MGCE: Multi-Granularity Conceptual Experts

• MGCE is a framework that employs multiple conceptual experts at varied semantic resolutions to dynamically mine and represent fine-grained categories.
• It integrates Dynamic Conceptual Contrastive Learning with Multi-Granularity Experts Collaborative Learning to extract and align features from fine to coarse levels.
• Empirical results demonstrate significant improvements in novel-class accuracy and robustness by leveraging expert specialization and cross-granularity coordination.

Searching arXiv for the cited MGCE-related papers to ground the article in current records. arXiv search query: "Multi-Granularity Conceptual Experts Generalized Fine-Grained Category Discovery (Zheng et al., 30 Sep 2025)" arXiv search query: "Multi-Granularity Representations of Dialog (Mehri et al., 2019)" Multi-Granularity Conceptual Experts (MGCE) denotes a class of architectures in which multiple experts are specialized to different levels of abstraction and then coordinated for representation learning, prediction, or both. In the strict sense, MGCE is the name of a framework for generalized fine-grained category discovery that combines Dynamic Conceptual Contrastive Learning (DCCL) with Multi-Granularity Experts Collaborative Learning (MECL), dynamically mines concepts at several semantic resolutions, and can automatically estimate the number of categories in unlabeled data (Zheng et al., 30 Sep 2025). Closely related work shows the same design principle in other domains: dialog retrieval via distance-conditioned negative sampling (Mehri et al., 2019), knowledge-graph concept extraction through overlapping span prediction and post-hoc selection (Yuan et al., 2022), accented speech correction through accent-specific LoRA experts and word/phoneme granularity (Mu et al., 12 Jul 2025), and remote sensing vision-language modeling through a granularity-oriented mixture of experts spanning image-, region-, and pixel-level tasks (Liu et al., 2024). Across these settings, MGCE is characterized by expert specialization, cross-granularity coordination, and the use of granularity itself as an explicit modeling variable.

1. Formal problem setting and definition

In generalized category discovery, the labeled dataset is

$$\mathcal{D}_{L} = \left\{ (\mathbf{x}^{l}_{i}, y^{l}_{i}) \mid (\mathbf{x}^{l}_{i}, y^{l}_{i}) \in \mathcal{X} \times \mathcal{Y}_{L} \right\}_{i=1}^{N},$$

while the unlabeled dataset is

$$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$

with

$\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$

The central difficulty is that the unlabeled pool contains both known and novel classes, and in realistic open-world settings the number of unlabeled categories is often unknown during training (Zheng et al., 30 Sep 2025).

MGCE addresses two limitations identified for prior generalized category discovery methods. First, prior methods were described as failing to exploit multi-granularity conceptual information in visual data, despite the fact that fine-grained recognition naturally contains coarse superclasses, specific subclasses, and intra-class sub-concepts. Second, many parametric approaches assume that the number of unlabeled categories is known, which is impractical in open-world use (Zheng et al., 30 Sep 2025).

Within that formulation, MGCE constructs concept experts at multiple granularity levels, alternates concept mining with dual-level representation learning, establishes collaboration between experts through a concept alignment matrix, and uses clustering structure to estimate the number of categories. The fine-, base-, and coarse-grained experts differ only in how local similarity graphs are constructed, so granularity is operationalized through graph topology rather than through a different backbone architecture (Zheng et al., 30 Sep 2025).

A recurrent clarification in the broader literature is that “multi-granularity” does not necessarily imply a single hierarchical encoder. In dialog retrieval, for example, the multi-granularity effect is induced by training separate models with different negative sampling distributions; the hierarchy exists at the ensemble level rather than inside one model (Mehri et al., 2019). This suggests a general MGCE principle: abstraction level can be controlled by training signal design as much as by architectural depth.

2. Core architecture of MGCE

The canonical MGCE framework comprises two modules: DCCL and MECL. DCCL builds a single concept expert through Dynamic Concept Generation (DCG) and Dual-Level Representation Learning (DRL). MECL extends that single-expert setting to multiple experts and adds cross-expert collaboration (Zheng et al., 30 Sep 2025).

