Classifier-based Domain Router (C-DR)

• C-DR is a supervised routing module that predicts domain distributions and directs input representations to specialized experts in diverse ASR settings.
• It employs both hard and soft routing strategies with a coarse-to-fine hierarchy to manage ambiguous domain boundaries and improve overall performance.
• The design integrates entropy-aware routing and staged training to enhance interpretability, robustness, and efficient expert selection across heterogeneous data.

Classifier-based Domain Router (C-DR) denotes an explicit, supervised routing module that predicts a distribution over domains from an input representation and uses that distribution to select or mix domain-specific experts. In the multi-domain child–adult ASR setting, C-DR is introduced as the classifier-driven router that replaces the usual learned MoE gate, with the stated goals of making routing both controllable and interpretable in heterogeneous domains whose boundaries are not equally sharp (Shi et al., 9 Jun 2026). Related work shows that the same general pattern—routing by supervised domain prediction rather than unguided latent gating—also appears in reward modeling and multi-domain reranking, although those systems differ in whether routing is external, internal, or derived from the backbone’s own latent space (Namgoong et al., 2024, Wang et al., 25 Nov 2025). At the same time, broader router-LLM studies indicate that classifier-like routers can be brittle, category-driven, and sensitive to superficial cues, making robustness evaluation integral to any serious use of C-DR-style designs (Kassem et al., 20 Mar 2025).

1. Definition and motivation

In the Speech-LLM framework for unified ASR across adult and child speech, C-DR is the explicit, supervised routing module that maps an utterance’s speech representation to one of several domain-specific experts (Shi et al., 9 Jun 2026). Its role is to choose among experts in a setting where domain diversity is structured rather than flat: some distinctions are coarse and easy to separate, such as adult vs. child or different recording environments, while others are finer and more ambiguous, such as nearby child age groups within OGI-S. The authors contrast this with vanilla-routing MoE baselines that use learned gating networks and note that such gates lack explicit domain supervision and may not route reliably, especially with limited data (Shi et al., 9 Jun 2026).

This supervised routing logic is not unique to ASR. In lightweight reward modeling, an external router is used to select among multiple domain-specific reward models, again motivated by the observation that a single model trained on mixed-domain data may fail to represent domain-specific preferences well and may require retraining from scratch when a new domain arrives (Namgoong et al., 2024). In multi-domain decoder-only reranking, the same high-level idea appears in the form of a domain classifier that predicts a domain label from a query and selects a domain-specific expert reranker, although that work emphasizes routing from the reranker’s own latent state rather than from external features (Wang et al., 25 Nov 2025).

A broader formalization appears in router-LLM analysis, where routing is framed as an $N$-way classification problem over a set of models $M=\{M_1,\dots,M_N\}$, with

$R: Q \times M_N \to \{1,\dots,N\},$

and answer selection

$a = M_{R(q)}(q).$

That paper states that the term “classifier” is used broadly as “any method that decides which LLM to utilize for the given input query” (Kassem et al., 20 Mar 2025). This suggests that C-DR can be understood as a domain-specialized instance of a larger class of classifier-based routing mechanisms.

2. Mathematical formulation and hierarchical routing

The ASR C-DR paper describes the Speech-LLM backbone as

$$h = \text{Encoder}(s), \quad e^s = \text{Projector}(h), \quad e^p = \text{Emb}(\text{Tokenizer}(prompt)), \quad \hat{t} = \text{LLM}(e^p, e^s),$$

where $h$ is the encoder representation of the input utterance (Shi et al., 9 Jun 2026). C-DR predicts a routing distribution

$p(d \mid h),$

with $d \in \{1,\dots,n\}$ indexing the domain (Shi et al., 9 Jun 2026).

The paper distinguishes hard and soft routing. Hard routing selects the most probable domain,

$d^{*} = \arg\max_d p(d \mid h),$

and uses the expert output of the selected domain. Soft routing instead averages expert outputs by routing probabilities:

$$\hat{z}_o^{\text{routing}} = \sum_{d=1}^{n} p(d \mid h)\, \hat{z}_{o}^{d}.$$

This difference matters because the framework later introduces an uncertainty-sensitive interpolation mechanism built on the soft routing distribution (Shi et al., 9 Jun 2026).

A defining feature of C-DR in this formulation is its coarse-to-fine hierarchy. A coarse-grained classifier first predicts high-level domains; for inputs assigned to coarse domains that still contain substructure, a fine-grained classifier then predicts the sub-domain; and the final routing probability over domains is formed by combining the coarse and fine classifier outputs (Shi et al., 9 Jun 2026). The hierarchy is concrete in the dataset design: coarse domains include MyST, OGI-S, and Libri-Clean, while fine domains within OGI-S are the three age groups 4–7, 8–10, and 11–15 (Shi et al., 9 Jun 2026).

