Semantic Information Flow Overview

• Semantic information flow is the directed propagation, transformation, and preservation of meaning-bearing structures across systems, uniting approaches from ML, document analysis, dialogue, and vision.
• Key frameworks capture semantic flow via latent state trajectories, graph transitions, and conditional influence measures, which enhance interpretability and security in AI applications.
• Applications range from ML execution tracing and scene segmentation to dialogue dynamics and institutional knowledge, driving practical advances in efficiency and analytical precision.

Semantic information flow denotes the directed propagation, transformation, or preservation of meaning-bearing structure within a system. The term does not name a single standardized formalism; rather, it spans a family of related constructions in which semantics is represented as latent states, semantic fields, utterance-level context shifts, security-relevant influence, or truth-preserving constraint transport. In machine-learning systems, semantic flow is defined as a sequence of semantic states $s_1,s_2,\ldots,s_n$ in latent spaces traversed during execution (Yoo et al., 13 Mar 2025). In document analysis, it is the ordered sequence of semantic fields traversed by a text across sentences (Jr. et al., 2019). In LLM security, it is quantified as the mutual information between perturbations of an input span and downstream outputs (Storek et al., 22 Jun 2026). This suggests that the unifying concern is not raw data movement, but the structure by which internal meaning changes, stabilizes, and becomes observable.

1. Conceptual scope and recurring formal patterns

Across the literature, semantic information flow is usually defined through three ingredients: a representation space for semantics, a directed relation over that space, and an aggregation or transport rule that connects local states into a larger structure. In ML-based systems, the basic objects are semantic states and latent mappings: “Let $e_i$ represent the execution data at step $i$,” let $embed(i,e_i)$ map that execution data into a semantic state $s_i \in LS_i$, and let $aggregate(i,s_i,S_i)$ cluster states into nodes of a Semantic Flow Graph (Yoo et al., 13 Mar 2025). In document networks, the semantic carrier is a sequence of community labels $c_1,c_2,\ldots,c_T$ induced from sentence embeddings, with transition probabilities computed as $P_{ij} = n_{ij}/\sum_j n_{ij}$ in a first-order Markov chain (Jr. et al., 2019). In semantic transfer-entropy models of communication, the carrier is the LLM-induced conditional distribution over utterances, and directed influence is measured by conditional log-likelihood differences (Goodall et al., 29 Jun 2026).

A second recurring pattern is asymmetry. Security-oriented work distinguishes source spans, perturbations, and sinks; semantic influence is directional and not reducible to co-occurrence or attention heuristics (Storek et al., 22 Jun 2026). Dialogue models distinguish the current conversational context from the semantic influence of the next utterance, defining $I_k = C_{k+1}-C_k$ at the level of utterance representations (Li et al., 2021). Cognitive work based on the Context-Content Uncertainty Principle similarly assumes a broken symmetry between high-entropy context $\Psi$ and low-entropy content $e_i$0, making iterative inference necessary (Li, 8 Jul 2025). A plausible implication is that semantic information flow is typically invoked when static representations are insufficient and one must model how meaning-bearing structure changes under time, computation, or interaction.

A third recurring pattern is the combination of semantic transport with external observables. Some frameworks infer semantics from internal activations or function calls (Yoo et al., 13 Mar 2025); some infer it from sentence embeddings and narrative order (Jr. et al., 2019); some infer it from cross-attention conditioning in diffusion models (Kaplan et al., 1 Apr 2025); others derive it from satisfaction-preserving translations between signatures, models, and theories (Kent, 2016). The resulting notion of “flow” ranges from graph transitions and optimal transport to institution-theoretic channels, but the central issue remains the same: how semantic structure moves without being reduced to surface form alone.

