Semantic-Aware Multicast Decomposition

• The paper introduces semantic-aware multicast decomposition, which splits source data into semantically meaningful components to significantly reduce redundant transmissions.
• It employs adaptive coding, power allocation, and receiver-specific segmentation to dynamically match transmission strategies with diverse quality of service requirements.
• Experimental evaluations reveal improvements such as a 15.4% latency reduction, 80.7% fewer transmissions in control scenarios, and up to a 10% boost in spectral efficiency.

Semantic-aware multicast decomposition is an advanced class of methodologies in multiuser wireless communications that leverages semantic-level information to optimize multicast transmission. Instead of transmitting bitstreams representing source data, semantic-aware multicast decomposition partitions information into semantically meaningful components, dynamically matches these to user interests or tasks, and adapts transmission strategies—including coding, power allocation, and scheduling—at the segment, class, token, or bit level. This enables significant gains in efficiency, perceptual quality, and user-specific utility, particularly in bandwidth- and latency-constrained scenarios.

1. Conceptual Foundations and System Modeling

Semantic-aware multicast decomposition fundamentally redefines what is transmitted in multiuser scenarios by prioritizing meaning and utility over strict bitwise fidelity. The canonical architecture involves a transmitter with access to a source signal $X$ (e.g., an image, control vector, or multi-modal data) and $K$ or $N$ receivers, each with potentially diverse tasks, QoS requirements, or semantic interests.

• Semantic Decomposition: The source $X$ is partitioned into $L$ disjoint semantic components $X_l$, for $l=1,...,L$. For visual data, this is achieved by applying a segmentation network or a generative AI model ("foundation model") to yield a semantic map $X_0$ and segment representations $\{X_l\}$. Control data, as in UAV systems, are structured into low-dimensional vectors per command dimension (Liu et al., 2024, Li et al., 29 Jan 2026, Ma et al., 22 Feb 2025).
• Intent or Interest Modeling: Receiver requirements are captured by matrices or functions such as an intent matrix $I \in \{0,1\}^{K \times L}$, where $K$0 if user $K$1 is interested in class/segment $K$2.
• Wireless Channel Model: Standard block-fading Rayleigh or AWGN models are employed, with user-dependent SNRs or channel gains, and global constraints on power, rates, or resource blocks (Liu et al., 2024, Ma et al., 22 Feb 2025).

2. Semantic Decomposition and Receiver-specific Granularity

Decomposition into semantically meaningful entities proceeds with pre-trained segmenters (e.g., DDRNet, Segment Anything Model), forming:

• Pixel or Token Masks: For images, a one-hot semantic mask $K$3 for each pixel $K$4 is collapsed into binary masks $K$5 indicating class $K$6: $K$7 iff $K$8 class $K$9; thus $N$0 (Liu et al., 2024).
• Command Vectors: For multi-agent control, each user's $N$1-th time interval command is $N$2 (Li et al., 29 Jan 2026).
• Tokenization: Semantic segments are further encoded into tokens $N$3 for each $N$4, supporting fine-grained transmission strategies (Ma et al., 22 Feb 2025).
• Importance Quantification: Each segment or token is assigned a distortion tolerance $N$5 or $N$6 (receiver-specific), and a weighting reflecting the segment's impact on utility or task completion (Ma et al., 22 Feb 2025).

3. Semantic Multicast Grouping and Synthesis

Semantic-aware multicast decomposition includes mechanisms for forming multicast groups and synthesizing missing or non-intended semantic components at the receiver:

• Group Formation: Receivers (e.g., UAVs) are grouped into multicast sets $N$7 such that for each pair $N$8 in group $N$9, their semantic difference $X$0 does not exceed a threshold $X$1, where $X$2 (Li et al., 29 Jan 2026).
• Synthesis at Receiver: For image multicasting, receivers reconstruct intended segments directly and synthesize non-intended segments from a compressed semantic map $X$3 using a generative diffusion model $X$4. The final received reconstruction for user $X$5 is a fusion $X$6 (transmitted) and $X$7 (synthesized) (Liu et al., 2024).

