Concept Flow Models (CFMs) Overview

• Concept Flow Models (CFMs) are flow-centered frameworks that use intermediate structures to constrain information propagation in tasks like conceptual modeling, ideology detection, and visual reasoning.
• CFMs employ distinct formulations—including thining-machine modeling, bidirectional iterative diffusion, and hierarchical decision trees—to address challenges in diverse application domains.
• Each variant of CFM enhances interpretability and performance by enforcing localized reasoning, while differing in semantics, architecture, and standardization.

Searching arXiv for the specified Concept Flow Models papers to ground the article and verify bibliographic details. arXiv search: "Concept Flow Models" Concept Flow Models (CFMs) denote a family of flow-centered representational frameworks rather than a single standardized formalism. In the arXiv literature surveyed here, the label has been used for at least three distinct constructions: a conceptual-modeling methodology built on the Thinging-Machine ontology and three levels of specification; a bidirectional hierarchical concept-propagation mechanism for multifaceted ideology detection; and a hierarchical bottleneck architecture for concept-based visual reasoning. Across these usages, “flow” refers to structured propagation through a diagram, schema tree, or decision hierarchy, but the modeled entities, objectives, and mathematical machinery differ substantially (Al-Fedaghi, 2020, Liu et al., 2024, Wang et al., 17 Jun 2026).

1. Terminological scope and recurring design pattern

The main usages of the term in this literature can be organized as follows.

Formulation	Core structure	Primary use
Thinging-Machine / Concept-Flow Model	thimacs, five generic operations, event chronology	conceptual modeling and system specification
Bidirectional Iterative Concept Flow	hierarchical schema tree with root-to-leaf diffusion and leaf-to-root aggregation	multifaceted ideology detection
Hierarchical bottleneck CFM	concept-driven decision tree with local concept subsets	interpretable image classification

In the 2020 conceptual-modeling line, a CFM is a methodology for representing static structure, dynamic events, and behavior through a single ontology of “things” and “machines” (Al-Fedaghi, 2020). In the 2024 NLP line, concept flow is a representation-learning mechanism in which concepts move across levels of a schema tree to enrich label embeddings with multi-granularity semantics (Liu et al., 2024). In the 2026 interpretable-vision line, CFMs are an extension of Concept Bottleneck Models (CBMs) in which a flat concept bottleneck is replaced with a hierarchical, concept-driven decision tree (Wang et al., 17 Jun 2026).

A common misconception is to treat these CFMs as a unified technical framework. The cited works instead use the same label for different ontological commitments and learning problems. A plausible unifying characterization is that each formulation constrains prediction or specification by routing information through intermediate concept-bearing structures rather than allowing an unconstrained end-to-end mapping.

2. Thinging-Machine CFM: static, dynamic, and behavioral specification

In the Thinging-Machine formulation, the single ontological primitive is the “thimac,” which is both a “thing” and a “machine.” As a thing, a thimac has parts, attributes, and “flow-trajectories” by which it receives or hands off subthings. As a machine, it performs five generic operations on things: Create, Process, Release, Transfer, and Receive (Al-Fedaghi, 2020).

This framework defines three levels of representation. The static structural description is a single diagram in which thimacs are nodes and flows are directed arcs. Formally, with $T$ the set of all things and $$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$, a static CFM model is

$M_{\text{static}}=(T,N,F),$

where $N\subseteq T$ is the set of named machines/things and

$F\subseteq N\times O\times N$

is the set of ordered triples $(t_i,\text{op},t_j)$ meaning that machine $t_i$ performs operation $\text{op}$ and hands off a thing to $t_j$ or to itself. Within each thimac, the internal five-stage flow imposes a local partial order

$$\text{create} \prec \text{process} \prec \text{release} \prec \text{transfer} \prec \text{receive}.$$

The dynamic model turns static operations into events. Each occurrence of a static arc is an event

$$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$0

with timestamp $$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$1, and the event set is collected as

$$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$2

Higher-level domain events are formed by grouping generic events. The paper gives examples such as

$$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$3

and

$$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$4

The behavioral level is the chronology of events. A behavior is a sequence

$$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$5

such that adjacent events are causally linked by a flow arc or an allowed nondeterministic trigger, each event obeys the local precedence from create to receive, and domain-specific constraints are satisfied. Because the static model contains all flows but no single forced ordering across unrelated flows, multiple behaviors may be extracted by choosing different topological sorts of the event graph.

