HERO: Length-Robust Machine-Influenced Text Detector

• HERO is a four-class detector that distinguishes human-written, machine-generated, machine-polished, and machine-translated texts, addressing nuanced AI involvement.
• It employs a hierarchical design with length-specialist models and Subcategory Guidance, ensuring effective detection across varied document lengths.
• Experimental results show a 2.5–3 mAP improvement over prior methods, highlighting its impact in authorship verification and AI-assistance applications.

Searching arXiv for the HERO paper and closely related machine-generated text detection work to ground the article in current literature. LLM use in document production increasingly spans not only fully synthetic writing but also narrower forms of intervention such as polishing and translation. The HiErarchical, Length-RObust Machine-Influenced Text Detector (HERO) was introduced to address that broader regime by separating text samples of varying lengths into four primary types: human-written, machine-generated, machine-polished, and machine-translated (Wang et al., 18 Sep 2025). Rather than treating authorship as a binary human-versus-machine problem, HERO is framed as a detector for machine-influenced text, with a hierarchical design intended to remain effective across document lengths and across fine-grained forms of machine involvement (Wang et al., 18 Sep 2025).

1. Task formulation and conceptual scope

HERO is defined around a four-way categorization problem: human-written, machine-generated, machine-polished, and machine-translated (Wang et al., 18 Sep 2025). This formulation differs from prior work in Machine Generated Text (MGT) detection that mostly focuses on simply identifying whether a document was human or machine written, thereby ignoring fine-grained uses such as grammar improvement or translation (Wang et al., 18 Sep 2025). The distinction matters because the paper explicitly contrasts benign uses of LLMs on human-written documents with cases in which a document is entirely produced by an LLM and may be more likely to be used to spread misinformation (Wang et al., 18 Sep 2025).

This framing places HERO in a broader authorship-and-manipulation regime than standard binary detectors. A plausible implication is that HERO treats machine influence as operationally heterogeneous rather than collapsing all non-human intervention into a single negative class. That interpretation is consistent with the contrast drawn in related work such as "GPT-who: An Information Density-based Machine-Generated Text Detector" (Venkatraman et al., 2023), which addresses both binary Turing Test detection and multi-class authorship attribution but remains fundamentally article-level and does not model machine-polished or machine-translated text as HERO does (Venkatraman et al., 2023).

The emphasis on machine-influenced text also aligns conceptually with later work arguing that MGT labels are often inexact because the boundary between human and machine text is ambiguous, especially under mixed or partially edited conditions (Wu et al., 2 Nov 2025). This suggests that HERO’s finer-grained labeling scheme is not merely taxonomic; it may also be a response to the inadequacy of binary supervision for realistic LLM-assisted writing settings.

2. Hierarchical design and length robustness

HERO is described as a HiErarchical, length-RObust machine-influenced text detector that learns to separate text samples of varying lengths by combining predictions from length-specialist models (Wang et al., 18 Sep 2025). The core architectural claim is therefore twofold: first, the system is hierarchical; second, the hierarchy is organized around robustness to text length (Wang et al., 18 Sep 2025).

The paper states that HERO accomplishes its classification objective by combining predictions from length-specialist models that have been trained with Subcategory Guidance (Wang et al., 18 Sep 2025). The available description does not specify the number of specialists, their length bins, or the precise fusion rule. Even so, the published summary makes clear that the hierarchy is not merely semantic; it is coupled to document length as an explicit design variable (Wang et al., 18 Sep 2025).

In the broader literature, length has emerged as a critical axis for detection quality. "Advancing Machine-Generated Text Detection from an Easy to Hard Supervision Perspective" argues that longer-text detection is easier and can serve as a more reliable supervisory source for harder short-text detection, with explicit evaluation at sentence and paragraph level (Wu et al., 2 Nov 2025). HERO’s “length-RObust” formulation is compatible with that line of reasoning, although the two works differ in scope: HERO is a detector for four machine-influence categories, whereas the easy-to-hard framework is a supervision strategy for binary machine-generated text detection (Wu et al., 2 Nov 2025).