2. Execution-centric semantics in machine-learning systems

A prominent software-engineering formulation appears in “Capturing Semantic Flow of ML-based Systems” (Yoo et al., 13 Mar 2025). The core claim is that traditional dynamic analysis observes only externally visible behavior, whereas decisive behavior in ML-based systems occurs in latent spaces such as DNN activations or LLM internal reasoning states. The paper defines a semantic state $e_i$1 as a point in a latent space and a semantic flow as a sequence of such states across execution phases. Nodes in Semantic Flow Graphs represent clusters of semantically related states, and edges represent observed or theoretically possible transitions across executions. Hybrid Semantic and Control Flow Graphs are proposed to combine these structures with ordinary CFGs, enabling coverage notions such as “$e_i$2-coverage” and finer-grained distance-based partial coverage. The paper is explicit that this is a conceptual framework: it sketches properties and applications, but does not provide full soundness or completeness guarantees.

The DNN example uses layers as the unit of analysis, activation values as execution data, and t-SNE to map layer-wise activations into semantic states. In the visualization, aggregation is “trivial” and based on ground-truth labels, showing that classes become increasingly distinct through the network even though the forward pass is linearly unbranched. The LLM-agent example uses AutoFL function calls as inference steps; function call type and arguments are encoded discretely with one-hot representations so that identical calls with the same arguments map to the same state. Kim et al.’s LLM Inference Graph is identified as “a specific instance of a semantic flow graph,” with merged nodes for identical function calls and edge weights reflecting transition frequency. In that setting, a GCN over the LIG achieved “precisions of over 0.8” for predicting whether agent executions could reach a correct answer (Yoo et al., 13 Mar 2025).

A complementary interpretability line appears in “VISIT: Visualizing and Interpreting the Semantic Information Flow of Transformers” (Katz et al., 2023). There, semantic flow is the movement and transformation of token-level semantics during a single forward pass of a transformer. Internal vectors—hidden states, attention head outputs, and memory values—are projected into vocabulary space via a logit lens. The paper shows that the OV circuit is essential for interpreting attention outputs semantically, that only top-norm heads meaningfully align with block outputs and final logits, and that top-ranked memory values by attention score exhibit strong semantic alignment with head outputs. It also reports that layer norms act as semantic filters, depressing probabilities for ubiquitous function words and boosting content words, while certain high-entropy FF2 neurons behave as regularization vectors rather than specific semantic contributors (Katz et al., 2023).

Taken together, these works treat semantic flow as an internal execution structure. Instead of asking only which output was produced, they ask which latent states, attention-mediated recalls, or clustered semantic trajectories made that output reachable. This recasts testing, coverage, fault localization, and interpretability as problems over semantic state transitions rather than over surface control flow alone.

3. Language, dialogue, and communicative dynamics

In text analysis, Corrêa Jr., Marinho, and Amancio define semantic flow as the trajectory of a document across semantic fields detected in a sentence-similarity network (Jr. et al., 2019). Sentences are embedded by averaging Word2Vec vectors after stopword removal and omission of out-of-vocabulary words, then linked by a $e_i$3-nearest-neighbor rule. Louvain communities define semantic fields, and the sequence of field labels in textual order becomes a first-order Markov chain. Motif counts in that chain are then used for classification. The strongest reported result is 92.5% accuracy for distinguishing Investigative from Philosophy books using simplified weighted motifs with threshold 0.120, while chronology classification also produced significant results, though weaker than thematic classification. The framework thus treats semantic flow as a mesoscopic dynamic structure: not merely which topics appear, but how a text revisits and shifts among them over time.

Dialogue work moves from document-level fields to turn-level contextual state. “Conversations Are Not Flat” defines the context after utterance $e_i$4 by a dense vector $e_i$5 and the semantic influence of utterance $e_i$6 as $e_i$7 (Li et al., 2021). DialoFlow predicts future context from prior contexts with a dedicated flow module and uses the predicted influence $e_i$8 to guide token generation. Three training objectives are used: Context Flow Modeling, Semantic Influence Modeling, and Response Generation Modeling. On DailyDialog, DialoFlow-large achieved NIST-4 4.08, BLEU-4 11.57%, and METEOR 17.0%, outperforming DialoGPT-large. For evaluation, the paper proposes a Flow score based on the similarity between predicted and realized influence vectors; on DSTC9 interactive evaluation data, the score reached chatbot-level Pearson $e_i$9 and Spearman $i$0 with human ratings (Li et al., 2021).

