Semantic-Anchor Expert Routing

• Semantic-anchor-guided expert routing is a family of methods that uses semantic signals, such as foreground masks and latent embeddings, to guide expert selection in MoE architectures.
• The approach employs techniques like logit perturbation, distribution alignment, and similarity-based routing to ensure that the routing mechanism aligns with semantically meaningful structures.
• This routing strategy finds applications across vision, language, and multimodal tasks, yielding enhanced performance, interpretability, and targeted computational efficiency.

Semantic-anchor-guided expert routing denotes a class of mixture-of-experts routing strategies in which expert selection is shaped by semantic signals rather than left to unconstrained gating alone. In the literature, these signals range from explicit foreground masks, cue-derived concept directions, and multilingual routing priors to latent semantic embeddings, instruction-derived task descriptions, anatomical partitions, and anchor tokens that mark answer or action boundaries. Across vision, vision-language reasoning, multilingual LLMs, heterogeneous graphs, retrieval, person re-identification, referring image segmentation, EEG decoding, and test-time reasoning control, the common objective is to make routing more consistent with semantically meaningful structure while preserving the computational advantages of sparse or soft expert activation (Min et al., 24 May 2025, Zeng et al., 18 Apr 2026, Dong et al., 24 Jun 2026).

1. Conceptual scope and anchor types

The central design principle is to treat some semantic signal as an “anchor” that should influence how expert capacity is allocated. In some systems the anchor is explicit supervision, such as a foreground mask or an answer-cue vector. In others it is a latent representation already present in the model, such as a contextual graph embedding, a routing prior extracted from a high-resource language, or a recurring delimiter token whose surrounding routing state marks a reasoning phase. This yields a broader notion of semantic guidance than classic token-wise gating, because the router is no longer driven only by the instantaneous hidden state of a token or patch.

Paper	Semantic anchor	Routing role
"Guiding the Experts: Semantic Priors for Efficient and Focused MoE Routing" (Min et al., 24 May 2025)	Foreground mask $\mathcal{M}$ from Grounding DINO + SAM	Align dispatch weights with semantic foreground
"CoGR-MoE: Concept-Guided Expert Routing with Consistent Selection and Flexible Reasoning for Visual Question Answering" (Zeng et al., 18 Apr 2026)	Cue-derived semantic directions $s_a, s_j$	Bias teacher logits and reweight shared Top-$K$ experts
"SARA: Unlocking Multilingual Knowledge in Mixture-of-Experts via Semantically Anchored Routing Alignment" (Dong et al., 24 Jun 2026)	High-resource sequence-level routing priors	Align low-resource routing with symmetric JS divergence
"One Router to Route Them All: Homogeneous Expert Routing for Heterogeneous Graph Transformers" (Shakirov et al., 10 Nov 2025)	Latent semantics in HGT embeddings $\mathbf{h}_v$	Regularize routing away from deterministic type dependence
"RouterRetriever: Routing over a Mixture of Expert Embedding Models" (Lee et al., 2024)	Pilot centroid embeddings of oracle-best queries	Hard-route a query to the nearest expert family
"Glider: Global and Local Instruction-Driven Expert Router" (Li et al., 2024)	LLM-generated task instructions embedded as global vectors	Provide a query-level semantic prior for all layers
"Does the Same Token Mean the Same State? MoE Routing as Signal for Reasoning Control" (Chen et al., 22 Jun 2026)	Boundary and delimiter anchor tokens	Read routing neighborhoods to select a rollout

A recurring distinction is between explicit and implicit anchors. Explicit anchors are externally constructed and directly injected into routing or routing losses. Implicit anchors are extracted from the model’s own representations or trajectories and used either to regularize routing or to interpret it. This suggests that semantic-anchor-guided routing is not a single algorithmic template but a family of mechanisms organized around the same premise: semantically matched inputs should induce more appropriate, more stable, or more interpretable expert usage than purely unguided gating.

2. Core routing mechanisms and mathematical patterns

Three technical patterns recur in the literature. The first is logit perturbation by semantic directions. In CoGR-MoE, the teacher router injects a correct-answer semantic direction $s_a$ into the base logits,

$$g^{T} = \mathrm{softmax}\!\left( z_{\mathrm{base}} + \lambda_a\, s_{a} \right),$$

while a student router learns to mimic this semantic guidance through a KL distillation loss (Zeng et al., 18 Apr 2026). A related pattern appears in GLIDER, where cosine similarity between a query-level semantic embedding and expert-level semantic embeddings yields a global routing score that is added to token-level local scores; the global term is sharply amplified when the maximum similarity exceeds $0.8$ (Li et al., 2024).