The classifier design uses weighted-sum encoder representations rather than only the top encoder layer. Both the coarse and fine classifiers are linear models operating on a weighted sum of multi-layer encoder representations, with learnable layer weights, followed by softmax normalization (Shi et al., 9 Jun 2026). This is a central design choice because the experiments explicitly compare top-layer encoder features, weighted-sum features with single-stage classification, and weighted-sum features with coarse-to-fine classification.

3. Expert selection, uncertainty handling, and staged training

C-DR does not compute the ASR output itself; it determines how the model should select or mix the experts in two expert banks: a Mixture-of-Projectors (MoP) and a Mixture-of-LoRAs (MoL) (Shi et al., 9 Jun 2026). Each domain is associated with a projector–LoRA pair, and the routing probabilities $M=\{M_1,\dots,M_N\}$0 are used to route the input to the appropriate projector expert in MoP and the appropriate LoRA expert in MoL. The outputs are then combined in the Speech-LLM stack (Shi et al., 9 Jun 2026).

To address uncertain routing near domain boundaries, the same framework adds Entropy-Aware Routing (EAR). EAR uses the normalized entropy of the routing distribution,

$M=\{M_1,\dots,M_N\}$1

where higher entropy indicates lower routing confidence (Shi et al., 9 Jun 2026). The routed output is then interpolated with a shared expert:

$M=\{M_1,\dots,M_N\}$2

The intended behavior is explicit: low entropy means the system trusts the domain expert routing more, while high entropy means it mixes in more of the shared expert (Shi et al., 9 Jun 2026).

The shared expert is a separate projector+LoRA pair trained on aggregated multi-domain data; in the reported implementation it is trained on the full OGI-S training set and EAR is applied only to OGI-S age groups, with entropy computed over the three age-group probabilities (Shi et al., 9 Jun 2026). The paper states that routing is most certain for 11–15, younger age groups show more uncertainty, and the 4–7 group benefits strongly from EAR (Shi et al., 9 Jun 2026).

Training is staged rather than end-to-end. The C-DR is trained separately from the MoE Speech-LLM, the encoder is frozen, the experts are trained separately, and the shared expert is trained separately on aggregated OGI-S data (Shi et al., 9 Jun 2026). The router is trained on the training sets of MyST, OGI-S, and LibriSpeech train-clean-100 using the proposed coarse-to-fine classification strategy, while the full system is not trained jointly with the ASR model in that paper (Shi et al., 9 Jun 2026).

4. Empirical behavior in multi-domain child–adult ASR

The reported experiments use Canary-Qwen-2.5B as the backbone and define five domains: MyST, OGI-S 4–7, OGI-S 8–10, OGI-S 11–15, and Libri-Clean (Shi et al., 9 Jun 2026). The projector and LoRA parameters of Canary-Qwen are replicated to initialize five experts; MyST and OGI-S age-group experts are fine-tuned with ground-truth hard routing, while the adult expert is kept fixed (Shi et al., 9 Jun 2026). Router training uses AdamW with peak learning rate $M=\{M_1,\dots,M_N\}$3, cosine warmup-decay, 500 warmup steps, 60,000 total steps, batch size 2, gradient accumulation 8, and a single NVIDIA A6000 GPU (Shi et al., 9 Jun 2026).

Several routing-focused findings are emphasized. Figure 1 compares top-layer encoder features, weighted-sum features with single-stage classification, and weighted-sum features with coarse-to-fine classification. The reported findings are that dataset-level classification is relatively easy, age-group classification within OGI-S is harder, weighted-sum features outperform top-layer features, and coarse-to-fine classification improves further, especially for OGI-S (Shi et al., 9 Jun 2026).

The paper also reports that better routing improves ASR. In Table 1, C-DR MoE consistently outperforms baselines, and the weighted-layer classifier with coarse-to-fine strategy is the best hard-routing variant; it achieves the best non-upper-bound result on MyST without hurting Libri-Clean (Shi et al., 9 Jun 2026). Under soft routing, adding EAR improves OGI-S performance further and achieves the best OGI-S result among the compared methods (Shi et al., 9 Jun 2026). A particularly notable observation is that, for the 4–7 age group, routing based on classifier-predicted probabilities can outperform routing based on ground-truth age labels; the authors interpret this as evidence that pronunciation development does not align perfectly with chronological age, so the classifier’s acoustic estimate may be more useful than the nominal label (Shi et al., 9 Jun 2026).

The paper presents these results as evidence that classifier-based, supervised routing can materially outperform unguided gating in a domain structure with both coarse separations and ambiguous fine boundaries (Shi et al., 9 Jun 2026).