4. Joint Resource Allocation and Optimization

A central research emphasis lies in jointly optimizing communication and computation resources to minimize latency, adhere to per-user distortion/perceptual requirements, and maximize overall utility. Typical approaches include:

• Rate and Power Allocation: Rates $X$8 and powers $X$9 are adapted per segment/class. The total constraint is $L$0, with per-class constraints on reconstruction $L$1 and overall perceptual quality $L$2. Data rates are derived from Shannon capacity, e.g., $L$3 (Liu et al., 2024).
• Latency Minimization: For each receiver $L$4, total retrieval time is $L$5, where $L$6 is the computation/diffusion latency (Liu et al., 2024).
• Optimization Objectives: Typical problems are posed as minimizing the sum $L$7 subject to latency, power, and quality constraints, using sequential quadratic programming or reinforcement learning-driven policies (e.g., PPO) (Liu et al., 2024, Li et al., 29 Jan 2026).
• Rate-Splitting and Progressive Transmission: For scenarios with receiver-specific importance, segments are rate-split into progresssive layers, leveraging Slepian–Wolf coding and real-time channel feedback to determine which layers are transmitted to which users (Ma et al., 22 Feb 2025).

5. Transmission and Coding Strategies

Semantic-aware multicast decomposition operates at multiple granularity levels:

• Segment-Level Importance: Each segment is compressed independently to a target distortion; all users receive the same representation unless receiver-specific codes are used (Ma et al., 22 Feb 2025).
• Receiver-Specific Importance: Segments are layered; each receiver decodes only as many layers as required to meet its individual $L$8. Power and error-correction resources are allocated accordingly (Ma et al., 22 Feb 2025).
• Token- and Bit-Level Importance: Tokens are classified and coded based on criticality. Strong FEC (e.g., polar or LDPC codes) is applied to critical tokens, weaker protection to more robust tokens, and unequal error protection (UEP) is achieved by partitioning MSBs and LSBs (Ma et al., 22 Feb 2025).
• Transmission Triggers and Skipping: For time-series or control applications, transmissions are only triggered if semantic drift $L$9 exceeds thresholds; otherwise, commands are skipped, reducing redundant transmissions (Li et al., 29 Jan 2026).

6. Evaluation and Performance Insights

Across diverse tasks—generative image multicast, multi-UAV command delivery, and multi-modal semantic communication—semantic-aware multicast decomposition achieves significant improvements over bit-oriented baselines:

• Latency Reduction: In generative multicasting with diffusion synthesis, average per-user latency is reduced by ≈15.4% compared to non-generative multicast, and operating at a given latency may require up to 50% less power than conventional approaches (Liu et al., 2024).
• Spectral Efficiency: Gains up to 10% are reported for larger user counts or when semantic intents overlap (Liu et al., 2024).
• Transmission Reduction: In control-centric multi-UAV applications, the framework achieves ≈80.7% fewer physical transmissions than bit-oriented schemes, while maintaining high service effectiveness (Li et al., 29 Jan 2026).
• Distortion Management: Importance-driven allocation preserves high-fidelity segments or critical command tokens at the expense of less important elements, as verified in toy experiments with dynamic segment- and bit-level coding (Ma et al., 22 Feb 2025).

7. Research Directions and Theoretical Guarantees

Current research emphasizes:

• End-to-End Joint Optimization: While piecewise rate-distortion and convex optimization routines are directly implementable (e.g., (Liu et al., 2024)), the rigorous mathematical characterization of multiuser, importance- and semantic-aware multicast, especially under dynamic or feedback-driven adaptation, remains an open direction (Ma et al., 22 Feb 2025).
• Model-based Grouping and Scheduling: Deep reinforcement learning (especially PPO) and MDP formulations are being utilized to jointly schedule transmission modes, RB allocation, and group formation towards long-term QoS maximization (Li et al., 29 Jan 2026).
• Segment- and User-adaptive Protection: Integration of progressive coding, layered multicast, and UEP at semantic granularity points toward highly flexible paradigms, with room for further exploration of tradeoffs between bandwidth, computation, and utility (Ma et al., 22 Feb 2025).

These trends underscore the centrality of semantic-aware multicast decomposition as an enabling principle for next-generation wireless systems, especially those leveraging generative models and task-specific, multi-user resource sharing (Liu et al., 2024, Li et al., 29 Jan 2026, Ma et al., 22 Feb 2025).