradict the interpretation that AGENT-X is a better detector. AUROC is threshold-independent, and under this metric Fast-DetectGPT and Likelihood outperform AGENT-X on several source-model settings (e.g., average ChatGPT AUROC 0.9615 and 0.9364 vs AGENT-X 0.9206; average Claude-3-Opus 0.9519 and 0.9220 vs AGENT-X 0.9081). The text in Section 6 says AUROC "inherently favors threshold-dependent methods," but AUROC does not require selecting a threshold; it measures ranking quality and should favor methods with better separation. Please explain why the method that ranks below baselines in AUROC nevertheless reports higher accuracy, and provide per-dataset threshold-matched accuracy numbers to resolve the discrepancy.
[Section 5.1, Section 6, Appendix A] The generalization claim is tested on only three domains (XSum, WritingPrompts, PubMedQA) and four commercial LLMs, while the method's behavior is entirely determined by hand-curated guidelines whose premises are that human text is "variable/spontaneous/less balanced" and AI text is "uniform/balanced/formal" (Appendix A). The paper provides no mechanism for the router or base agents to detect when these premises fail, e.g., concise formal human writing or newer generator styles that follow the "AI" profile. The Limitations section mentions only the capability of the base LLM. Please either evaluate on out-of-distribution domains and generators or explicitly limit the generalization claim to the tested distribution.

minor comments (5)

[Section 4.2] The term κans is defined as the maximum fraction of steering prompts that agree with the majority class, so it is not a measure of answer consistency in the usual sense; it is always at least 0.5 and reduces to a simple scaling factor. Please rename or redefine it, and state how the final decision is computed when the calibrated confidence is used to pick k*.
[Tables 5-6] The caption states that baseline AUROC results are "cited from (Bao et al., 2024)" while Tables 1-2 appear to report newly computed accuracy numbers; please state explicitly which numbers are reproduced and which are newly computed, since the two tables should come from the same evaluation pipeline.
[Appendix D and Tables 1-3] Please fix the typo "zeo-shot" in Appendix D and standardize the dataset name, which appears as both "Writing" and "WritingPrompts" in the table headers and text.
[Figure 2 and Appendix B] The worked example quotes a news-style article without source attribution; if the text is taken from a published article, please add a citation or a note that it is an illustrative example.
[Appendix E, Table 4] The "Steer token prob" row reports accuracy and AUROC, but the text does not define how token-probability confidence is extracted for the deepseek-chat API; please add a sentence describing the extraction procedure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: AGENT-X's detection pipeline and experimental claims are self-contained and do not reduce to their own inputs.

full rationale

The paper's central derivation is an empirically evaluated method, not a closed logical chain that presupposes its conclusion. The hand-curated detection guidelines (Appendix A) are presented as inputs drawn from stylistics literature and are not derived from, or fitted to, the test datasets; the SteeringConf calibration is explicitly attributed to prior work by Zhou et al. (2025); and the meta-agent aggregation is an architectural procedure rather than a restatement of the desired accuracy result. The reported accuracy comparisons in Tables 1-2 do depend on a threshold calibration choice for baseline methods, and the authors' decision to prioritize accuracy over AUROC is debatable as an evaluation-standard matter, but that is a fair-comparison or external-validity concern, not a circularity: no fitted parameter is renamed as a prediction, no target quantity is used in its own definition, and no load-bearing premise is justified only by the authors' own prior work. The AUROC results in Appendix F, where some baselines outperform AGENT-X, actually provide independent, threshold-free evidence that the method's rank-ordering ability is competitive rather than forced. Under the stated hard rules, especially the requirement to quote a specific reduction and to avoid treating evaluation-standard disagreements as circularity, no circular step can be identified.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The framework introduces no scientific entities or fitted parameters. It relies on four domain assumptions about the validity of the guidelines, the steering calibration, the router, and the meta agent. These are practical design choices, not parameters learned from the evaluation data, so they are listed as assumptions rather than free parameters.

assumptions (4)

domain assumption The curated stylistic guidelines are reliable discriminators of AI-generated vs. human text across domains and generators.
Section 4.2 and Appendix A: each base agent applies a fixed guideline; the whole framework depends on their validity beyond the three tested domains and four LLMs.
domain assumption Semantic steering confidence calibration yields trustworthy confidence estimates.
Section 4.2 and Appendix E: the method borrows SteeringConf from Zhou et al. (2025) and assumes the calibrated scores are meaningful for weighted aggregation.
domain assumption The router correctly infers domain and stylistic features, and the selected guideline subset improves detection.
Section 4.1: the router is a prompted LLM with no verification of its selection quality; the ablation shows removing it hurts, but the selection itself is not validated.
domain assumption The meta agent's confidence-weighted aggregation produces correct final decisions.
Section 4.3: the meta agent is another prompted LLM; its aggregation rule is not formally specified and is only tested implicitly through the main results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AGENT-X: Adaptive Guideline-based Expert Network for Threshold-free AI-generated teXt detection." pith.science (2026). https://pith.science/paper/BTQ5SUMX