A more explicitly information-theoretic formulation appears in “Information Dynamics of Language Communication” (Goodall et al., 29 Jun 2026). There, semantic information flow is the directed predictive coupling between interlocutors measured from LLM-conditioned probabilities. Semantic Transfer Entropy is defined as

$i$1

and Semantic Partial Information Decomposition resolves redundant, unique, and synergistic contributions of multiple sources to a target. Across four experiments, the framework detected reduced information flow in cognitively rigid dialogue, found a strong persuader-to-persuadee asymmetry in PersuasionForGood, distinguished high- from low-quality psychotherapy by therapist-client directionality, and showed positive synergy for two-premise argumentative claims with increasing redundancy as premise count grew (Goodall et al., 29 Jun 2026).

These language-oriented formulations differ in granularity but converge on a common claim: semantic flow is not exhausted by sequential token prediction. It is the structured evolution of discourse state, field transitions, or speaker-conditioned predictability, and it can be operationalized with graphs, context vectors, or conditional mutual information.

4. Vision, scene semantics, and multimodal generation

In computer vision, “Semantic Flow: Learning Semantic Field of Dynamic Scenes from Monocular Videos” defines semantic flow as a neural semantic field attached to continuous 3D motion trajectories rather than to isolated points (Tian et al., 2024). A point trajectory is written as $i$2 with $i$3, and semantic logits are rendered along rays using density and transmittance as opacity priors. The paper argues that these opacity priors resolve the 2D-to-3D ambiguity that arises when supervising with 2D semantic labels. Experimentally, the model achieved higher mIoU and accuracy than baselines on the Semantic Dynamic Scene dataset; example figures reported include Balloon1 0.919 / 0.844, Balloon2 0.967 / 0.839, and Umbrella 0.970 / 0.884. The learned field also supports instance-level scene editing, semantic completion, dynamic scene tracking, and semantic adaptation on novel scenes (Tian et al., 2024).

Text-to-image work studies semantic information flow at the text-encoder interface. “Follow the Flow” examines contextualized token representations in diffusion models and reports two main phenomena: strong in-item token redundancy and inter-item semantic leakage (Kaplan et al., 1 Apr 2025). In 89% of lexical items, at least one token’s contextualized representation alone suffices to render the item; removing non-representative tokens preserves alignment in 98% of aligned cases and reduces error rate by 21% when the original image failed to align. Inter-item leakage appears in 11% of cases, and replacing a leaked item’s contextualized representation with its uncontextualized counterpart reduces leakage by 85% on the 104 leaking prompts. The paper’s central methodological point is that semantic distortions often originate before cross-attention, inside the contextualized token representations themselves (Kaplan et al., 1 Apr 2025).

A different but related use of the term appears in “SFNet: Faster and Accurate Semantic Segmentation via Semantic Flow” (Li et al., 2022). Here semantic flow is a learnable 2D offset field that aligns adjacent levels of a feature pyramid, allowing high-level semantics to be dynamically warped onto high-resolution features. The Flow Alignment Module predicts offsets $i$4 and performs differentiable bilinear warping, while SFNet-Lite introduces a Gated Dual Flow Alignment Module for direct alignment of low- and high-resolution features. On Cityscapes test, SFNet-Lite achieved 80.1 mIoU while running at roughly 60 FPS with a ResNet-18 backbone, and 78.8 mIoU at 120 FPS with an STDC backbone; on the Unified Driving Segmentation dataset, both SFNet and SFNet-Lite were reported as strong speed-accuracy baselines (Li et al., 2022).

