A Survey of Optimal Resource Allocation in Semantic Communications: Technologies, Development Trends, and Applications

Abstract

As an emerging communication paradigm, semantic communications (SC) focuses on the semantic content of information transmission, aiming to achieve more efficient and accurate information interaction. However, SC requires an in-depth analysis of information semantics and accurate adaptation to application scenarios. With the substantial growth in information volume, richness, and application diversity, the resources, constraints, and requirements associated with SC systems have also increased. Optimal resource allocation (ORA) of SC can effectively deal with these practical problems, a key technology for improving communication efficiency. In complex network environments and special scenarios, communication resources are limited. Increasing transmission efficiency, ensuring accuracy and quality of information, and reducing energy consumption can be achieved by rationally allocating bandwidth, power, and related resources. Since SC technology is still in its early stages of development, there is a lack of a comprehensive review of ORA for SC in the existing literature. This paper provides a comprehensive review of ORA for SC. First, the basic concepts, characteristics and development motivations for SC and ORA are reviewed. Then, a comprehensive analysis of the key technologies for ORA in SC is presented, covering end-to-end and semantic network multi-link ORA. These technologies include technology for predicting resource demand based on semantic understanding, resource optimization technology for semantic information (SI) processing and transmission, and technology for dynamic resource adjustment. Then, a conceptual ORA framework is synthesized from the reviewed technologies to unify key design principles and provide a foundation for future research. In addition, this paper provides research prospects in future trends of ORA for SC, comprising emerging artificial intelligence (AI) and machine learning, laying the foundation for next-generation intelligent communication networks. Finally, this paper points out the main application direction of ORA in SC, which reflects the practical significance of this study.

1. Introduction

The development of technologies such as the Internet of Things (IoT) and artificial intelligence (AI) has led to an explosion of data, overwhelming the capabilities of traditional communications (TC) systems. Rooted in Shannon’s information theory, classical TC systems are primarily designed to accurately transmit symbols, rather than interpret or act on the meaning of the information. It should be acknowledged that advanced conventional systems have increasingly incorporated application-aware scheduling, QoS/QoE-driven resource allocation (RA), and cross-layer optimization to better serve intelligent applications. Nevertheless, these enhancements still operate within the bit-centric paradigm, where the ultimate measure of successful transmission remains symbol fidelity rather than the utility of the conveyed meaning. As intelligent applications and complex communication environments become increasingly prevalent, the limitations of this purely syntactic focus are becoming more evident.

Figure 1 depicts a simplified schematic diagram of a TC system [1]. The TC concept focuses on the grammatical level of the transmission of information—ensuring that symbols are accurately conveyed—but it often neglects the semantic content those symbols represent [2]. In practical applications, users are generally more concerned with the meaning of the information than with the mere accuracy of transmitted symbols. TC methods, which lack semantic understanding, often struggle to meet real-world application needs. This shortcoming can lead to the transmission of redundant information, resulting in increased bandwidth consumption, system burden, latency, and degraded performance.

With the rise of task-driven applications, TC mechanisms increasingly face challenges such as limited bandwidth, high transmission delays, and inefficient resource utilization [3]. Although conventional systems have developed task-oriented and priority-based RA strategies (e.g., differentiated services, adaptive modulation, and content-aware caching), these approaches still treat all bits as equally important at the fundamental level and lack an explicit semantic understanding of information content. Consequently, transmitting high-value data under these constraints remains difficult, as modern systems often send all available information without prioritizing task relevance or semantic importance, wasting resources.

In this context, semantic communications (SC) has emerged as an effective solution to the limitations of TC systems. SC is task-oriented and follows a “comprehend-then-transmit” approach. It aims to extract, compress, encode and transmit the features of the original signal based on a shared “semantic consensus” between the sender and receiver. Communication is then achieved using information at the semantic level, rather than merely transmitting raw symbols. A simplified schematic of an SC system is illustrated in Figure 2 [1]. Unlike physical noise that corrupts symbols during transmission, semantic noise arises primarily from the mismatch between the sender’s and receiver’s shared knowledge representations. Consequently, maintaining knowledge-base alignment is as critical for reliable SC as combating channel fading is for TC. Compared to TC, SC emphasizes the meaning of information, effectively reduces redundancy, significantly enhances communication efficiency, and enables more efficient resource utilization along with higher-quality communication services. While TC methods prioritize the speed of information transmission and the integrity of data packets, SC focuses on conveying the true intent or meaning behind the information [4]. This paradigm shift not only minimizes unnecessary data transmission but also ensures that the most relevant and valuable information is delivered using limited resources.