The second pattern is routing-distribution alignment to an anchor distribution. In semantic-guided Soft MoE for vision, the router is not constrained at the level of discrete expert identities but through the spatial structure of dispatch weights. A token-level aggregate

$$\mathbf{W}_i = \frac{1}{N\times p} \sum_{j=1}^{N\times p} \mathbf{D}_{i,j}$$

is thresholded into a dispatch mask and aligned to a foreground mask with an importance-weighted IoU-like score, yielding

$$\mathcal{L}_\text{total} = \mathcal{L}_\text{cls} + \lambda \mathcal{L}_\text{aux},$$

with $\lambda=0.01$ performing best in the reported experiments (Min et al., 24 May 2025). In SARA, the aligned object is the dense routing distribution itself: sequence-level routing priors from high-resource anchor-language traces are matched to low-resource counterparts with a symmetric Jensen–Shannon penalty over selected intermediate layers (Dong et al., 24 Jun 2026).

The third pattern is similarity-based routing against semantic prototypes. RouterRetriever is a non-parametric example: for each expert, base-encoder embeddings of queries on which that expert is oracle-best are averaged into pilot centroids, and a new query is routed to the expert whose pilot set has the highest average cosine similarity to the query embedding (Lee et al., 2024). This is semantically guided in the sense that the centroids summarize embedding-space regions where a given expert is empirically optimal.

A more interpretive but technically important pattern treats routing itself as a semantic signal. In the control-space decomposition proposed in "Polysemantic Experts, Monosemantic Paths: Routing as Control in MoEs" (Ye et al., 20 Apr 2026), the hidden state at layer $s_a, s_j$0 is split into a router-visible control component and an orthogonal router-blind content component through the row space of the routing matrix. Because routing logits depend only on the control component, semantic steering can in principle target that subspace without directly altering the content channel. By contrast, "The Myth of Expert Specialization in MoEs" argues that all such effects ultimately reduce to the geometry of hidden-state space under a linear router, so semantic anchors must be understood as geometric constraints on $s_a, s_j$1 and $s_a, s_j$2, not as intrinsic semantic labels attached to experts (Wang et al., 10 Apr 2026).

3. Vision and multimodal perception

In vision classification, semantic-anchor guidance has been implemented most explicitly in Soft MoE. The key observation is that dispatch weights in Soft MoE form segmentation-like patterns even when the model is trained only for classification. The proposed intervention is to treat class-conditioned foreground masks as anchors and align high-dispatch regions with them at the last Soft MoE layer, while introducing a LayerScale-modified residual

$s_a, s_j$3

to prevent residual dominance from diluting the auxiliary supervision (Min et al., 24 May 2025). On ImageNet-1K, this raises the Soft MoE baseline from 73.9% Top-1 and 91.5% Top-5 to 74.5% and 91.8%; on ImageNet-100, the method improves 75.4 / 92.4 to 76.8 / 92.8. The same study reports faster convergence, more interpretable dispatch maps, and stronger gains when semantic guidance is used in both pretraining and fine-tuning.

Referring image segmentation provides a second vision-language instantiation. SERA inserts expression-aware expert refinement in two places: a soft-routed SERA-Adapter inside selected DINOv2 blocks and a Top-$s_a, s_j$4 SERA-Fusion module over geometry-preserving experts for spatial, context, boundary, and shape refinement (Dalaq et al., 13 Mar 2026). Here the semantic anchors are the expression embeddings and cross-attended visual tokens, while routing is driven by global descriptors derived from language-conditioned visual features. In the reported parameter-efficient setting, SERA updates less than 1% of the backbone parameters and reaches average mIoU 71.1, compared with 70.4 for DETRIS-B and 62.8 for ETRIS.

In Any-Time Person Re-Identification, STFER uses LVLM-generated identity-consistent text as the anchor. The text is designed to capture near-invariant biometric constants rather than clothing-dependent cues, and its mean-pooled embedding is concatenated with scenario-specific CLS features inside a Semantic-driven Expert Routing module,

$s_a, s_j$5

with stochastic masking of the semantic vector to avoid over-reliance on text (Li et al., 16 Apr 2026). The full system combines semantic-driven visual token filtering and semantic-driven routing, improving Any-Time R1 from 55.80 for the baseline to 94.54 on AT-USTC, with consistent additional gains over the text-only variant.