5. Related router designs in reranking and reward modeling

A closely related formulation appears in multi-domain decoder-only reranking. That work states that the method is compatible with the C-DR view because it predicts a domain label $M=\{M_1,\dots,M_N\}$4 from a query $M=\{M_1,\dots,M_N\}$5, uses a classifier-like head

$M=\{M_1,\dots,M_N\}$6

selects a domain expert via argmax over domain probabilities, and uses the selected domain to determine which expert reranker is used (Wang et al., 25 Nov 2025). The difference is that the classifier is not built on external features: the Latent Semantic Router reads the frozen reranker’s own internal query representation $M=\{M_1,\dots,M_N\}$7, described as “a high-dimensional summary of the query’s semantic intent,” and dynamically activates the corresponding LoRA module $M=\{M_1,\dots,M_N\}$8 for reranking (Wang et al., 25 Nov 2025). In the reported aggregated cross-domain setup, the Latent Semantic Router achieves Acc 97.4 and Macro F1 97.3, compared with Acc 84.3 and Macro F1 82.2 for the standalone MLP router and Acc 97.3 and Macro F1 97.3 for the LLM-as-router; its extra parameter cost is 0.2B, versus 6.0B for the MLP router and 685B for the LLM-as-router (Wang et al., 25 Nov 2025). The paper explicitly contrasts this with a standard classifier-based router and presents the latent-semantic version as a tighter coupling between routing and the reranker’s internal semantic space.

Another related line appears in lightweight reward modeling. The RODOS architecture uses multiple domain-specific reward models and an external router that predicts which domain model should be used for a given prompt (Namgoong et al., 2024). Router training is classification-based over five domains—Anthropic, SHP, HellaSwag, Dahoas, and Oasst—and inference proceeds by having the router predict the domain and selecting the corresponding reward model (Namgoong et al., 2024). ARLISS preserves the external routing idea but stores one shared small LLM backbone plus multiple reward adapters and a router adapter, dynamically switching adapters at inference time through PEFT/LoRA (Namgoong et al., 2024). In the reported averages, Baseline DeB$M=\{M_1,\dots,M_N\}$9 achieves 0.6972, Baseline DeB$R: Q \times M_N \to \{1,\dots,N\},$0 0.6948, Base$R: Q \times M_N \to \{1,\dots,N\},$1 0.6914, MoRE DeB$R: Q \times M_N \to \{1,\dots,N\},$2 0.6961, RODOS DeB$R: Q \times M_N \to \{1,\dots,N\},$3 0.7003, and ARLISS DeB$R: Q \times M_N \to \{1,\dots,N\},$4 0.6975; RODOS has the best average accuracy, while ARLISS is lighter in parameter footprint than storing multiple full domain-specific models (Namgoong et al., 2024).

Taken together, these systems indicate that C-DR is not a single architecture but a family of supervised routing patterns. Some versions are explicit external routers over domain-specific models, some are hierarchical classifiers over expert banks, and some refine the same logic by probing the backbone’s latent space rather than adding an external classifier (Namgoong et al., 2024, Wang et al., 25 Nov 2025).

6. Robustness, fragility, and evaluation concerns

Work on router-LLM robustness raises several cautions that apply directly to classifier-based routing mechanisms. In that analysis, routing is formalized as selecting among models to optimize

$R: Q \times M_N \to \{1,\dots,N\},$5

where $R: Q \times M_N \to \{1,\dots,N\},$6 is routing quality, $R: Q \times M_N \to \{1,\dots,N\},$7 is cost, and $R: Q \times M_N \to \{1,\dots,N\},$8 is the quality–cost tradeoff parameter (Kassem et al., 20 Mar 2025). The benchmark introduced there, DSC, is designed around diverse, simple, and categorized evaluation, precisely because aggregate scores can conceal systematic routing biases.

The core empirical claim is that current routers are often not truly routing by difficulty, but instead behave like category-based classifiers with hardwired biases (Kassem et al., 20 Mar 2025). The paper states that a BERT-based router directs all coding and mathematics queries to the most powerful LLM—even when simpler models would suffice—while routing jailbreaking attempts to weaker models, thereby elevating safety risks. The reported routing proportions for the BERT router include MT-Bench Math 90%, MT-Bench Code 75%, SimpleCode 100%, LeetCode 100%, SimpleMath 98%, WildTrans 93.8%, and AdvBench 8% to the strong model (Kassem et al., 20 Mar 2025). The same study reports an average flipping rate of 98% when adding math/code keywords, indicating extreme sensitivity to lexical cues, and notes that threshold changes can shift math routing from 100% to 0% strong-model usage (Kassem et al., 20 Mar 2025).

Safety findings are equally central. The paper evaluates harmful prompts with Attack Success Rate (ASR) measured by LLM-as-a-judge and reports that Mistral-7B has 100% ASR on all attacks, while GPT-4o has 0% ASR on plain-text and GCG but 60% ASR on PAP (Kassem et al., 20 Mar 2025). Because most AdvBench prompts are routed to the weak model, the routing policy can increase attack success by selecting a less safe model (Kassem et al., 20 Mar 2025). Commercial routers are reported to face similar limitations.

For C-DR specifically, these findings do not invalidate supervised domain routing, but they do show that classifier-based routing should not be assumed to be semantic, calibrated, or robust merely because it is trained with labels. A plausible implication is that C-DR systems require categorized benchmarks, simple-task evaluation, privacy and safety tests, and audits for lexical sensitivity and domain-label ambiguity, especially when routing decisions control access to specialized or safety-critical experts (Kassem et al., 20 Mar 2025).