Hybrid Semantic Communication

• Hybrid Semantic Communication is a family of architectures that combine task-relevant semantic transmission with conventional bit methods for enhanced network performance.
• Key architectural patterns include semantic/bit coexistence, auxiliary information integration, and cross-layer control to dynamically allocate resources and optimize throughput.
• Innovative evaluation metrics and optimization schemes, such as per-subcarrier switching and adaptive retransmission, improve reliability and efficiency in complex communication systems.

Hybrid Semantic Communication (HSC) is used in the recent literature for systems that do not rely on a single semantic transmission mechanism. In different formulations, hybridization appears as coexistence of semantic communication and conventional bit communication in the same network, semantic information inserted into conventional bit payloads, per-subcarrier switching between semantic and Shannon transmission, semantic representation supplemented by a complementary representation, and semantic/non-semantic radio access operation inside one protocol stack. This suggests that HSC is best understood as a family of architectures that couple task-relevant semantic transmission with conventional communication, auxiliary information, or cross-layer control rather than as a single standardized transceiver model (Xia et al., 2024, Yu et al., 2024, Evgenidis et al., 2024, Nam et al., 23 Jul 2025, wang et al., 10 Apr 2025).

1. Conceptual scope and terminology

Semantic communication departs from the Shannon-style objective of reliably transporting symbols irrespective of meaning. In the signaling-game formulation, the semantic objective is the correct recovery of a semantic type $T$ through a signal $S$ and a response $R$, with performance measured by the Success Rate of Semantic Agreement, $\mathrm{SRSA}=\Pr(T=\hat T)$. The decoder is explicitly context-dependent, $\phi:\mathcal{S}\times\mathcal{K}\to\mathcal{R}$, so semantic performance depends on both transmitted signals and receiver knowledge (Choi et al., 2022).

A second foundational strand makes the hybrid character explicit by separating a semantic layer from a technical communication layer. In the ProbLog-based formulation, the TC layer still handles message length, noisy channels, and retransmission cost, while the SC layer handles knowledge bases, clauses, inference, and message value. This is not a replacement of conventional communication theory; it is a two-layer architecture in which semantic reasoning determines what is worth sending and the technical layer determines how efficiently and reliably it can be sent (Choi et al., 2022).

A common misconception is that HSC necessarily means a hybrid semantic/bit coding format. Some papers use “hybrid” in a different sense. In UAV-assisted semantic communication, for example, the hybrid aspect is primarily hybrid discrete-continuous optimization for semantic communication system control, not a hybrid semantic/syntactic coding architecture. There the discrete action is channel allocation, while the continuous actions are semantic model scale, transmission power, and UAV movement (Si et al., 2023).

2. Main architectural patterns

The literature instantiates HSC through several recurring patterns.

HSC pattern	Representative mechanism	Example papers
Semantic/bit coexistence	Mode selection, same-spectrum co-transmission, semantic/non-semantic scheduling	(Xia et al., 2024, Evgenidis et al., 2024, Ahmed et al., 6 May 2025, wang et al., 10 Apr 2025, Yu et al., 2024)
Semantic plus auxiliary information	Complementary representation, check codewords, decoder/model transfer	(Nam et al., 23 Jul 2025, Li et al., 31 May 2026, Dong et al., 2022)
Semantic plus shared context or knowledge	Correlated knowledge bases, ProbLog knowledge layers, proactive knowledge sharing	(Choi et al., 2022, Choi et al., 2022, Chen et al., 3 Jan 2025)
Semantic plus human or control loop	Human decision-making interface, mixed discrete-continuous RL control	(Beck et al., 2024, Si et al., 2023)
Semantic plus hybrid media substrate	Acoustic-optical-RF meaning-driven connectivity	(Khalil et al., 19 Jan 2026)

Within the coexistence family, the most explicit network-level formulation is the hybrid semantic/bit communication network, or HSB-Net, in which each mobile user may use SemCom or BitCom and the network jointly optimizes user association, mode selection, and bandwidth allocation (Xia et al., 2024). A closely related per-subcarrier formulation is “Hybrid Semantic-Shannon Communications,” where each subcarrier chooses between semantic transmission and Shannon bit transmission under strict similarity thresholds (Evgenidis et al., 2024). At the physical-access level, hybrid NOMA serves semantic users and bit users together, using NOMA in shared slots and OMA in bit-user-only slots, with bit-to-semantic decoding order in mixed slots (Ahmed et al., 6 May 2025).

A second pattern supplements the semantic representation with additional information. “Hybrid Semantic-Complementary Transmission” defines HSC as semantic representation supplemented by a complementary representation that captures residual image-specific information, with controllable fidelity through the complementary load parameter $d$ (Nam et al., 23 Jul 2025). The reliability perspective likewise proposes joint source-channel-check coding and adaptive retransmission, where check codewords are transmitted as complementary information rather than as conventional parity detached from semantics (Li et al., 31 May 2026). A different variant transmits not only semantic code but also semantic decoder parameters through a conventional digital link, making the model itself part of the communicated object (Dong et al., 2022).