BrainStack extends the same logic to EEG-based language decoding. Its experts are anatomically anchored regional subnetworks—Prefrontal, Frontal, Central, Left-Temporal, Right-Temporal, Parietal, and Occipital—together with a global CTNet expert. A softmax gate computes

$s_a, s_j$6

over expert outputs, and a global-to-local distillation term aligns regional experts with the global expert’s semantic space (Zhao et al., 29 Jan 2026). This functionally anchored routing raises average word classification accuracy on SS-EEG to 41.87%, compared with 29.50% for TCNet and 35.50% for the global-only CTNet.

4. Language, retrieval, graphs, and multilingual transfer

Semantic-anchor-guided routing in language reasoning is exemplified by CoGR-MoE for multiple-choice VQA. The anchor is not the image-question representation alone but concept information derived from answer options. Positive and negative cues are generated offline by an LLM, projected into routing space as

$s_a, s_j$7

and injected into teacher logits to stabilize expert selection for the correct concept while preserving option-level flexibility through option-aware reweighting inside a shared Top-$s_a, s_j$8 expert set (Zeng et al., 18 Apr 2026). The model reports VQAv2 88.5, GQA 83.2, VizWiz 84.8, MMStar 52.0, and on MRAG-Bench improves a MoE-LLaVA backbone from approximately 62.76 to 63.25 overall and a Qwen3-VL-30B backbone from 66.78 to 68.96.

In multilingual sparse LLMs, SARA uses high-resource language routing traces themselves as semantic anchors. For each correctly solved English or Chinese interaction, it extracts dense sequence-level routing priors and aligns low-resource translations toward them with a symmetric JS divergence over the model’s intermediate “semantic valley” layers (Dong et al., 24 Jun 2026). This yields average gains such as 70.73 to 71.53 on Global-MMLU for Qwen3-30B-A3B with English anchors, and 41.63 to 42.80 on Phi-3.5-MoE-instruct, with the paper emphasizing that self-anchoring from the model’s own routing priors is much more effective than external anchors.

Heterogeneous graphs pose a different problem: type labels can dominate routing. HER addresses this by using one shared expert pool for all node types and stochastically masking type embeddings so that routing cannot collapse into a deterministic type-to-expert map (Shakirov et al., 10 Nov 2025). The anchors are the latent semantic patterns already encoded in HGT embeddings rather than explicit semantic labels. On IMDB, SharedMoE with $s_a, s_j$9 reaches ROC-AUC 92.63, compared with 87.57 for HGT and 90.34 for a type-separated MoE baseline, and the resulting experts align more with genres than with node types.

In information retrieval, RouterRetriever operationalizes semantic anchors as pilot centroid embeddings summarizing regions in base-encoder space where each domain-specific retriever is oracle-best (Lee et al., 2024). Routing is hard and query-level rather than token-level, but the principle is identical: route the query toward the expert whose semantic prototypes are closest. On the seven-dataset BEIR subset reported in the paper, RouterRetriever reaches 49.6 nDCG@10, compared with 47.5 for a single MSMARCO-trained retriever and 46.4 for a multi-task model.

GLIDER pushes this idea further by explicitly generating semantic task instructions for both experts and incoming queries. Expert-specific LoRA modules are paired with local token-level routing vectors and with global instruction embeddings produced by prompting an external LLM on three input-output pairs per task (Li et al., 2024). The resulting multi-scale router sharply improves held-in routing: on T0 Held-In, GLIDER reaches 68.04 versus 61.42 for Phatgoose, while also improving T0 Held-Out from 56.90 to 57.78. The reported heatmaps indicate that for held-in tasks the global semantic prior strongly recovers the oracle expert, whereas for held-out tasks it biases the system toward semantically adjacent experts such as causal-reasoning tasks for COPA.

5. Interpretability, context sensitivity, and path-level structure

Several probing studies show that modern MoE routing does respond to semantics and context, but not always at the granularity of individual experts. "Probing Semantic Routing in Large Mixture-of-Expert Models" compares expert overlap under tightly controlled lexical conditions and finds statistically significant semantic effects across large open MoEs (Olson et al., 15 Feb 2025). In the WiC setup, same-sense pairs exhibit higher overlap than different-sense pairs with $K$0; in SWORDS, semantically equivalent substitutions yield higher overlap than semantically different substitutions with $K$1. The effect is strongest in middle layers.

