D2C-Routing: Dimension-to-Composition Evidence Routing for Mixed-Origin AI-Generated Text Detection

AI summary

Overview

Research area: AI-generated text detection, specifically mixed-origin authorship attribution in natural language processing.

Technical level: Advanced. The paper assumes familiarity with transformer encoders, cross-entropy and ranking losses, gated fusion, and low-false-positive-rate detector evaluation metrics.

One-sentence scope: The paper recasts AI-text detection from a binary human-versus-machine judgment into a two-dimension source attribution problem (content origin and expression origin) composed into four collaboration types, and introduces D2C-Routing to solve it.

What This Paper Is About

Most AI-text detectors answer a single question: was this document written by a human or a machine? That framing fails for mixed-origin writing, where the origin of the content and the origin of the expression can differ—for example, human-originated ideas written up in AI-polished wording, or AI-originated content rewritten by a person. The authors adopt HART's four-label taxonomy (HH, HA, AH, AA) and build D2C-Routing, which routes document evidence into separate content-side and expression-side pathways, supervises each source dimension directly, and then composes the two into a final four-way label with a learned gate.

Key Contributions

1. Linguistically motivated evidence organization. The authors split document-internal evidence into a content branch (entity-chain coherence, RST discourse motifs) and an expression branch (lexical-connective choices, rhythm/POS patterns, surface regularity), giving each family a dimension-specific role rather than flat-concatenating features.

2. The D2C-Routing architecture. A dimension-to-composition model that combines dimension-specific pathways, supervised content-origin and expression-origin heads, and a learned gated composition layer that outputs HH/HA/AH/AA.

3. Empirical evidence under the MixD2C split. D2C-Routing improves average and AA low-FPR ranking over same-split single-model controls, while a D2C-Routing-based detector system (D2C-base Fusion) reaches 0.8603 four-way Avg TPR@1%FPR, 6.5 points above the same-split RACE-local rerun.

4. A transparent reconstructed split and protocol-separated baselines. MixD2C is a 70/10/20 split (11,200/1,600/3,200 train/dev/test) derived from the released HART development and test files, with published RACE values kept as external references and a local RACE rerun used as the same-split comparison.

Main Findings

• The headline detector-system number: D2C-base Fusion reaches 0.9889 macro AUROC, 0.9195 macro-F1, and 0.8603 four-way Avg TPR@1%FPR, with class-wise TPR@1%FPR of 0.9925 (HH), 0.8888 (HA), 0.7892 (AH), and 0.7708 (AA).

• The gain is on average and AA, not uniform superiority: The single-model D2C-Routing (3-seed, dual encoder) reaches 0.8440 Avg TPR@1%FPR and 0.7701 AA, but stays below RACE-local on AH (0.7435 versus 0.7696). RACE-local is strong on ordinary metrics (0.9849 AUROC, 0.9030 F1-score) but weak on AA TPR@1%FPR at 0.5752.

• AH is the hardest boundary: In the branch diagnosis, content-origin accuracy on AH is 0.9477 while expression-origin accuracy is only 0.6438, localizing the single-model weakness to expression-origin recognition.

• Dimension supervision and learned composition both help: Removing the dimension loss lowers Avg TPR@1%FPR to 0.7801 and AA to 0.6662; removing the gate lowers them to 0.7838 and 0.6870, compared with 0.8110 and 0.6841 for the shared-encoder routed model. The learned composition control reports Avg 0.8110 versus 0.7754 (fixed), 0.7942 (concat), and 0.7868 (hard).

• The source dimensions are learnable: The content head reaches 0.9937 AUROC and the expression head 0.9871 AUROC.

• Cost-matched controls temper the interpretation: At roughly matched parameter scale, the D2C cost-audit system (625.34M parameters, 3 forward calls) reaches 0.9894 AUROC, 0.9185 F1, 0.8601 Avg TPR@1, and 0.7686 AA, versus flat-features 5x uniform (623.87M, 5 calls) at 0.9880 / 0.9093 / 0.8443 / 0.7772. The Avg TPR advantage over flat 5x is positive but not significant at 95%, and flat 5x has higher AA TPR.

• Frozen probes are far weaker: Freezing D2C-Routing at the last layer drops Avg TPR@1%FPR from 0.8206 to 0.5444 and AA from 0.6777 to 0.3854.

• Feature routing alignment is not proven: Correct, swapped, and fixed-random routing are statistically similar on the principal low-FPR metric; the authors state the evidence does not establish a uniquely optimal or causally interpretable handcrafted feature assignment.

• Backbone results: DeBERTa-base D2C reaches 0.9877 macro AUROC, 0.9094 F1, 0.8320 Avg TPR@1, 0.7614 AH, and 0.7051 AA across three seeds. ModernBERT D2C (149.73M) gives 0.9851 / 0.9026 / 0.8073 / 0.7435 / 0.6796, versus ModernBERT text-only (149.02M) at 0.9850 / 0.9089 / 0.8032 / 0.7484 / 0.6562—ModernBERT does not yield a reliable D2C-over-text advantage.

• Transfer is dimension-specific, not general: Zero-shot expression-origin transfer gives AUROC 0.8993 on APT and 0.9392 on PAN; zero-shot content transfer on PAN is 0.8012 versus 0.7961 for a matched text-only score. Direct HART-to-MixSet transfer reaches only 0.2262 Macro-F1 and 0.4372 Macro AUROC.

• Official HART task collapses favor the fused system slightly: For Level-1/2/3 (AUROC / F1 / TPR@5%FPR), RACE-local reports 0.9985 / 0.9925 / 0.9967, 0.9943 / 0.9641 / 0.9731, and 0.9841 / 0.9365 / 0.9398; D2C-base Fusion reports 0.9988 / 0.9925 / 0.9962, 0.9959 / 0.9702 / 0.9850, and 0.9906 / 0.9498 / 0.9685.