3. Theoretical foundations and evaluation metrics

The most compact information-theoretic statement of semantic hybridization appears in the signaling-game model with correlated knowledge bases. Under error-free semantic encoding, the receiver’s recovered semantic information satisfies

$I(T;\hat T)=H(S)+I(K_A;K_B\mid S),$

which separates explicit signaling content $H(S)$ from the gain contributed by correlated knowledge $I(K_A;K_B\mid S)$. In this view, semantic recovery can exceed the entropy of the explicit signal because part of the meaning is resolved by receiver-side context rather than by transmitted symbols alone (Choi et al., 2022).

The ProbLog-based two-layer formulation introduces semantic measures defined over knowledge bases rather than raw source symbols. Clause entropy is

$H_f(c)=-(p_c\log p_c+(1-p_c)\log(1-p_c)),$

knowledge-base uncertainty is

$S$0

and the semantic content of a message is

$S$1

This makes message value a function of inference and knowledge-base improvement rather than of symbol occurrence alone (Choi et al., 2022).

When semantic and bit transmissions coexist, unified performance metrics become necessary. HSB-Net defines a unified message rate

$S$2

where semantic throughput depends on a bit-rate-to-message-rate transformation and knowledge matching, while bit throughput is converted to message throughput through a bit-to-message coefficient (Xia et al., 2024). Hybrid NOMA defines an equivalent semantic rate for the bit user,

$S$3

so that semantic and bit users can be optimized in a common semantic-efficiency space (Ahmed et al., 6 May 2025). Knowledge sharing-enabled HSC in multi-cell networks defines a generalized effective semantic transmission rate $S$4 that combines semantic information, semantic accuracy, and the residual bit payload under one objective (Chen et al., 3 Jan 2025).

Resource-control formulations introduce additional mixed semantic-classical QoS measures. SBQ-compatible adaptive resource allocation defines semantic quantization efficiency, latency, and the overall SC-QoS objective

$S$5

which directly couples semantic quality per bit to transmission delay (Wang et al., 2023). Reliable semantic communication adds lower-tail metrics such as semantic distortion outage probability and tail PSNR, explicitly arguing that mean PSNR or mean task accuracy are insufficient to characterize reliability under adverse conditions (Li et al., 31 May 2026). In underwater IoT, semantic efficiency is summarized by

$S$6

linking conveyed semantic information to bandwidth-time resource consumption (Khalil et al., 19 Jan 2026).

4. Cross-layer optimization and control

A large portion of HSC research is organized around joint optimization across semantic, wireless, and computing variables. In HSB-Net, the network jointly solves user association, mode selection, and bandwidth allocation under average queuing latency requirement $S$7, average packet loss ratio requirement $S$8, and minimum message-throughput requirement $S$9. The key modeling device is a knowledge matching-aware two-stage tandem queue for SemCom links, which differs structurally from the single transmission queue of BitCom links (Xia et al., 2024).

In multi-carrier Hybrid Semantic-Shannon communication, each subcarrier chooses either Shannon mode or semantic mode through binary variables $R$0 with $R$1, and the subcarrier delay is

$R$2

Semantic feasibility is governed by the sentence-group threshold $R$3 and the DeepSC saturation level $R$4, so semantic mode is admissible only when the required similarity can be met (Evgenidis et al., 2024).

Knowledge sharing-enabled task-oriented HSC introduces a different control knob, the semantic extraction ratio $R$5, together with the current matched knowledge set $R$6. The total completion time is the sum of knowledge upload time, semantic transmission time, bit transmission time for residual unmatched data, semantic reconstruction time, and raw-data execution time. The optimization jointly selects SBS association, knowledge sharing, and $R$7 to maximize generalized effective semantic transmission rate under delay and semantic-accuracy constraints (Chen et al., 3 Jan 2025).

The SBQ-compatible adaptive resource allocation paradigm pushes the hybridization down to the semantic-bit interface. The base station jointly optimizes transmit beamforming, semantic-bit depth, subchannel assignment, and bandwidth allocation. The paper formulates a non-convex SC-QoS maximization problem and solves it by a hybrid DRL scheme whose agent perceives both semantic tasks and dynamic wireless environments, with up to 13% SC-QoS improvement over mapping-guided resource allocation schemes (Wang et al., 2023).

Hybrid action can also refer to the structure of the control space itself. In UAV-assisted semantic communication, the continuous action is

$R$8

where $R$9 is semantic model scale, $\mathrm{SRSA}=\Pr(T=\hat T)$0 is transmit power, and $\mathrm{SRSA}=\Pr(T=\hat T)$1 are UAV movements, while the discrete action is channel allocation. The system therefore couples semantic fidelity control, physical-layer allocation, and platform mobility in one mixed action space (Si et al., 2023).