The backbone in the reported visual instantiation is a ViT-B/16 encoder pre-trained with DINO, with only the last transformer block fine-tuned. Image features are denoted by $\mathbf{z}_i = \mathcal{E}(\mathbf{x}_i)$, instance-level contrastive features by $\mathbf{h}_i = \phi(\mathcal{E}(\mathbf{x}_i))$, and expert-specific concept features by $\mathbf{v}_i^r = \phi_C^r(\mathbf{z}_i)$ for expert $r \in \{1,2,3\}$ (Zheng et al., 30 Sep 2025).

Instance-level representation learning combines self-supervised and supervised contrastive terms with classification. The total instance-level objective is

$$\mathcal{L}^I = \mathcal{L}_{con} + \mathcal{L}_{cls}.$$

When the number of unlabeled categories is unknown, the classification loss reduces to labeled supervision; when it is known, an additional unsupervised classification objective from SimGCD is included (Zheng et al., 30 Sep 2025).

Dynamic Concept Generation constructs a similarity graph over all samples. Similarity is defined as

$$s_{ij}=[(\mathbf{z}_{i}/{\left \| \mathbf{z}_{i} \right \| })\bm{\cdot} (\mathbf{z}_{j}/\left \| \mathbf{z}_{j} \right \|) + 1]/2\in [0,1].$$

The adjacency matrix strengthens labeled same-class pairs with maximal edge weight and retains unlabeled connections only above a threshold $\delta$. A $$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$0-nearest-neighbor restriction then yields a local adjacency $$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$1, and Semi-Infomap partitions that graph into concepts $$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$2 (Zheng et al., 30 Sep 2025).

Concept prototypes are initialized as cluster centroids: $$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$3 Concept-level supervision is then applied through a prototype-based contrastive objective

$$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$4

and concept memory is updated by momentum: $$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$5 The full training objective is

$$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$6

where $$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$7 is the collaborative loss across experts (Zheng et al., 30 Sep 2025).

3. Granularity specialization and expert collaboration

MGCE operationalizes granularity through the neighborhood size $$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$8 used in graph construction. Small $$\mathcal{D}_{U} = \left\{ \mathbf{x}^{u}_{i} \mid \mathbf{x}^{u}_{i} \in \mathcal{X} \right\}_{i=1}^{M},$$9 yields sparse neighborhoods, more clusters, and finer granularity; large $\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$0 yields denser neighborhoods, fewer clusters, and coarser granularity (Zheng et al., 30 Sep 2025). The base expert uses an adaptively selected $\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$1, and the additional experts are defined by

$\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$2

with $\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$3 in the reported experiments (Zheng et al., 30 Sep 2025).

Cross-expert interaction is not implemented as simple voting. Instead, MGCE aligns expert-specific concept spaces through a concept alignment matrix

$\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$4

which measures cosine similarity between concept prototypes from expert $\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$5 and the base expert. Each expert produces a concept-similarity distribution over its own prototypes, and the fine- and coarse-level experts are projected into the base concept space via $\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$6. Consistency is then enforced with a symmetric KL divergence (Zheng et al., 30 Sep 2025).

This collaborative mechanism is one of the defining properties of MGCE. The experts are not independent ensemble members trained once and averaged at inference; they are coupled by shared instance-level learning, synchronized dynamic concept mining, and explicit alignment of concept spaces. In the authors’ interpretation, this allows the base expert to absorb both coarse stability and fine discrimination, which is especially important for subtle inter-class boundaries in fine-grained recognition (Zheng et al., 30 Sep 2025).

A frequent misconception is that MGCE is synonymous with fixed taxonomic supervision. In the reported formulation, the method does not require external hierarchical labels. Concepts are mined dynamically from the evolving feature space, and the number of concepts $\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$7 is allowed to change across epochs. The hierarchy is therefore emergent rather than pre-annotated (Zheng et al., 30 Sep 2025).

4. Cross-domain formulations of multi-granularity experts

The name MGCE is explicit in generalized fine-grained category discovery, but closely related formulations appear in multiple research areas. In some cases the correspondence is explicit in the technical commentary: the dialog paper’s Multi-Granularity Training “can be read as a concrete instantiation” of MGCE, and the knowledge-graph extraction paper “essentially describes what you could reinterpret as a multi-granularity conceptual expert system” (Mehri et al., 2019, Yuan et al., 2022).