The restaurant script is the central illustration. Its principal machines are Customer, Restaurant, Waiter, Cook, CashRegister, and sub-machines under Customer for states “hungry” and “has-money.” Example flows include $$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$6, $$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$7, $$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$8, and $$O=\{\text{create},\text{process},\text{release},\text{transfer},\text{receive}\}$$9. At the dynamic level, elementary events are labeled $M_{\text{static}}=(T,N,F),$0 through $M_{\text{static}}=(T,N,F),$1, from customer receipt into the restaurant to customer departure. From the same static model, four distinct service behaviors are derived: FastFood ($M_{\text{static}}=(T,N,F),$2), Restaurant ($M_{\text{static}}=(T,N,F),$3), Buffet ($M_{\text{static}}=(T,N,F),$4), and ChurchSupper ($M_{\text{static}}=(T,N,F),$5) (Al-Fedaghi, 2020).

The stated advantages are a uniform ontology, a clear separation of static structure vs. run-time events vs. overall behavior, support for multiple, alternative behaviors without duplicating static diagrams, and avoidance of UML’s mixing of “activity” and “event.” The stated limitations are a steep learning curve, potentially overwhelming diagrams, implicit concurrency semantics, the need for a translation layer to executable or standard modeling languages, and the lack of a standardized notation or tool-chain for large industrial applications.

3. Bidirectional iterative concept flow in ideology detection

A second usage of concept flow appears in multifaceted ideology detection (MID), where label semantics and explanatory descriptions are encoded through a hierarchical schema. The schema tree is

$M_{\text{static}}=(T,N,F),$6

with four node types: Root, Domain, Facet, and Ideology. Edges represent parent-to-child subordination. Facet and Ideology nodes are initialized by feeding their natural-language definitions into a shared pre-trained encoder such as BERTweet and taking the $M_{\text{static}}=(T,N,F),$7 hidden state, producing $M_{\text{static}}=(T,N,F),$8. Domain and Root nodes average-pool their immediate children (Liu et al., 2024).

The core mechanism is Bidirectional Iterative Concept Flow (BICo). At each level $M_{\text{static}}=(T,N,F),$9, node embeddings are maintained as $N\subseteq T$0, and each edge type $N\subseteq T$1 has a learnable relation-rotation vector $N\subseteq T$2. The root-to-leaf phase, called Concept Metapath Diffusion, updates a path $N\subseteq T$3 by

$N\subseteq T$4

The leaf-to-root phase, called Concept Hierarchy Aggregation, uses attention-style aggregation:

$N\subseteq T$5

$N\subseteq T$6

$N\subseteq T$7

The iterative update alternates these two directions for $N\subseteq T$8 iterations and outputs enriched facet embeddings $N\subseteq T$9 and ideology embeddings $F\subseteq N\times O\times N$0.

These concept-flow representations are integrated into a two-stage MID architecture. In Subtask 1: Relevance Recognition, text token embeddings $F\subseteq N\times O\times N$1 are matched against facet embeddings through Concept Attentive Matching,

$F\subseteq N\times O\times N$2

followed by a facet-specific classifier and an auxiliary contrastive loss $F\subseteq N\times O\times N$3. In Subtask 2: Ideology Analysis, for each facet judged related, the model encodes the concatenated sequence $F\subseteq N\times O\times N$4 text $F\subseteq N\times O\times N$5 facet-concept, applies a facet-specific classifier over $F\subseteq N\times O\times N$6, and adds Concept-Guided Contrastive Learning (CGCL) with ideology concept anchors.

The semantic motivation is explicit. Upward aggregation pulls fine-grained distinctions into parent nodes, while downward diffusion injects broad abstract semantics into specific facets and ideologies. Iterating the two processes yields embeddings that simultaneously encode the abstract viewpoint from above and the specialized perspective from below. The paper argues that this enriched representation improves robustness under topic shift by allowing the model to use the relation between a specific label and a broader domain structure (Liu et al., 2024).

Empirically, experiments on MITweet with 12 594 tweets, 12 facets, 14 topics report state-of-the-art performance. For Relevance Recognition, BERTweetInd obtains Macro-F1 = 57.48 and Micro-F1 = 70.32, while BICo obtains Macro-F1 = 59.22 and Micro-F1 = 72.14. For Ideology Analysis, BERTweetInd obtains Macro-F1 = 42.68, Micro-F1 = 69.28, and Micro-Acc = 76.38, while BICo obtains Macro-F1 = 47.32, Micro-F1 = 70.90, and Micro-Acc = 78.60. In cross-topic transfer, the reported improvements are statistically significant at $F\subseteq N\times O\times N$7, and ablations show degradation from removing BICo, removing either diffusion or aggregation, or disabling CGCL (Liu et al., 2024).

