Minimizing Mismatch Risk: A Prototype-Based Routing Framework for Zero-shot LLM-generated Text Detection

TL;DR

This work identifies a fundamental mismatch risk in zero-shot LLM-generated text detection when using a fixed surrogate across diverse sources. It introduces DetectRouter, a two-stage prototype-based routing framework that learns text-detector affinity and aligns distributions to route inputs to the most compatible detector, formalized by a mismatch bound $|\,\mu^* - \mu_{proxy}\,| \le B \cdot \sqrt{2 D_{KL}(P_{src} \| P_{sur})}$. The approach combines discriminative prototype construction from white-box models with distributional alignment to black-box sources, achieving state-of-the-art AUROC on EvoBench (90.85%) and MAGE (77.92%) and providing universal enhancement across six zero-shot detection criteria. This has practical impact for robust, scalable detection in real-world deployments where source models are unknown or evolving, by enabling adaptive, evidence-based routing to complementary detectors.

Abstract

Zero-shot methods detect LLM-generated text by computing statistical signatures using a surrogate model. Existing approaches typically employ a fixed surrogate for all inputs regardless of the unknown source. We systematically examine this design and find that detection performance varies substantially depending on surrogate-source alignment. We observe that while no single surrogate achieves optimal performance universally, a well-matched surrogate typically exists within a diverse pool for any given input. This finding transforms robust detection into a routing problem: selecting the most appropriate surrogate for each input. We propose DetectRouter, a prototype-based framework that learns text-detector affinity through two-stage training. The first stage constructs discriminative prototypes from white-box models; the second generalizes to black-box sources by aligning geometric distances with observed detection scores. Experiments on EvoBench and MAGE benchmarks demonstrate consistent improvements across multiple detection criteria and model families.

Figures (9)

• Figure 1: Detection variance of Fast-DetectGPT across nine black-box LLM families on the MIRAGE benchmark. Each bar spans from the minimum to maximum AUROC achieved by different surrogate models, with circles indicating the median. The percentage on the right denotes the performance gap between the best and worst surrogate for each source.
• Figure 2: Affinity matrix of detection performance using Fast-DetectGPT. Nine open-source models serve as both generators (rows) and surrogates (columns). Diagonal dominance confirms optimal performance requires source-surrogate alignment, while off-diagonal patterns reveal structured affinity governed by architectural similarity.
• Figure 3: Overview of DetectRouter. Stage 1 constructs discriminative prototypes from white-box models via multi-task generations and distance-based classification. Stage 2 generalizes to black-box sources by aligning geometric distances with detection score distributions. At inference, the router selects the optimal detector based on nearest-prototype affinity.
• Figure 4: Performance improvement from applying DetectRouter to each zero-shot detection method. Bars show average AUROC on EvoBench and MAGE; numbers indicate relative improvement.
• Figure 5: Distribution of selected surrogates for each source model family, confirming source-specific routing preferences.

Theorems & Definitions (6)

• Proposition 3.1: Mismatch Risk Bound
• Definition 1.1: Total Variation Distance
• Definition 1.2: Kullback-Leibler Divergence
• Lemma 1.3: Pinsker's Inequality
• proof
• Corollary 1.4: Finite Sample Bound