Domain	Granularity or expert structure	Specialization mechanism
Fine-grained category discovery (Zheng et al., 30 Sep 2025)	Fine, base, and coarse experts	Different $\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$8, concept alignment, collaborative KL loss
Dialog retrieval (Mehri et al., 2019)	High-, intermediate-, and low-granularity dialog representations	Different negative sampling distance buckets and ensemble averaging
Knowledge-graph concept extraction (Yuan et al., 2022)	Coarse and fine concept spans, including overlaps	BERT-MRC span scoring, random forest selection, rule-based pruning
Accented speech correction (Mu et al., 12 Jul 2025)	Accent experts plus word/phoneme granularity	Mono-accent LoRA experts, HDMoLE routing, multi-granularity GER
Remote sensing VLM (Liu et al., 2024)	Image-, region-, and pixel-level understanding	Granularity-oriented mixture of experts

In dialog retrieval, the central claim is that the kind of negative examples a retrieval model sees during training determines what dialog information its latent representations specialize in. Training $\mathcal{Y}_{L} \subset \mathcal{Y}_{U}.$9 independent dual encoders on different bands of semantic distances between the ground-truth response and negatives yields experts specializing in lexical and utterance-level coherence, paraphrase-like variations, or global dialog properties such as domain, goal, and act type. Their predictions are combined by averaging probabilities (Mehri et al., 2019).

In knowledge-graph completion, MRC-CE uses a BERT-based machine reading comprehension model with a pointer network to extract multiple overlapping concept spans from entity descriptions, then refines those candidates with a random forest selector and rule-based pruning. Because the extractor can return “company,” “technology company,” and “multinational technology company” simultaneously, the system supports multi-granular conceptual predictions without enforcing non-overlap constraints (Yuan et al., 2022).

In accented speech recognition, the expert axis is partly domain-specific and partly granularity-specific. Mono-accent LoRA experts encode accent-specific pronunciation and lexical patterns, HDMoLE mixes those experts through global and local routing with dynamic thresholds, and a separate multi-granularity GER stage combines N-best word-level and phoneme-level hypotheses to generate the final transcription (Mu et al., 12 Jul 2025). A related formulation in remote sensing uses a granularity-oriented mixture of experts to support image-level, region-level, pixel-level, and multi-image tasks within an end-to-end RS VLM constrained to about 1 billion parameters (Liu et al., 2024).

Taken together, these systems support a broader interpretation of MGCE: experts can be indexed by abstraction level, by domain, or by both, provided that the training or routing mechanism forces persistent specialization and the inference procedure preserves complementary information across levels.

5. Empirical evidence and representative results

The most direct MGCE evidence comes from fine-grained category discovery. On small-scale fine-grained datasets, MGCE improves over the strongest parametric HypCD by 4.3% in All ACC and 6.3% in New ACC on average, and over the strongest non-parametric CMS by 3.5% in All ACC and 8.3% in New ACC. On large-scale fine-grained datasets, the corresponding gains are +6.5% All ACC and +10.9% New ACC over HypCD, and +10.0% All ACC and +15.8% New ACC over CMS. In the unknown-$\mathbf{z}_i = \mathcal{E}(\mathbf{x}_i)$0 setting, MGCE improves over PIM by +7.6% All ACC and +10.9% New ACC, over GPC by +13.0% All ACC and +14.9% New ACC, and over its own DCCL baseline by +12.3% All ACC and +11.2% New ACC. It also reduces average cluster-number error rate to 25% (Zheng et al., 30 Sep 2025).

Ablation results in the same framework localize the effect of each component. Adding concept-level learning to instance-level learning improves All ACC by 7.5% over DCCL and 2.2% over an instance-only variant; adding Multi-Expert Conceptual Learning yields a further 2.6% gain in All ACC; and adding collaborative learning gives an additional +0.9% All ACC and +1.2% New ACC. The reported CUB case study shows an early coarse expert merging three similar bird species and finer experts progressively separating them, after which collaborative learning aligns the experts toward the correct three-species partition (Zheng et al., 30 Sep 2025).