Building upon this foundational advantage, another significant reason to study SC is its potential to meet the increasingly complex communication demands. By processing information semantically, SC can filter valuable data in big data environments, enhancing communication efficiency. It enables intelligent integration and smart communication, offering precise semantic support for 6G applications. SC meets diverse task-specific needs, empowering 6G networks with superior performance and diversified applications, driving 6G development [5]. As a result, SC is emerging as a key technology for next-generation communication systems, enabling more efficient resource management and information transmission, and has become a prominent research focus.

In SC systems, RA extends beyond traditional grammar-level considerations to include bandwidth, computing power, time, storage, and energy, focusing on information content and importance. This makes RA for SC more promising, effective, and complex. Unlike traditional methods targeting transmission networks or human users, SC emphasizes semantic understanding between intelligent agents. The core objective of future SC systems is to transition from traditional “bit-level data transmission” to “efficient semantic information (SI) transmission and understanding,” and its RA prioritize task-driven and goal-oriented optimization. While TC systems focus on physical-layer metrics like channel capacity, SC emphasizes semantic extraction, transmission, and interpretation for intelligent tasks. Therefore, RA strategies based solely on service quality (QoS) or user experience (QoE) are no longer optimal.

To address this, an optimal resource allocation (ORA) model tailored for SC must be developed [6]. Under limited resource conditions, SC performance can be enhanced by optimizing allocation strategies—improving transmission efficiency, reducing delays, and increasing reliability. This involves multi-objective, multi-dimensional decision-making, requiring dynamic adjustment based on SI, channel state, QoS requirements, and other factors to achieve optimal network resource utilization.

Currently, the most common approaches include reinforcement learning (RL), game theory-based competition mechanisms, and optimization theory-based models [7]. RL learns optimal strategies through trial and error, while game theory focuses on multi-agent interactions to achieve Nash equilibrium. In multi-agent networking scenarios, these paradigms can be formally modeled as Decentralized Partially Observable Markov Decision Processes (Dec-POMDPs) or Partially Observable Markov Games (POMGs), which provide mathematical frameworks for distributed decision-making under uncertainty [8]. With the ongoing advancement of AI, SC systems face increasing demands for real-time performance and reliability, requiring more adaptive and efficient resource management strategies [9]. This involves optimizing allocation algorithms, enhancing semantic understanding through structured representations for better resource matching, and improving communication architectures, protocols, and evaluation frameworks to support efficient SC systems.

Existing review papers on SC primarily focus on three key dimensions: basic theoretical research, key enabling technologies, and practical applications.

