Stylistics-Aware Router (SAR)

• SAR is a dynamic routing module that selects stylistically-homogeneous references through prototype-based clustering in latent semantic space.
• It improves AI-generated text detection accuracy by enabling adaptive conditional threshold calibration in heterogeneous and low-resource settings.
• SAR leverages unsupervised clustering and PCA-reduced deep embeddings to efficiently route inputs, reducing computational complexity.

The Stylistics-Aware Router (SAR) is a core component of the Mixture of Stylistic Experts (MoSEs) framework for uncertainty-aware AI-generated text detection. SAR addresses limitations of existing detection methods that either neglect stylistic modeling entirely or rely on static global references, which can impair detection accuracy, particularly in heterogeneous or low-resource settings. SAR efficiently routes each test input to a dynamically selected, stylistically-homogeneous subset of reference texts by leveraging unsupervised clustering in latent semantic space, thereby improving conditional threshold estimation and ultimately system robustness (Wu et al., 2 Sep 2025).

1. Functional Role within MoSEs

MoSEs integrates three primary modules: the Stylistics Reference Repository (SRR), the Stylistics-Aware Router (SAR), and the Conditional Threshold Estimator (CTE). SRR comprises a large, multi-style corpus of reference texts annotated with human/AI labels and associated conditional features (surface-level statistics, n-gram repetition, type-token ratio, and deep semantic embeddings). SAR acts as the intermediary, taking an input text’s semantic embedding and efficiently selecting a small, locally-homogeneous subset of reference samples from SRR. This subset is then supplied to the CTE, which utilizes it (in conjunction with the model’s discrimination score) to fit an input-specific threshold. The combination of these modules ensures that detection decisions are informed by stylistically-congruent, localized evidence, bridging the gap between global reference memory and fine-grained, per-example calibration (Wu et al., 2 Sep 2025).

2. Prototype-Based Semantic Clustering

SAR frames the routing problem as prototype-based clustering within each distinct “style” partition of SRR. For style $s$, the semantic embeddings of SRR samples are represented as $X^s \in \mathbb{R}^{d \times N^s}$. These are partitioned into $K$ clusters using relaxed-optimal-transport clustering with the Sinkhorn-Knopp algorithm ($\varepsilon = 0.05$), producing cluster centroids (“prototypes”) $$P^s = [p^s_1, ..., p^s_K] \in \mathbb{R}^{d \times K}$$ and assignment matrix $Q^s \in \{0,1\}^{K \times N^s}$, where $q^s_{k,i}=1$ if sample $i$ is assigned to prototype $k$. The clustering objective is

$$Q^{s*} = \operatorname{diag}(\alpha)\, \exp\left((P^s)^\top X^s / \varepsilon\right)\, \operatorname{diag}(\beta),$$

where $\alpha$, $\beta$ are scaling factors enforcing balanced clusters. Prototype representations may be periodically refined using momentum updates:

$$p^s_k \leftarrow \mu p^s_k + (1-\mu) \left(\frac{1}{|\mathcal{I}_k|} \sum_{i \in \mathcal{I}_k} X^s_{:,i}\right)\ ,$$

where $\mu$ is the momentum (e.g., 0.9) and $\mathcal{I}_k$ are indices assigned to prototype $k$.

This structure reduces reference retrieval from $\mathcal{O}(N)$ to $\mathcal{O}(SK)$ (with $S$ styles, $K$ prototypes per style), providing an efficient and manageable routing backbone (Wu et al., 2 Sep 2025).

3. Routing Algorithm and Computational Workflow

Given a test input $x$, SAR first encodes it into a deep semantic vector $e \in \mathbb{R}^d$ via a pre-trained encoder (such as BGE-M3). In practice, principal component analysis (PCA) is performed to reduce the embedding dimensionality (typically to 32) for computational efficiency.