4. Hierarchical bottleneck CFMs for concept-based reasoning

A third formulation treats Concept Flow Models as a hierarchical extension of Concept Bottleneck Models. The motivating claim is that flat concept bottlenecks suffer from information leakage when the number of concepts approaches the embedding dimension, allowing the model to exploit spurious or semantically irrelevant correlations. The proposed remedy is to replace the flat concept-to-label bottleneck with a hierarchical, concept-driven decision tree in which each internal node focuses on a localized subset of discriminative concepts (Wang et al., 17 Jun 2026).

Architecturally, an input image is encoded into a visual embedding $F\subseteq N\times O\times N$8. Rather than projecting once into all $F\subseteq N\times O\times N$9 global concepts, a CFM organizes the concepts into $(t_i,\text{op},t_j)$0 internal tree nodes. Each internal node $(t_i,\text{op},t_j)$1 has a local concept matrix $(t_i,\text{op},t_j)$2 with $(t_i,\text{op},t_j)$3, a trainable weight matrix $(t_i,\text{op},t_j)$4, and an optional temperature $(t_i,\text{op},t_j)$5. Using normalized image features $(t_i,\text{op},t_j)$6, each node computes

$(t_i,\text{op},t_j)$7

For a leaf $(t_i,\text{op},t_j)$8, the class probability is the product of branching probabilities along the root-to-leaf path, multiplied by a learnable leaf bias $(t_i,\text{op},t_j)$9:

$t_i$0

Prediction is $t_i$1.

Hierarchy construction proceeds from visual statistics. The method computes per-class CLIP image centroids $t_i$2, applies agglomerative clustering with Ward’s linkage, prunes nodes with merge distance below the $t_i$3-th percentile or low LDA separability, and may annotate internal nodes with concise natural-language labels using an LLM prompt. Concept generation is also explicit: for each leaf class, an LLM such as GPT-4 generates 20–30 candidate noun-phrase concepts using class name, sample images, and sibling context. Candidates are aggregated bottom-up to each internal node into a pool $t_i$4 of approximately $t_i$5 phrases, embedded via CLIP’s text encoder, filtered for class-name leakage and near-duplicates, and then pruned by a multi-class logistic-regression Lasso. Each concept $t_i$6 is scored by

$t_i$7

and the top $t_i$8 concepts are selected. The per-node concept budget follows

$t_i$9

Training freezes the CLIP encoder and learns only $\text{op}$0 by minimizing the negative log-likelihood

$\text{op}$1

optionally with hierarchical cross-entropy and Elastic Net regularization. A key interpretability metric is NEC (Effective Concepts), the average number of concepts with non-negligible weights at inference. Because a prediction only uses concepts on its path, NEC scales as $\text{op}$2. The paper further states a theoretical contrast: to preserve separability with random concepts, a flat CBM needs $\text{op}$3, whereas a CFM with $\text{op}$4 nodes requires $\text{op}$5. Conversely, if semantic concepts exist that locally separate siblings, a CFM can match a CBM’s accuracy with $\text{op}$6 semantic concepts while reducing per-prediction usage to $\text{op}$7 (Wang et al., 17 Jun 2026).

The experimental evaluation covers CIFAR-10, CIFAR-100, UCF-101, CUB-200-2011, TinyImageNet, and ImageNet-1K in the appendix. On CIFAR-10, the main reported results are: LinearProbe accuracy 94.6%; CBM(Random) NEC 56.8, accuracy 89.2%, SIR 0%; CFM(Random) NEC 6.9, accuracy 76.5%, SIR 0%; PCBM NEC 56.8, accuracy 91.2%, SIR 2.2%; Labo NEC 57.7, accuracy 90.9%, SIR 1.9%; and CFM (Ours) NEC 6.9, accuracy 91.6%, SIR 19.8%. On TinyImageNet, CFMs are reported to match or exceed CBM accuracy while reducing NEC from approximately 362 to approximately 92 and increasing SIR from 1–3% to 9%. Ablations attribute gains to hierarchy quality, Lasso-based concept selection, depth pruning, temperature scaling, and leaf-bias calibration (Wang et al., 17 Jun 2026).

This formulation makes the strongest direct interpretability claim among the surveyed CFMs: predictions are associated with stepwise decision flows that are said to be transparent, auditable, and editable. The interpretability mechanism is therefore structural rather than post hoc: the model is prevented from consulting all concepts simultaneously.

5. Comparative structure, semantics, and interpretability

The three principal CFM formulations differ in mathematical substrate, but each introduces intermediate structure that constrains how information may propagate.

