Concept Granularity: Levels of Detail

• Concept granularity is the level of detail at which a domain is represented, controlling how much structure is exposed or compressed in various contexts.
• Researchers formalize granularity using quantitative measures like hyperbolic embedding and activation metrics, alongside hierarchical and qualitative categorizations.
• Applications span adaptive proof presentation, multilingual modeling, fine-grained visual search, and medical diagnosis, balancing coarse and fine elements for improved task performance.

Concept granularity denotes the level of abstraction or detail at which a domain is represented, queried, explained, or manipulated. In the literature, the term appears in several technically distinct but structurally related senses: as the fineness of proof steps in automated proof presentation, as the shift from token-level to sentence-level semantic units in concept modeling, as the hierarchical specificity of linguistic expressions, as the distinction between coarse and fine visual concepts, and as the resolution at which spatial, medical, or rough-set concepts are organized (0903.0314, Musacchio et al., 24 May 2026, Ellinger et al., 26 May 2026, Zhong et al., 17 Apr 2026). Taken together, these works suggest that concept granularity is best understood not as a single formal primitive, but as a family of mechanisms for controlling how much structure is exposed, compressed, or ignored in a representation.

1. Domain-general meaning

In language and discourse analysis, granularity is “the level of abstraction at which entities or events are represented in language,” with fine-grained expressions denoting highly specific referents and coarse-grained expressions denoting more abstract or generic ones (Ellinger et al., 26 May 2026). In proof presentation, the same basic distinction appears as the difference between fine-grained proofs, which expose many low-level inferences, and coarse-grained proofs, which bundle those inferences into larger conceptual steps adapted to a reader’s background and the didactic goal (0903.0314). In multilingual concept modeling, the shift is even more explicit: standard LLMs operate at token granularity, whereas Mimir treats “concepts as sentences” and performs next-concept prediction in a sentence-embedding space (Musacchio et al., 24 May 2026).

In vision, granularity often describes the mismatch between coarse global representations and fine local semantics. CoAt-CBM identifies “granularity misalignment” in CLIP-based concept bottleneck models because coarse-grained global image features are often paired with fine-grained textual concepts such as local attributes or parts (Zhong et al., 17 Apr 2026). In visual search, SGML characterizes granularity as the level at which two images are similar, ranging from the exact same instance to similar design or common category (Manandhar et al., 2019). In open-world counting, the relevant distinction is between identity, attribute, category, instance, and concept levels, all of which determine what the system is expected to count (Liu et al., 11 May 2026).

In knowledge representation and qualitative reasoning, granularity is explicitly hierarchical. A rare-disease diagnosis framework organizes concepts across disease families, named diseases and phenotypes, and concrete instances, and uses a three-layer knowledge graph over taxonomy, clinical features, and instances (Zhang et al., 11 Jul 2025). In qualitative spatial reasoning, adjustable granularity determines how finely direction and distance are partitioned into qualitative categories (Moratz, 2010). In rough-set theory, the notion is generalized further: granules may be explicit, but co-granular approximations use ideals and neighborhood conditions to determine when deviations from a concept are negligible (Mani, 2017).

2. Formalization and measurement

Several recent works define explicit quantitative measures of granularity. In activation steering, concept granularity at layer $\ell$ is defined as

$$\mathcal{G}_c(\ell)=\frac{\gamma_c(\ell)}{\mathcal{A}_c(\ell)},$$

where $\gamma_c(\ell)$ is within-question directional agreement and $\mathcal{A}_c(\ell)$ is overall prompt-boundary alignment (Robertson et al., 9 May 2026). High values indicate that prompts agree locally for each question but that the utility-maximizing direction rotates across questions; low values indicate a stable global direction. This formulation treats granularity as directional heterogeneity across contrastive contexts rather than as a symbolic hierarchy.

Granuscore formalizes textual granularity through the geometry of a hyperbolic embedding space. It uses HiT embeddings, radial depth from the origin (Dist0), relational features to a fixed set of random anchors, and a LightGBM regressor, then calibrates the output as a percentile over all WordNet nouns so that lower scores correspond to finer granularity and higher scores to coarser granularity (Ellinger et al., 26 May 2026). Unlike length-based heuristics, this definition is explicitly hierarchical: a concept’s granularity is its position in a semantic hierarchy realized in embedding space.

Other fields use domain-specific formalizations. SGML defines a semantic granularity similarity mapping as cosine similarity between predicted attribute vectors, allowing training pairs to encode multiple levels of similarity rather than a binary same/different label (Manandhar et al., 2019). In rough-set theory, co-granular approximations are defined by

$$X^{l_*}=\{a\in X:\gamma(a)\cap X^c\in \mathcal{I}_\sigma(S)\},\qquad X^{u_*}=\{a\in S:\gamma(a)\cap X\notin \mathcal{I}_\sigma(S)\}\cup X,$$

so membership depends on whether the part of a local neighborhood outside or inside the target set belongs to a suitable ideal (Mani, 2017). In qualitative spatial calculi, the parameter $m$ controls the number of directional sectors and the fineness of distance classes, so higher $m$ yields finer qualitative distinctions (Moratz, 2010).