SAR then computes the distance (typically Euclidean: $d(s,k) = \|e - p^s_k\|_2$) between $e$ and every prototype $p^s_k$ across all styles. The $m$ nearest prototypes are retrieved, their indices forming the set $M = \{(s_1, k_1), ..., (s_m, k_m)\}$. SAR then gathers all SRR samples assigned to those prototypes, constructing set $R = \bigcup_{(s,k)\in M} \{ i\,|\,q^s_{k,i}=1 \}$, which is returned for downstream threshold estimation by CTE.

Pseudocode for the routing logic is as follows:

def StylisticsAwareRoute(x, m): e = Encoder(x) # e ∈ ℝ^d D = [] # list of (s,k,dist) for s in range(1, S+1): for k in range(1, K+1): dist = np.linalg.norm(e - P^s[:,k]) # or 1–cosine similarity D.append((s, k, dist)) D.sort(key=lambda tup: tup[2]) # ascending by dist M = D[:m] # indices of m nearest prototypes R = set() for (s, k, _) in M: R.update([i for i, val in enumerate(Q^s[k,:]) if val==1]) return R

This selective, similarity-driven activation ensures that only stylistically relevant references inform decision thresholding, which is particularly advantageous in variable or sparse-data regimes (Wu et al., 2 Sep 2025).

4. Input Features and Unsupervised Prototype Learning

SAR operates exclusively on deep semantic embeddings obtained from pre-trained LLMs. These embeddings typically reside in high-dimensional space ($d \approx 1024$), but are compressed via PCA to 32 dimensions to optimize computational throughput. No surface-level statistics, n-gram counts, or related features are input to SAR; such signals are exploited downstream by the CTE only.

Prototype learning within SAM is fully unsupervised, relying solely on the structure in the embedding space within each style—without recourse to output labels or static reference assignments. Key hyperparameters include the number of prototypes per style ($K$, typically 10–50, chosen via cross-validation), the number of nearest prototypes to activate ($m$, e.g., 5–20, based on ablation), the specific distance metric (Euclidean or cosine), and the PCA-reduced embedding dimension (fixed at 32). There is no supervised loss on SAR itself; cluster structure is purely emergent from unsupervised latent geometry (Wu et al., 2 Sep 2025).

5. Empirical Evaluation and Impact

Ablation results demonstrate consistent gains when integrating SAR into the detection pipeline. The table below summarizes average performance on main datasets (“Lastde” model), with and without SAR activation:

Method	w.o. SAR	w. SAR
Static Threshold	0.8388	0.8488
Nearest Voting	0.8388	0.8525
MoSEs-lr	0.9250	0.9350
MoSEs-xg	0.9450	0.9475

These results indicate 1–1.5% absolute improvement across both naïve and advanced baselines, substantiating the value of stylistics-aware routing. In a selection-strategy ablation (Lastde average), activating all samples of a single predicted style yields 0.9350 accuracy, while the m-nearest-prototype strategy (as implemented by SAR) achieves 0.9475—confirming the benefit of finer-grained, prototype-based selection.

In the low-resource regime (200 reference texts), the SAR-equipped MoSEs-xg scores up to +39.15% improvement over static thresholds (notably on RoBERTa), highlighting its efficacy under data scarcity (Wu et al., 2 Sep 2025).

6. Significance and Applications

SAR enables MoSEs to exploit a large, diverse SRR by: (1) clustering stylistically heterogeneous references into compact prototypes, (2) efficiently retrieving the most relevant prototypes per input using unsupervised latent-space geometry, and (3) ensuring that only the nearest style-homogeneous references inform threshold adaptation for each test instance.

This methodology is particularly advantageous for scalable, uncertainty-aware detection of AI-generated texts, enabling robust generalization even when reference domains are broad or labeled data is limited. Furthermore, the unsupervised approach to prototype discovery provides prospect for extensibility to other detection and retrieval applications requiring dynamic, style- or domain-sensitive reference activation (Wu et al., 2 Sep 2025).