@misc{pith2026250515261,
  author       = {Pith},
  title        = {Pith review of: AGENT-X: Adaptive Guideline-based Expert Network for Threshold-free AI-generated teXt detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BTQ5SUMX}},
  note         = {Machine review of arXiv:2505.15261}
}

read the original abstract

Existing AI-generated text detection methods heavily depend on large annotated datasets and external threshold tuning, restricting interpretability, adaptability, and zero-shot effectiveness. To address these limitations, we propose AGENT-X, a zero-shot multi-agent framework informed by classical rhetoric and systemic functional linguistics. Specifically, we organize detection guidelines into semantic, stylistic, and structural dimensions, each independently evaluated by specialized linguistic agents that provide explicit reasoning and robust calibrated confidence via semantic steering. A meta agent integrates these assessments through confidence-aware aggregation, enabling threshold-free, interpretable classification. Additionally, an adaptive Mixture-of-Agent router dynamically selects guidelines based on inferred textual characteristics. Experiments on diverse datasets demonstrate that AGENT-X substantially surpasses state-of-the-art supervised and zero-shot approaches in accuracy, interpretability, and generalization.

Figures

Figures reproduced from arXiv: 2505.15261 by the authors.

**Figure 2.** Example detection case of AGENT-X. The router agent dynamically activates relevant stylistic guidelines based on inferred textual characteristics. Each guideline-specific agent independently assesses the input text, providing calibrated confidence scores and detailed linguistic reasoning. The meta agent aggregates these evaluations via confidence-weighted voting, making the final interpretable classification decisio… view at source ↗

**Figure 3.** Probability histogram and Expected Calibration Error (ECE) for (a)(c) LLM generated verbalised [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 5 canonical work pages

[1]

The text’s domain (e.g., medical abstract, literary fiction, news article, etc.)

work page
[2]

[Input Text Placeholder]

Key stylistic features that strongly characterize the text. Input Text: “[Input Text Placeholder]” Respond strictly in the following JSON format: { "Domain": "<identified domain>", "Stylistic Features" : ["<feature 1>", "<feature 2>", ". . . "] } 19 Prompt G.2: Router Agent Guideline Selection Prompt Given a text’s domain and stylistic features, your task...

work page
[9]

concluded that the text is AI-generated, with confidence levels ranging from moderate to high (0.4246 to 0.7574). The two agents that identified the text as human-written had confidence levels of 0.7560 and 0.7620, which are relatively high but not enough to outweigh the consensus of the other agents. The text's balanced narrative, polished language, and ...

work page 2008
[10]

RoBERTa Classifier (Park et al., 2021) A pop- ular and competitive detector method

as the reference model. RoBERTa Classifier (Park et al., 2021) A pop- ular and competitive detector method. Recognize LLM generated text by fine-tuning the RoBERTa classifier on large amounts of labeled text. E Additional Confidence Calibration Experiments We performed additional experiments to evalu- ate the intrinsic confidence calibration capabili- tie...

work page 2021
[1954]

Modern library of the world’s best books

Rhetoric. Modern library of the world’s best books. Modern Library. Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, Marc’Aurelio Ranzato, and Arthur Szlam

work page
[2008]

In Proceedings of the ECAI’08 Workshop on Uncovering Plagiarism, Authorship and Social Software Misuse, Patras, Greece, July 22, 2008, vol- ume 377 of CEUR Workshop Proceedings

Detecting fake content with relative entropy scoring. In Proceedings of the ECAI’08 Workshop on Uncovering Plagiarism, Authorship and Social Software Misuse, Patras, Greece, July 22, 2008, vol- ume 377 of CEUR Workshop Proceedings. CEUR- WS.org. Moxin Li, Wenjie Wang, Fuli Feng, Fengbin Zhu, Qifan Wang, and Tat-Seng Chua. 2024. Think twice before trusting...

arXiv 2008
[2018]

Preprint, arXiv:1808.08745

Don’t give me the details, just the summary! topic-aware convolutional neural networks for ex- treme summarization. Preprint, arXiv:1808.08745. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Ale- man, Diogo Almeida, Janko Altenschmidt, Sam Alt- man, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, ...

arXiv 2024
[2019]

arXiv preprint arXiv:1906.03351

Real or fake? learning to discriminate ma- chine from human generated text. arXiv preprint arXiv:1906.03351. Guangsheng Bao, Yanbin Zhao, Zhiyang Teng, Linyi Yang, and Yue Zhang. 2023. Fast-detectgpt: Ef- ficient zero-shot detection of machine-generated text via conditional probability curvature. CoRR, abs/2310.05130. Guangsheng Bao, Yanbin Zhao, Zhiyang ...

arXiv 1906

Show all 12 references

[2021]

Preprint, arXiv:2105.09680

Klue: Korean language understanding evalua- tion. Preprint, arXiv:2105.09680. Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng 11 Su, Xin Cong, et al. 2024. Chatdev: Communicative agents for software development. In Proceeding...