These formulations are not equivalent, but they share a common structure. In each case, granularity is operationalized through an ordering over representational units: proof segments, sentence embeddings, contrastive activation directions, attribute similarities, spatial sectors, or ideal-filtered neighborhoods. This suggests that concept granularity is typically a relation between a target concept and a representational scaffold rather than a property of the concept alone.

3. Granularity in learned representations

In learned representation systems, granularity often determines the basic predictive unit. Mimir replaces next-token prediction with next-concept prediction, where each concept is a sentence-level SONAR embedding and the training objective is mean squared error in embedding space rather than token cross-entropy (Musacchio et al., 24 May 2026). This shifts the model’s inductive bias from local lexical continuation to sentence-level semantic continuity. The paper reports that such coarse concept granularity is advantageous for semantically abstractive, long-context tasks such as multilingual summarization, but weaker on tasks that require span-accurate or sense-precise outputs.

Concept bottleneck models provide a second major setting. CoAt-CBM addresses granularity misalignment by introducing one learnable query per concept, computing concept-wise attention over patch-level visual features, and then optimizing concept scores with a multi-positive contrastive loss rather than Binary Cross-Entropy (Zhong et al., 17 Apr 2026). The architectural claim is that concept-wise attention yields adaptive fine-grained image-concept alignment, while the optimization claim is that relative ranking among positive and negative concepts better reflects mutual exclusivity among fine-grained attributes. The ablations further show that increasing the ratio of queries to concepts improves performance, indicating that finer concept decoupling materially affects both interpretability and classification.

Several other architectures make granularity itself adaptive. MCBM orders concepts by minimum redundancy maximum relevance and trains nested concept prefixes so that any prefix can serve as a valid bottleneck, yielding multiple levels of conceptual granularity within a single model (Chen et al., 20 May 2026). The paper proves that, under its geometric decay assumptions, expected intervention cost can drop from linear to logarithmic order, $O(\log K)$, while maintaining monotonic performance improvement as more concepts are revealed. MCM likewise uses multiple layers of concept tokens learned from masked images, with different encoder layers providing concepts at different levels of abstraction for the decoder; the resulting concept tokens support both reconstruction and controllable concept editing (Sun et al., 1 Feb 2025). In open-world category discovery, MGCE instantiates multiple conceptual experts at different graph neighborhood scales, so that each expert induces a different clustering resolution and cross-expert alignment couples coarse and fine conceptual structure (Zheng et al., 30 Sep 2025).

4. Language, explanation, and extraction

In human-oriented explanation, concept granularity governs what counts as “one step.” Work on granularity-adaptive proof presentation argues that automated theorem provers usually present proofs at a fixed level of detail that is often unsuitable for human use, whereas mathematicians adapt proof detail to audience knowledge and didactic goals (0903.0314). The proposed approach learns a model of proof granularity from examples and uses that model for the automated generation of further proofs at an adapted level of granularity. This places concept granularity at the boundary between formal derivation and mathematical exposition.

In knowledge-graph completion, MRC-CE treats multi-granular concepts as overlapping text spans in entity descriptions. The framework combines a BERT-based machine reading comprehension model with a pointer network, a random forest selector, and rule-based pruning in order to extract both existing and new instanceOf concepts from descriptive text (Yuan et al., 2022). Because the same text can contain nested concepts such as station, railway station, and the railway station of JR East Japan, the extraction problem is inherently multi-granular. After running MRC-CE for each entity in CN-DBpedia, the system adds more than 7,053,900 new concepts, with a higher average concept length than the original graph, which the paper interprets as improved fine-grained concept coverage (Yuan et al., 2022).

Granuscore extends the same problem to discourse and question answering. It shows that granularity is not equivalent to sentence specificity: a sentence can gain detail without changing the hierarchical level of its referential expressions (Ellinger et al., 26 May 2026). Using scientific articles, the paper finds that introduction sections are typically coarser than related-work sections, and on four QA benchmarks it shows that finer-grained questions and gold answers are associated with lower model accuracy. It also reports systematic differences between question, gold-answer, and output granularity across correct, incorrect, and not-attempted responses, providing a reference-free way to characterize QA difficulty and model behavior.

5. Reasoning, control, and knowledge organization

Granularity also determines whether simple control or reasoning mechanisms are viable. In activation steering, high concept granularity means that a concept is locally linear within each context but globally rotating across contexts, so a single steering direction is an imperfect compromise (Robertson et al., 9 May 2026). Empirically, higher granularity is associated with slower convergence and lower best-found steering utility, with Pearson $r=0.44$ for trials-to-95% and $r=-0.46$ for best-found utility, both with $$\mathcal{G}_c(\ell)=\frac{\gamma_c(\ell)}{\mathcal{A}_c(\ell)},$$0 (Robertson et al., 9 May 2026). GRACE uses this diagnosis to distinguish removable noise from intrinsic cross-context rotation and to decide when rank-1 steering is likely to be cheap and stable.