Related segmentation work also models semantic information flow from global concepts to local predictions. “Recalling Holistic Information for Semantic Segmentation” uses a two-stream network in which image-level class confidences $i$5 gate per-pixel probabilities $i$6 through $i$7, explicitly favoring high-recall holistic inference (Hu et al., 2016). The paper reports that segmentation performance is far more sensitive to recall than precision in this global semantic prior. In autonomous driving, “Exploiting Semantic Information and Deep Matching for Optical Flow” uses instance-level segmentation to partition scenes into background and independently moving objects, thereby converting optical flow estimation into object-wise epipolar subproblems (Bai et al., 2016). These latter works do not define semantic flow in exactly the same way as (Tian et al., 2024) or (Kaplan et al., 1 Apr 2025), but they share the broader idea that semantics can structure geometric or pixel-level information transport.

5. Security, leakage, and information-flow control

The most formally explicit security account in the corpus is “GIF: Locally Sound Geometric Information Flow Control for LLMs” (Storek et al., 22 Jun 2026). GIF treats an LLM invocation as an information-theoretic channel and defines semantic flow from an input span $i$8 to an output $i$9 as $embed(i,e_i)$0, where $embed(i,e_i)$1 is a Gaussian perturbation of the span embeddings. Using the span-restricted Jacobian, the output-head geometry, and the softmax Fisher information, the paper defines a pullback matrix $embed(i,e_i)$2 and a capacity-like score

$embed(i,e_i)$3

with the main local soundness statement

$embed(i,e_i)$4

Unlike attention-based or single-logit heuristics, GIF is given a fully mechanized Lean 4 proof of local geometric soundness under local regularity assumptions. Empirically, it achieved near-perfect recall without a downstream declassifier, matched or exceeded direct LLM-as-judge baselines when combined with lightweight declassifiers, reduced token cost by up to 81x, and transferred effectively from surrogates up to 200x smaller than the target model (Storek et al., 22 Jun 2026).

At the architectural level, “Information-flow Interfaces and Security Lattices” provides a semantics for IFIs in terms of security lattices (Bartocci et al., 2024). Interfaces specify assumptions and guarantees as no-flow relations over ports; the paper replaces the original contract semantics with lattice contracts in which permitted flow is the partial order itself. For labels $embed(i,e_i)$5, “$embed(i,e_i)$6 can-flow to $embed(i,e_i)$7” iff $embed(i,e_i)$8, and the product lattice combines confidentiality and integrity so that $embed(i,e_i)$9 iff $s_i \in LS_i$0 and $s_i \in LS_i$1. A translation algorithm constructs lattices from flow relations and is proved correct, giving a faithful order-theoretic interpretation of interface-level security policies (Bartocci et al., 2024).

At the programming-language level, “Types for Information Flow Control: Labeling Granularity and Semantic Models” shows that fine-grained and coarse-grained IFC type systems are equally expressive (Rajani et al., 2018). Labels range over a security lattice $s_i \in LS_i$2; noninterference is proved semantically with step-indexed Kripke logical relations supporting higher-order state. The paper proves type- and semantics-preserving translations both from a coarse-grained system to a fine-grained one and back, and argues that the backward translation requires a construct such as toLabeled to limit the scope of context taint. This body of work treats semantic information flow as dependence control rather than latent-state geometry, but the common theme remains the formal characterization of which meaning-bearing distinctions are allowed to influence which observables (Rajani et al., 2018).

6. Ontologies, institutions, and distributed knowledge

A separate tradition treats semantic information flow as truth-preserving transport across distributed knowledge structures. “Information Flow in Logical Environments” generalizes Barwise and Seligman’s channel theory from classifications to arbitrary logical environments or institutions (Kent, 2016). An institution is given as $s_i \in LS_i$3, with the satisfaction condition

$s_i \in LS_i$4

for any signature morphism $s_i \in LS_i$5. In this setting, a distributed system is a diagram of structures, a channel is a cocone to a common core, and semantic integration is performed by transporting local constraints to the core via direct image and then taking their meet. Information flow is therefore semantic in the strict sense that it preserves satisfaction rather than only symbolic form (Kent, 2016).