In the Thinging-Machine line, the primitives are ontological: thimacs and the five generic operations define admissible movement and transformation of “things.” The distinction between static model, dynamic event hierarchy, and behavioral chronology is explicit, and behavior is extracted as permissible sequences from a single static diagram (Al-Fedaghi, 2020). In BICo, the primitives are hierarchical label embeddings in a schema tree, and “flow” is a learned bidirectional exchange between levels realized through complex-space metapath diffusion and attention-style aggregation (Liu et al., 2024). In the hierarchical bottleneck CFM, the primitives are local concept subsets attached to internal decision nodes, and flow is a probabilistic traversal from root to leaf driven by local softmax routers (Wang et al., 17 Jun 2026).

The role of hierarchy also varies. In the 2020 model, hierarchy organizes events and submachines within a conceptual specification. In the 2024 model, hierarchy encodes semantic granularity from root, domain, facet, to ideology. In the 2026 model, hierarchy narrows class scope and isolates concepts to control effective concept usage. This suggests that “concept flow” functions less as a single mathematical operator than as a design principle for localized reasoning under explicit structural constraints.

Interpretability is likewise non-uniform. The Thinging-Machine formulation emphasizes unambiguous specification and separation of structural, dynamic, and behavioral levels. BICo emphasizes semantically enriched concept embeddings that act as anchors in a contrastive space, where ideologically similar texts cluster around concept vectors. The hierarchical bottleneck formulation emphasizes auditable decision paths and mitigation of information leakage by limiting the concepts consulted at any one decision stage. These are related but distinct senses of interpretability: specification clarity, semantic alignment, and constrained decision exposure.

A further point of divergence is standardization. One line explicitly states that the notation and tool-chain are not standardized for large industrial applications, and another line is presented as a new extension of CBMs rather than as a continuation of the earlier conceptual-modeling ontology. The literature therefore does not yet support the view that CFMs constitute a single consolidated research program.

6. Acronym overlap with other flow-map literatures

The acronym CFM is also used in nearby literatures for models that are not Concept Flow Models in the senses above. This is a persistent source of confusion.

Acronym usage	Expansion	Domain
CFM	Conditional Flow Matching	generative modeling
CFM	Categorical Flow Maps	non-autoregressive language modeling
CFM	Contextual Flow Maps	dynamical-systems theory of transformers

In Conditional Flow Matching, a model learns a vector field $\text{op}$8 whose ODE transports samples from a simple prior $\text{op}$9 to a data distribution $t_j$0. The paper on Implicit Dynamical Flow Fusion (IDFF) addresses the sampling inefficiency of such models by adding a Hamiltonian-inspired momentum term and using an implicit ODE update

$t_j$1

which allows $t_j$2 with $t_j$3. The reported empirical claim is up to a 10× NFE reduction relative to CFM baselines on image and time-series tasks, with examples including CIFAR-10, CelebA-64, polyalanine, and NOAA OISSTv2 (Rezaei et al., 2024).

In Categorical Flow Maps, the target is discrete sequence generation. A differentiable interpolation schedule $t_j$4 defines a stochastic interpolant between a Gaussian prior and one-hot data:

$t_j$5

A two-time flow-map parameterization $t_j$6 is then learned, together with diagonal and self-distillation losses. The reported large-scale result is a 1.7B-parameter base flow model trained on 2.1T tokens, self-distilled into a CFM that generates high-quality text in as few as 4 inference steps while maintaining near-data-level token entropy, and that can be scored with a semi-discrete ELBO on standard LM benchmarks (Davis et al., 8 May 2026).

In Contextual Flow Maps, transformers are analyzed as interacting-particle dynamical systems. A distinguished token $t_j$7 evolves under a velocity field $t_j$8 coupled to a contextual empirical measure $t_j$9, and the infinite-context limit replaces the empirical measure by the underlying population measure. The core theoretical results are propagation-of-chaos bounds with rate $$\text{create} \prec \text{process} \prec \text{release} \prec \text{transfer} \prec \text{receive}.$$0 in general and $$\text{create} \prec \text{process} \prec \text{release} \prec \text{transfer} \prec \text{receive}.$$1 under a kernel-linear structural assumption that includes single-head and multi-head self-attention, for both forward inference and online gradient-descent training trajectories (Chen et al., 16 May 2026).

The overlap in acronym usage does not imply conceptual equivalence. In the conceptual-modeling, ideology-detection, and hierarchical-bottleneck literatures, “concept” is central to the model’s ontology or supervision interface. In the conditional, categorical, and contextual flow-map literatures, the central object is instead a transport or dynamical flow. The shared abbreviation therefore obscures substantial differences in objectives, mathematics, and interpretability claims.