A plausible implication is that HERO’s length-specialist structure is intended to reduce degradation caused by applying a single classifier uniformly across short and long documents. The abstract supports this interpretation only indirectly, through the explicit phrase “length-specialist models” and the claim that the system is robust to varying text lengths (Wang et al., 18 Sep 2025).

3. Subcategory Guidance

A central component of HERO is its Subcategory Guidance module (Wang et al., 18 Sep 2025). The paper specifies that, for categories that are easily confused, such as different source languages, the module encourages separation of the fine-grained categories, boosting performance (Wang et al., 18 Sep 2025).

This description establishes Subcategory Guidance as an auxiliary mechanism for disambiguation within coarse classes. The example of “different source languages” indicates that the subcategories are not limited to the four top-level labels, but may encode finer provenance distinctions beneath them (Wang et al., 18 Sep 2025). Because the available text does not provide equations, losses, or a formal hierarchy, the exact implementation is unspecified. Nevertheless, the intent is clear: when coarse classes exhibit internal heterogeneity that leads to confusability, HERO introduces additional guidance so that fine-grained distinctions become easier to preserve at the top level (Wang et al., 18 Sep 2025).

This design can be situated relative to other detection paradigms. GPT-who uses a fixed 44-dimensional feature representation derived from token surprisal statistics and maximum/minimum UID spans, followed by logistic regression (Venkatraman et al., 2023). That approach is interpretable and low-cost, but it has no explicit hierarchical composition mechanism and no analogue of Subcategory Guidance (Venkatraman et al., 2023). By contrast, HERO’s published summary suggests a learned detector in which auxiliary fine-grained structure plays a role in improving the coarse four-way decision boundary (Wang et al., 18 Sep 2025).

A plausible implication is that HERO is aimed not only at discriminating machine influence from human writing, but also at reducing confusions among superficially similar machine-influence modes, especially those sharing lexical or stylistic artifacts. The source-language example is the clearest indication of that objective in the published description (Wang et al., 18 Sep 2025).

4. Experimental regime and empirical performance

The paper reports extensive experiments across five LLMs and six domains (Wang et al., 18 Sep 2025). Within that evaluation regime, HERO is said to demonstrate the benefits of its design and to outperform the state-of-the-art by 2.5–3 mAP on average (Wang et al., 18 Sep 2025).

These are the central quantitative claims presently available for HERO. The summary does not enumerate the five LLMs, the six domains, the specific baselines, or per-class metrics. It does, however, identify mean average precision as the aggregate measure in which the average improvement is reported (Wang et al., 18 Sep 2025). Because the paper focuses on four primary text types rather than a binary label, mAP is consistent with a multi-class evaluation setting, although that interpretation should remain inferential rather than asserted.

The reported gain situates HERO against a recent landscape of supervised and statistical detectors. GPT-who, for example, is a supervised, statistical detector using UID-inspired features derived from GPT2-XL surprisal, and is evaluated on four benchmark datasets spanning over 15 domains and 35 recent LMs (Venkatraman et al., 2023). That paper emphasizes document-level authorship signatures and article-level aggregation rather than fine-grained machine-influence classes (Venkatraman et al., 2023). HERO’s reported 2.5–3 mAP average improvement therefore appears to reflect a different target problem as well as a different model design (Wang et al., 18 Sep 2025).

Later work on easy-to-hard supervision reports robustness gains under cross-LLM, cross-domain, mixed text, and paraphrase attacks, with evaluation using TPR@FPR-1% and AUROC rather than mAP (Wu et al., 2 Nov 2025). This highlights that HERO’s empirical contribution lies less in adversarial or sentence-level binary detection and more in four-way machine-influence discrimination over varying lengths and domains (Wang et al., 18 Sep 2025).