The dialog instantiation provides direct evidence that separate granularities really are learned. On MultiWOZ, the dual-encoder baseline obtains MRR 79.55 and H@1 66.13, whereas Multi-Granularity (5) obtains MRR 82.74 and H@1 72.18. On Ubuntu, Multi-Granularity (5) improves the dual encoder from MRR 76.84 to 80.10 and $\mathbf{z}_i = \mathcal{E}(\mathbf{x}_i)$1 from 63.6 to 68.7. Frozen-feature analysis shows that bag-of-words prediction improves monotonically toward higher granularity, peaking at BoW F1 59.43, while dialog-act prediction peaks at the highest abstraction with DA F1 19.24. That pattern demonstrates that different experts encode distinct levels of detail and abstraction (Mehri et al., 2019).

In knowledge-graph completion, MRC-CE on 500 CN-DBpedia test entities reports EC # 519, NC # 323, NC length 4.91, OC ratio 29.35%, Precision 92.22%, R-Recall 39.63%, and R-F1 55.44%. When applied to all entities in CN-DBpedia, it extracts 9,021,805 instanceOf relations, of which 7,053,900 are new, increases unique concepts to 894,689, and raises average extracted concept length to 4.92 characters, indicating finer granularity (Yuan et al., 2022).

In accented speech recognition, the best system—HDMoLE plus multi-granularity GER—reaches 2.07% average WER, compared with 6.34% for vanilla Whisper-large-v3, corresponding to a relative WER reduction of 67.35%. Ablations show that word-only N-best hypotheses outperform phoneme-only hypotheses, but the best result comes from combining word and phoneme N-best lists with $\mathbf{z}_i = \mathcal{E}(\mathbf{x}_i)$2, giving 2.07% WER (Mu et al., 12 Jul 2025).

These results collectively support a common empirical thesis: when expert specialization is tied to granularity rather than introduced as generic ensemble diversity, the resulting system tends to improve novel-class accuracy, transfer, fine-grained discrimination, or error correction, depending on the domain.

6. Limitations, misconceptions, and research directions

MGCE is not a single fixed architecture, and its limitations depend on the domain-specific instantiation. In fine-grained category discovery, the gains are strongest where rich multi-granularity structure is abundant; on generic datasets with less pronounced hierarchy or very large intra-class variation, improvements over strong baselines are smaller. The method also requires graph construction and clustering at every epoch, which is more computationally expensive than purely parametric alternatives, even with FAISS acceleration (Zheng et al., 30 Sep 2025).

In dialog retrieval, the limitations are explicit: combination is a simple equal-weight ensemble; granularity control is indirect because it is entirely mediated by distance-based negative sampling; interpretability remains limited; and only one axis of variation—semantic similarity of candidate responses—is used. The paper also notes that there is no learned gating or more sophisticated fusion of granularities (Mehri et al., 2019). A common misconception is therefore to treat any multi-model ensemble as MGCE; the dialog results specifically argue that diversity aligned with granularity is more useful than generic ensemble diversity.

In concept extraction for knowledge graphs, the main challenges concern explicit hierarchy modeling, ambiguity, scalability of span enumeration, and language-specific rules. MRC-CE extracts concept phrases but does not build or exploit a structured concept hierarchy, and its rule-based pruning is ontology- and language-dependent (Yuan et al., 2022). This suggests that a stricter MGCE formulation for KGs would require explicit hierarchy constraints rather than overlap handling alone.

In accented speech, limitations include coverage of only nine English accents, no explicit treatment of noise or channel variation, dependence on the quality of the accent-recognition router, and substantial stage-wise complexity due to connector pretraining, mono-accent expert specialization, HDMoLE training, and the separate multi-granularity GER stage (Mu et al., 12 Jul 2025). Here a plausible implication is that MGCE systems with both domain experts and granularity experts benefit from stronger routing supervision but also inherit the brittleness of their routers.

Across these lines of work, several future directions recur. One is richer routing, moving beyond uniform averaging toward learned gating or hierarchical composition. Another is extending expert structure beyond a single granularity axis to include domain, style, intent, hierarchy, or modality. A third is strengthening consistency constraints so that coarse and fine experts remain compatible rather than merely complementary. In the visual GCD setting, that role is played by the concept alignment matrix and symmetric KL loss; in speech, by hierarchical routing and dynamic thresholds; in KG extraction, by selector and pruning layers; and in remote sensing, by a granularity-oriented mixture of experts spanning image-, region-, and pixel-level reasoning (Zheng et al., 30 Sep 2025, Mu et al., 12 Jul 2025, Yuan et al., 2022, Liu et al., 2024).