A complementary encoder-decoder study reaches a similar conclusion with Switch Transformers. In the encoder, routing similarity correlates more with WordSim than with SimLex and becomes substantially more aligned with SCWS judgments when context is present, with average correlations rising from 0.2773 without context to 0.3785 with context; WiC effect sizes are strongest in middle encoder layers (Arnold et al., 2024). The decoder is markedly less context-sensitive. This suggests that semantic-anchor-guided routing is naturally more compatible with encoder or mid-layer routing than with late autoregressive decoding.

Path-based analyses refine this picture. "Polysemantic Experts, Monosemantic Paths" argues that individual experts remain polysemantic, while trajectories through experts become monosemantic because the low-bandwidth control signal rotates across layers (Ye et al., 20 Apr 2026). Clustering in the control channel yields approximately 4.3 unique token IDs per cluster, compared with about 1.1 in the router-blind content channel and 1.4 in the full representation, indicating that control-space clusters group different surface forms by semantic function. The same token, such as “:”, follows distinct expert trajectories when used as a type annotation, an introductory colon, or a time separator.

RoE provides a related but architectural perspective: a dense MLLM can be treated as a depth-wise mixture where each layer chooses between the original block and a lightweight adapter, with routing driven by special router tokens that summarize image and question context (Wu et al., 2024). The paper does not define semantic anchors explicitly, but it shows question-dependent and difficulty-dependent route patterns, and reports an average performance gain of 3.3% on 5 benchmarks while being faster than MoE-LLaVA. This suggests a latent-anchor interpretation in which joint multimodal context determines an example-specific compute path.

The most explicit use of semantic anchors for reasoning control appears in RAD. Here the anchors are not external labels but recurring delimiter or boundary tokens, such as \boxed{, code fences, or the thinking-to-final transition (Chen et al., 22 Jun 2026). Routing vectors in a fixed window around these anchors form neighborhoods that align with final-answer basins, and selecting the densest routing neighborhood yields answer-string-free rollout selection. Across 10 sparse-MoE configurations on six datasets, RAD reaches 73.9 and RAD+DC 74.2, compared with Majority 73.6 on well-posed answer-string tasks, while also improving best-of-16 patch selection on SWE-bench Verified over random when no canonical answer string exists.

6. Limitations, debates, and future directions

The literature is not uniform in how strongly it interprets routing as semantic specialization. Many application papers report that explicit semantic priors improve accuracy, convergence, or interpretability, but several also note dependencies that limit generalization. Vision methods depend on mask quality and offline preprocessing through systems such as Grounding DINO and SAM; SARA depends on stable high-resource anchors and translation quality; GLIDER depends on the quality and cost of external LLM-generated task descriptions; RAD depends on reliable anchor tokens and is explicitly “not a verifier” because a dense wrong basin can still win (Min et al., 24 May 2025, Dong et al., 24 Jun 2026, Li et al., 2024, Chen et al., 22 Jun 2026).

A deeper controversy concerns whether experts themselves are semantically meaningful. "The Myth of Expert Specialization in MoEs" argues that routing is a linear readout of hidden-state geometry and that specialization is an emergent property of representation space rather than of the routing architecture (Wang et al., 10 Apr 2026). It reports that expert overlap between different models answering the same question is approximately 60%, no higher than between entirely different questions; prompt-level routing does not predict rollout-level routing; and deeper layers can exhibit nearly identical expert activation across semantically unrelated inputs during prefilling. On this account, semantic-anchor-guided routing is still possible, but only by shaping hidden-state geometry and the router’s linear partition of that geometry.

A plausible implication is that the strongest future work will combine explicit anchors with mechanistic accounts of routing geometry. Several directions are already suggested in the surveyed papers: extending semantic-guided routing from classification to detection and segmentation, using self-supervised or attention-derived anchors instead of external segmentors, incorporating richer multi-anchor structures for multi-object scenes or multilingual parallelism, aligning internal circuits beyond routers, and using anchor-local routing states as a control interface for verifier-free selection or intervention (Min et al., 24 May 2025, Dong et al., 24 Jun 2026, Ye et al., 20 Apr 2026). The surveyed results collectively indicate that semantic anchors can make routing more focused, more transferable, and more interpretable, but they also indicate that semantic control in MoEs remains inseparable from the geometry, depth dependence, and training dynamics of the underlying model.