2024 arXiv
[2024]

Preprint, arXiv:2402.00251

Efficient Non-Parametric Uncertainty Quan- tification for Black-Box Large Language Models and Decision Planning. Preprint, arXiv:2402.00251. Eduard Tulchinskii, Kristian Kuznetsov, Laida Kushnareva, Daniil Cherniavskii, Sergey Nikolenko, Evgeny Burnaev, Serguei Barannikov, and...

2024 arXiv
[2025]

Yes," "No,

Calibrating llm confidence with semantic steering: A multi-prompt aggregation framework. Preprint, arXiv:2503.02863. A Detection Guidelines Detection Guidelines Semantic Dimension Guideline 1 : Domain-Specific References Human-written: Briefly mentions specialized terms withou...

arXiv
[2309]

Mingjian Jiang, Yangjun Ruan, Sicong Huang, Saifei Liao, Silviu Pitis, Roger Baker Grosse, and Jimmy Ba

International Committee on Computational Linguistics. Mingjian Jiang, Yangjun Ruan, Sicong Huang, Saifei Liao, Silviu Pitis, Roger Baker Grosse, and Jimmy Ba. 2023. Calibrating language models via aug- mented prompt ensembles. Zhengbao Jiang, Jun Araki, Haibo Ding, and Graham ...

2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

The text’s domain (e.g., medical abstract, literary fiction, news article, etc.)

work page

[2] [2]

[Input Text Placeholder]

Key stylistic features that strongly characterize the text. Input Text: “[Input Text Placeholder]” Respond strictly in the following JSON format: { "Domain": "<identified domain>", "Stylistic Features" : ["<feature 1>", "<feature 2>", ". . . "] } 19 Prompt G.2: Router Agent Guideline Selection Prompt Given a text’s domain and stylistic features, your task...

work page

[3] [9]

concluded that the text is AI-generated, with confidence levels ranging from moderate to high (0.4246 to 0.7574). The two agents that identified the text as human-written had confidence levels of 0.7560 and 0.7620, which are relatively high but not enough to outweigh the consensus of the other agents. The text's balanced narrative, polished language, and ...

work page 2008

[4] [10]

RoBERTa Classifier (Park et al., 2021) A pop- ular and competitive detector method

as the reference model. RoBERTa Classifier (Park et al., 2021) A pop- ular and competitive detector method. Recognize LLM generated text by fine-tuning the RoBERTa classifier on large amounts of labeled text. E Additional Confidence Calibration Experiments We performed additional experiments to evalu- ate the intrinsic confidence calibration capabili- tie...

work page 2021

[5] [1954]

Modern library of the world’s best books

Rhetoric. Modern library of the world’s best books. Modern Library. Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, Marc’Aurelio Ranzato, and Arthur Szlam

work page

[6] [2008]

In Proceedings of the ECAI’08 Workshop on Uncovering Plagiarism, Authorship and Social Software Misuse, Patras, Greece, July 22, 2008, vol- ume 377 of CEUR Workshop Proceedings

Detecting fake content with relative entropy scoring. In Proceedings of the ECAI’08 Workshop on Uncovering Plagiarism, Authorship and Social Software Misuse, Patras, Greece, July 22, 2008, vol- ume 377 of CEUR Workshop Proceedings. CEUR- WS.org. Moxin Li, Wenjie Wang, Fuli Feng, Fengbin Zhu, Qifan Wang, and Tat-Seng Chua. 2024. Think twice before trusting...

arXiv 2008

[7] [2018]

Preprint, arXiv:1808.08745

Don’t give me the details, just the summary! topic-aware convolutional neural networks for ex- treme summarization. Preprint, arXiv:1808.08745. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Ale- man, Diogo Almeida, Janko Altenschmidt, Sam Alt- man, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, ...

arXiv 2024

[8] [2019]

arXiv preprint arXiv:1906.03351

Real or fake? learning to discriminate ma- chine from human generated text. arXiv preprint arXiv:1906.03351. Guangsheng Bao, Yanbin Zhao, Zhiyang Teng, Linyi Yang, and Yue Zhang. 2023. Fast-detectgpt: Ef- ficient zero-shot detection of machine-generated text via conditional probability curvature. CoRR, abs/2310.05130. Guangsheng Bao, Yanbin Zhao, Zhiyang ...

arXiv 1906