5. Relation to prior detection paradigms

Prior work in MGT detection has largely emphasized binary human-versus-machine discrimination (Wang et al., 18 Sep 2025). HERO is explicitly positioned against that tradition by addressing fine-grained uses of LLMs rather than only full generation (Wang et al., 18 Sep 2025). In that respect, it overlaps partly with authorship-attribution and AI-assistance detection, but its four-class framing is more specific than many earlier detectors.

The contrast with GPT-who is especially instructive. GPT-who is a supervised, domain-agnostic statistical detector that models the statistical signature of each LLM and human author using UID-based features, and it is explicitly intended to address both Turing Test detection and authorship attribution (Venkatraman et al., 2023). Its unit of prediction is the article/document, and it does not provide a sentence-level or token-level detector output, nor any explicit hierarchical mechanism (Venkatraman et al., 2023). HERO, by contrast, is hierarchical, uses length-specialist models, and is aimed at separating machine-generated, machine-polished, and machine-translated text from human-written text (Wang et al., 18 Sep 2025).

The contrast with the easy-to-hard framework is different. "Advancing Machine-Generated Text Detection from an Easy to Hard Supervision Perspective" focuses on binary machine-generated text detection under inexact supervision, introducing an easy supervisor on longer texts to enhance a harder target detector on shorter/original texts (Wu et al., 2 Nov 2025). That work is conceptually hierarchical in supervision and explicitly motivated by text-length robustness, but it does not introduce a detector explicitly named HERO and does not provide a conventional hierarchical inference architecture (Wu et al., 2 Nov 2025). HERO can therefore be distinguished as a detector architecture for multi-class machine influence, whereas the easy-to-hard method is a training framework for binary MGT detection (Wang et al., 18 Sep 2025); (Wu et al., 2 Nov 2025).

A plausible implication is that HERO occupies an intermediate position between broad document-level statistical detectors and supervision-centric robustness frameworks: it is more semantically granular than the former and more directly architectural than the latter.

6. Significance, limitations, and interpretive issues

HERO’s significance lies in formalizing machine influence as a four-class problem and in tying that problem to length-specialized hierarchical prediction (Wang et al., 18 Sep 2025). The inclusion of machine-polished and machine-translated categories is especially important for settings in which LLM assistance is permitted in some forms but not others, or in which provenance analysis requires distinguishing original generation from post hoc transformation (Wang et al., 18 Sep 2025). This suggests practical relevance for editorial policy, authorship verification, and multilingual content analysis, though such applications are an inference from the task definition rather than an explicitly listed deployment scenario.

Several limitations follow from the currently available published summary. The paper excerpt provides no equations, no model architecture diagram, no explicit formulation of the Subcategory Guidance objective, no dataset construction details, and no ablation breakdown (Wang et al., 18 Sep 2025). As a result, claims about fusion strategy, backbone choice, specialist count, or training loss would exceed the presently recoverable evidence. The safest characterization is therefore that HERO combines length-specialist models with Subcategory Guidance to separate four primary classes over varying lengths, and that this combination yields an average 2.5–3 mAP improvement over prior state of the art across five LLMs and six domains (Wang et al., 18 Sep 2025).

A further interpretive issue concerns the broader debate on label granularity. Work on inexact supervision argues that hard binary labels obscure a latent continuum of machine-likeness, especially for mixed text and partial machine involvement (Wu et al., 2 Nov 2025). HERO’s categorical treatment of machine-generated, machine-polished, and machine-translated text does not eliminate that continuum, but it does move beyond a single binary boundary (Wang et al., 18 Sep 2025). This suggests that HERO may be viewed as part of a broader methodological shift from detecting “AI text” to characterizing how a machine influenced the text.

In that sense, HERO extends the field’s problem definition as much as its model design. Its main contribution is not merely higher detection accuracy, but the proposition that machine involvement in text should be represented hierarchically, evaluated across variable lengths, and partitioned into operationally distinct influence types (Wang et al., 18 Sep 2025).