In medical reasoning, multi-granularity is treated as a design requirement. A rare-disease diagnosis framework couples sparse activation of concepts with four complementary matching algorithms—standardized coding matching, compound terminology segmentation matching, biomedical variant matching, and multilingual cross-cultural similarity—and then grounds activated concepts in a three-layer knowledge graph over taxonomy, clinical manifestations, and instances (Zhang et al., 11 Jul 2025). The full system reports BLEU gains of 0.09, ROUGE gains of 0.05, and accuracy gains of 0.12 on the BioASQ rare-disease QA set, with peak accuracy of 0.89 approaching the 0.90 clinical threshold (Zhang et al., 11 Jul 2025). Here granularity is not merely representational: diagnosis requires movement across disease families, phenotypes, genes, and concrete cases.

The same pattern appears in perception and spatial reasoning. “Count Anything at Any Granularity” argues that open-world counting fails because counting granularity is usually left implicit, and it formalizes five explicit levels—identity, attribute, category, instance, and concept—through controlled target and distractor subsets (Liu et al., 11 May 2026). KubriCount supplies the multi-category scenes and annotations required to test these distinctions, and HieraCount improves multi-grained counting accuracy with MAE 4.67 and RMSE 11.07 on KubriCount, while level 4 remains the hardest due to fine-grained within-category instance-type discrimination (Liu et al., 11 May 2026). In qualitative spatial calculi, adjustable granularity similarly regulates expressiveness: $$\mathcal{G}_c(\ell)=\frac{\gamma_c(\ell)}{\mathcal{A}_c(\ell)},$$1 partitions direction into more or fewer sectors as $$\mathcal{G}_c(\ell)=\frac{\gamma_c(\ell)}{\mathcal{A}_c(\ell)},$$2 varies, and $$\mathcal{G}_c(\ell)=\frac{\gamma_c(\ell)}{\mathcal{A}_c(\ell)},$$3 extends this principle to distance through “hidden feature attachment,” adding an internal reference distance or elevation to each point (Moratz, 2010). Rough-set theory generalizes the idea further by making ideals, antichains, and contact relations part of the approximation mechanism, so that concept boundaries depend on what deviations are treated as negligible (Mani, 2017).

6. Trade-offs, misconceptions, and open directions

A recurrent finding is that coarse and fine granularity have complementary strengths and weaknesses. In Mimir, sentence-level concept modeling improves multilingual summarization relative to a comparable token-based baseline but remains weaker on MLQA and XL-WSD, where precise phrasing and fine sense distinctions matter (Musacchio et al., 24 May 2026). In SGML, visual similarity at multiple granularities improves retrieval because training pairs are not equally informative; treating graded semantic similarity explicitly yields Recall@1 gains on DeepFashion In-Shop (Manandhar et al., 2019). In MCBM, a coarse prefix of high-relevance concepts can reduce expert intervention cost, while larger prefixes recover accuracy when needed (Chen et al., 20 May 2026). In activation steering, high granularity is not evidence that a concept is absent; the paper argues that it often reflects search difficulty and contextual rotation rather than the nonexistence of useful rank-1 interventions (Robertson et al., 9 May 2026).

A common misconception is to equate granularity with either surface length or generic “specificity.” Granuscore shows that sentence length alone does not explain granularity and that granularity captures non-linear variation in specificity beyond word count (Ellinger et al., 26 May 2026). In counting, the failure of both multimodal LLMs and specialist counting models under fine-grained distinctions shows that prompt following at the intended granularity is a separate competence from generic open-world recognition (Liu et al., 11 May 2026). In concept bottleneck models, CoAt-CBM likewise argues that a frozen global embedding is not simply a weak feature; it is a representation at the wrong granularity for localized or mutually exclusive concepts (Zhong et al., 17 Apr 2026).

The literature also points to several open directions. High-granularity steering suggests multi-vector or context-adaptive interventions rather than single global directions (Robertson et al., 9 May 2026). Mimir explicitly proposes larger-scale multilingual concept models and multimodal extensions (Musacchio et al., 24 May 2026). MCM highlights the need for dynamically expanding concept spaces for continual learning (Sun et al., 1 Feb 2025). Work on proof presentation suggests richer user models that adapt proof granularity per concept rather than only per user level (0903.0314). Rare-disease diagnosis emphasizes the continuing problem of aligning heterogeneous knowledge sources at multiple levels of abstraction (Zhang et al., 11 Jul 2025). Across these cases, the central problem remains the same: concept granularity is useful only when a system can choose, infer, or control the right level of abstraction for the task at hand.