The Information Flow Framework and related ontology-sharing work extend this logic to conceptual integration. “Semantic Integration in the Information Flow Framework” describes institutions, institution morphisms, and theory colimits as the basic machinery for semantic integration, with fusion implemented by transporting theories to a colimit signature and taking their meet after closure (Kent, 2018). “The Information Flow Foundation for Conceptual Knowledge Organization” formalizes ontology sharing by modeling participant ontologies as IF logics, a common ontology as an IF theory, and the shared virtual ontology as the fusion of participant ontologies: the quotient of their sum modulo synonymic type equivalence and sharing links (Kent, 2018). In both cases, semantic flow is not a trajectory in latent space but a categorical transport of constraints, instances, and type correspondences across a shared semantic core.

These frameworks are structurally distant from neural-state or dialogue models, yet they preserve a recognizable definition of flow. Meaning is localized in types, sentences, and theories; directed mappings preserve truth or consequence; and global integration arises by composing local semantic channels. A plausible implication is that “semantic information flow” can denote either dynamic state evolution or semantic-preserving translation, depending on whether the underlying representation is probabilistic, graph-theoretic, or institutional.

7. Extended theoretical directions and open issues

Several works broaden the topic beyond immediate engineering applications. “Wikipedia information flow analysis reveals the scale-free architecture of the Semantic Space” constructs a directed network of semantic flows between Wikipedia entries using size-normalized Jensen–Shannon divergence and an entropy-based directionality index (Masucci et al., 2011). At the percolation threshold, the network contains approximately 362,000 pages, a giant cluster of 57,800 pages, about $s_i \in LS_i$6 directed links, average degree approximately 1743, and 44,500 disconnected clusters. The cluster-size distribution follows $s_i \in LS_i$7, the in- and out-degree exponents are approximately 1.27–1.28, the directed diameter is 20, and the average clustering coefficient is 0.87 versus 0.17 for a random baseline (Masucci et al., 2011). Here semantic information flow is a large-scale content network rather than an execution trace.

“Delocalized and Dynamical Catalytic Randomness and Information Flow” offers a quantum-information interpretation (Lie et al., 2022). There information flow is modeled by catalytic randomness: information spreads from a source to a recipient while leaving the source’s local state invariant. The paper argues that classical information can spread without altering its source or surrounding context, whereas quantum information cannot; it also proposes an account of semantic quantum information in which using semantic information is equivalent to using a partially depleted information source (Lie et al., 2022). This is a highly abstract formulation, but it preserves the directional and resource-sensitive character seen in other literatures.

A recent cognitive synthesis appears in “Information Must Flow: Recursive Bootstrapping for Information Bottleneck in Optimal Transport” (Li, 8 Jul 2025). The Context-Content Uncertainty Principle models cognition as directed flow between high-entropy context $s_i \in LS_i$8 and low-entropy content $s_i \in LS_i$9, with joint uncertainty minimized by alternating bottom-up contextual disambiguation and top-down content reconstruction. The paper states a Delta Convergence Theorem under which recursive bootstrapping drives $aggregate(i,s_i,S_i)$0 to a delta-like attractor, and extends the framework to language as a symbolic transport system for synchronizing inference cycles across individuals (Li, 8 Jul 2025). This is a speculative and strongly unifying proposal, but it illustrates how the term can be extended from representation analysis to a general theory of cognition and collective intelligence.

Across the surveyed work, several limits recur. Some frameworks are explicitly conceptual and illustrative rather than fully formalized, as in semantic flow graphs for ML-based systems (Yoo et al., 13 Mar 2025). Some depend strongly on embedding choice, clustering stability, or the stochasticity of projection methods such as t-SNE (Yoo et al., 13 Mar 2025). Some depend on VLM or LLM judgments whose noise is acknowledged by the authors (Kaplan et al., 1 Apr 2025). Some provide only local, asymptotic guarantees, as in GIF’s small-perturbation soundness theorem (Storek et al., 22 Jun 2026). In communicative settings, causal interpretation remains contingent on assumptions such as stationarity, absence of hidden confounders, and model adequacy (Goodall et al., 29 Jun 2026). This suggests that semantic information flow is best treated as a research program with multiple formal instantiations rather than as a settled universal formalism.