5. Reliability, protocol integration, and deployability

Reliability-oriented work extends HSC beyond average performance by combining semantic coding with conventional wireless control loops. A recent perspective organizes reliable semantic communication into three categories: channel-aware adaptation, robustness-oriented codec design, and HARQ-based retransmission. It then proposes two directions, robust adaptive semantic communication under imperfect CSI and joint source-channel-check coding with adaptive retransmission, and argues for reliability metrics beyond averages and compatibility with existing digital wireless networks (Li et al., 31 May 2026).

A concrete HARQ realization appears in SNN-SC-HARQ for collaborative intelligence. There, intermediate features are transmitted semantically, but the number of SNN time steps is allowed to vary. A policy model estimates semantic similarity after each accumulated transmission and sends ACK/NACK so that bandwidth can be increased incrementally only when needed. The reported result is that SNN-SC-HARQ can dynamically adjust the bandwidth according to the channel conditions without performance loss (Wang et al., 2023).

Several papers focus explicitly on deployment over existing digital stacks. HybridBSC inserts encoded semantic information into bit information for transmission via conventional digital communication systems utilizing the same spectrum resources. In a pluto-based SDR experiment over a real wireless channel, the average recovered quality of the bit images is PSNR $\mathrm{SRSA}=\Pr(T=\hat T)$2 and SSIM $\mathrm{SRSA}=\Pr(T=\hat T)$3, while the average recovered quality of the semantic images is PSNR $\mathrm{SRSA}=\Pr(T=\hat T)$4 and SSIM $\mathrm{SRSA}=\Pr(T=\hat T)$5 (Yu et al., 2024).

At the protocol-stack level, HSC-RAN extends a 6G-style downlink stack by adding a service classification layer and a semantic layer, partitioning radio resources into semantic RBs and non-semantic RBs, and introducing a 1-bit DCI field ResourceType with $\mathrm{SRSA}=\Pr(T=\hat T)$6 for non-semantic and $\mathrm{SRSA}=\Pr(T=\hat T)$7 for semantic resources. In the reported demo, semantic real-time video and non-semantic text are transmitted simultaneously through a PDSCH/OFDM-based framework, with Channel Bandwidth Ratio $\mathrm{SRSA}=\Pr(T=\hat T)$8 approximately for each I-frame, average PSNR $\mathrm{SRSA}=\Pr(T=\hat T)$9, and average MS-SSIM $\phi:\mathcal{S}\times\mathcal{K}\to\mathcal{R}$0 (wang et al., 10 Apr 2025).

A further deployability issue is decoder availability. One image-oriented semantic system addresses this by transmitting not only semantic code $\phi:\mathcal{S}\times\mathcal{K}\to\mathcal{R}$1 but also decoder parameters $\phi:\mathcal{S}\times\mathcal{K}\to\mathcal{R}$2, so that the communicated object is already hybridized into semantic data plus model information over a conventional digital link (Dong et al., 2022).

6. Applications, misconceptions, and open directions

The application space is broad and highly heterogeneous. In downlink heterogeneous access, one access point supports one bit user and two semantic users through hybrid NOMA, using NOMA in shared slots and OMA in bit-user-only slots, with equivalent ergodic semantic spectral efficiency as the objective (Ahmed et al., 6 May 2025). In human-in-the-loop systems, semantic communication is integrated with human decision-making through the probabilistic chain

$\phi:\mathcal{S}\times\mathcal{K}\to\mathcal{R}$3

showing a trade-off between maximizing relevant semantic information and matching the cognitive capabilities of the HDM model (Beck et al., 2024). In underwater IoT, hybrid acoustic-optical-RF architectures and edge-intelligent semantic encoders are presented as enablers of sustainable, adaptive operations, with examples in underwater archaeology, marine ecology, and AUV coordination (Khalil et al., 19 Jan 2026).

The literature also corrects several simplistic interpretations of HSC. HSC is not equivalent to “send more semantic detail.” In human-aware semantic communication, richer intermediate features can saturate and, under limited expertise, fewer features can outperform more features. HSC is also not reducible to “semantic plus bits” in a single coding layout. It can mean proactive knowledge sharing plus residual bit transmission, mixed control spaces, or complementary residual transmission. In high-fidelity image reconstruction, for example, semantic representation alone can saturate at a nonzero MSE even when its load exceeds the original image size, whereas complementary representation provides an explicit fidelity-load control mechanism through the parameter $\phi:\mathcal{S}\times\mathcal{K}\to\mathcal{R}$4 (Chen et al., 3 Jan 2025, Nam et al., 23 Jul 2025).

Open problems are correspondingly cross-disciplinary. Reliable semantic communication still needs joint robustness and retransmission design, stronger reliability metrics beyond averages, and explicit compatibility with existing digital wireless networks (Li et al., 31 May 2026). Other recurring directions include semantic representation standardization, cross-domain interpolation, privacy-support schemes, better human models, joint optimization of machine communication and human decisions, and more scalable hybrid-action learning with stronger theoretical grounding. Taken together, these directions indicate that HSC is evolving from isolated semantic transceivers into a systems field concerned with interoperability, control, reliability, and task execution across heterogeneous networks and receivers.