• Scalar detectors collapse on the four-way setting: Published RACE Table 2 references include Binoculars-MLP at 0.1170 Avg TPR@1%FPR, Fast-DetectGPT-MLP at 0.1080, and TDT-SVC at 0.0233.

• Grouped robustness is competitive but not uniformly significant: F1 0.9100 and Avg 0.8507 under the stricter in-benchmark grouped split.

Methodology in Plain English

The authors take a document and refuse to reduce it to a single "AI-likeness" score. Instead, they define two questions: where did the content come from, and where did the expression come from? Each gets a binary human/AI answer, and the pair composes into the four labels HH, HA, AH, and AA.

To answer those questions, they organize observable text cues into two evidence pathways. The content pathway looks at how information is held together—whether entities recur consistently, whether sentences overlap in their referents, and how discourse relations and motifs are arranged. The expression pathway looks at how the wording is realized—connective and lexical choices, rhythm and part-of-speech patterns, and surface regularity. A RoBERTa encoder supplies the base document representation, learned projectors map each evidence group into its pathway, and each pathway feeds a binary head supervised directly on its source dimension.

A learned gate then mixes the two pathway states—using the document anchor plus the two head scores to decide the weighting—before a four-way classifier produces the final HH/HA/AH/AA distribution. Training combines cross-entropy on content origin, cross-entropy on expression origin, a term specifically separating AH from AA, cross-entropy on the four-way label, and an AA one-vs-rest ranking term aligned with low-FPR evaluation.

For the final reported system, they fuse several RoBERTa-base-family member models by weighted probability interpolation. Candidate pools and weights are fixed on the development set and evaluated once on test, with no test-set tuning; the final fusion has three nonzero-weight members. They also build MixD2C by merging the released HART development and test files and stratifying by domain and class into the 70/10/20 split.

Why This Matters

Impact on research: The paper argues that the field's default binary framing is structurally incomplete for collaborative writing, and that a scalar AI-likeness score cannot say which source dimension changed. It offers a testable alternative: intermediate source dimensions should be learnable, and composing them should help at strict false-positive operating points. It also provides a protocol-separated benchmark setup that distinguishes published reference values from same-split reruns, and reports negative results (grouped robustness, ModernBERT, routing-alignment controls, external transfer) alongside positive ones.

Real-world applications:

• Auditing editorial and publishing pipelines where a human drafted content and an AI polished the prose, or vice versa.
• Academic-integrity review, where the useful question is often whether the ideas or the wording were machine-originated.
• Content-provenance and platform moderation tooling that needs finer provenance than a document-level AI flag.
• Research on writing-process analysis, where distinguishing content origin from expression origin supports different interventions.

Industry relevance: Teams deploying AI writing assistants, newsroom verification tools, or enterprise content-governance systems need graded provenance signals rather than one flag. The paper's cost disclosures—parameter counts and online forward calls for cost-matched controls—let practitioners judge whether the accuracy gains justify the compute. The authors explicitly caution that such detectors should be decision support and auditing aids, not sole evidence for punitive decisions about authorship, academic integrity, employment, or access to services.

Future Directions

• Fix the expression bottleneck on AH. Expression-origin accuracy of 0.6438 on AH versus content-origin accuracy of 0.9477 is the clearest localized failure, and the authors name AH expression recognition as the main remaining challenge.
• Make routing causally interpretable. Correct, swapped, and fixed-random routing perform similarly on the principal low-FPR metric, so the evidence does not yet support a uniquely optimal or causally meaningful handcrafted feature assignment.
• Improve out-of-domain four-way generalization. Direct HART-to-MixSet transfer of 0.2262 Macro-F1 and 0.4372 Macro AUROC means external experiments currently support dimension-specific transfer and supervised adaptation rather than broad out-of-domain four-way performance.
• Establish a stricter comparison against published RACE numbers. Because the release does not package frozen sample IDs or checkpoints, the local run is a same-split rerun rather than a reproduction, leaving the published RACE values as related-reconstruction context.
• Find backbones that reliably benefit from the D2C design. ModernBERT does not show a reliable D2C-over-text advantage, which leaves open which encoder families best support dimension-supervised routing.

Target Audience

Researchers working on AI-generated text detection and authorship attribution will get the most from this paper, especially those already familiar with the HART taxonomy or the RACE line of work. It is also relevant to practitioners building provenance or content-governance systems who need to understand what low-FPR four-way detection can and cannot currently deliver, and to NLP researchers interested in multi-view architectures that supervise intermediate source factors before composing a final label. Readers without a background in transformer fine-tuning and detector evaluation metrics will find the protocol details demanding.

Authors’ abstract

AI-generated text detection is commonly framed as a binary document-level judgment about whether a text is human-written or machine-generated. This framing breaks down for mixed-origin writing, where content origin and expression origin may differ. We cast mixed-origin detection as dimension-to-composition source attribution, inferring content origin and expression origin before composing them into four collaboration types. We propose Dimension-to-Composition Routing (D2C-Routing), which routes content-side and expression-side evidence to supervised dimension heads before a learned gated composition layer predicts the final label. On MixD2C, a reconstructed split derived from the HART mixed-origin benchmark, our disclosed D2C-Routing-based detector system reaches 0.8603 four-way Avg TPR@1%FPR, 6.5 points above the same-split RACE-local rerun. Core ablations support the routing design, while error analysis shows that distinguishing AI-content/human-expression from fully AI-generated text remains the hardest boundary. Code is available at https://github.com/bystander563/d2c-routing-artifact.