• Basic theoretical research
  Several works have laid the theoretical foundation of SC. In [10], the theoretical landscape of SC is systematically reviewed, including the core concepts of semantic entropy, semantic rate distortion, and semantic channel capacity. The study also discusses challenges such as quantifying SI, developing task-oriented communication frameworks, and understanding the effects of semantic noise—emphasizing that SC is still in its early developmental stage. Similarly, ref. [11] highlights the fundamental shift in SC, where AI technologies are used to extract and transmit semantics rather than raw bits. The paper compares SC and TC in terms of architecture, performance metrics, and resilience to noise. In [12], the authors summarize the progress of SI theory and SC systems, proposing a semantic-based transmission framework and discussing potential 6G applications. These studies primarily focus on theoretical explorations in SC, as well as AI-enabled semantic extraction frameworks and 6G network visions. However, these studies still exhibit certain limitations. For instance, the theoretical framework remains incomplete, especially the theory related to ORA in SC. There is a lack of universal performance evaluation standards, as well as insufficient research on multi-user scenarios and complex tasks.
• Key enabling technologies
  Research has also addressed SC-enabling technologies. In [13], the authors review advancements in multi-modal SC, covering critical technologies like multi-modal data fusion and secure SC, and explore its potential within 6G networks. Paper [14] focuses on semantic knowledge bases, discussing their role in semantic encoding/decoding, cross-protocol optimization, and resource management, while highlighting performance indicators and key design challenges. In [15], the role of large language models (LLMs) in enabling SC is examined, along with associated challenges in computing resources, generalization, and data security. Paper [16] investigates security threats in SC and explores protection mechanisms including federated learning, adversarial defense, and semantic encryption. Meanwhile, ref. [17] offers a comprehensive overview of SC’s evolution and engineering methodologies, categorizing them into four approaches: classical SI theory, knowledge graphs, machine learning, and information saliency. A context-aware SC framework is also proposed to enhance communication efficiency. Furthermore, ref. [18] reviews deep learning (DL) applications at the physical layer, showing how data-driven methods enable end-to-end SC for multi-modal transmission. The above research has primarily focused on investigating key technologies in SC. However, studies remain insufficient regarding prediction, processing, and transmission techniques for ORA in SC. Furthermore, research on multimodal semantic fusion lacks a unified framework, and investigations into dynamic updating of knowledge bases and distributed coordination mechanisms are also inadequate.
• Application-oriented research
  Application-focused reviews emphasize SC’s integration into real-world systems. Paper [19] presents a comprehensive survey that establishes SC as a paradigm shift from conventional “data pipe” transmission to context-aware meaning exchange. Their work critically analyzes major technological trends and use cases, exposes fundamental theoretical and implementation challenges, and delineates concrete future research directions. In [20], SC is applied to the internet of vehicles (IoV), covering architecture, communication technologies, and its role in traffic perception and intelligent driving. The paper in [21] explores the integration of SC with edge computing, analyzing semantic extraction when combined with edge computing’s low latency and high privacy, ORA and task offloading in domains like industrial IoT and IoV. Additionally, ref. [22] discusses SC’s potential in enabling intelligent tasks within the context of 6G-AI convergence. Current research on SC applications primarily focuses on practical implementations in typical domains such as vehicular networks, edge computing networks, and industrial Internet systems. However, research on application scenarios incorporating ORA remains scarce.

In the classical Shannon communication paradigm, the core objective of resource optimization is to achieve error-free and high-rate transmission of bit-streams under given channel capacity constraints. Its optimization criteria are typically based on physical layer performance metrics, such as bit error rate or spectral efficiency. However, the fundamental transformation of the SC paradigm lies in the fact that the value of its transmission is no longer the fidelity of bits, but rather the degree to which the semantics conveyed by information of specific tasks are successfully extracted and utilized. This paradigm shift endows ORA with entirely new connotations and strong motivations. Meanwhile, this field is undergoing rapid technological evolution. Therefore, a systematic review paper is needed to sort out the key technologies in ORA strategies for SC, gain insights into its development trends, and integrate this highly interdisciplinary research field has become increasingly meaningful. Given the current research landscape, existing articles lack a comprehensive overview of the key technologies, development trends, and practical applications of ORA in SC, as shown in

Table 1. Summary and comparison of review papers on SC.

Survey	Theory	Technology	Application	Future	ORA
[9,11]	★ Theoretical framework of SI theory	▲ Key technologies for realizing the theory	• Transmission of images, texts, and audio	• Theory, system and practice	/
[10]	★ The overall framework of SC, and an in-depth discussion on specific mathematical theories such as “semantic entropy”	★ Semantic encoding and decoding, knowledge bases, DL	/	• Theories, protocol design, integration of emerging technologies, etc.)	/
[12,14,17]	• Basic theory of SC and end-to-end (E2E) communication systems	★ Latest technologies such as multi-modal data fusion, secure, E2E communication systems and those based on large language models	▲ Explosive data transmission, human-machine communication, and communication in harsh environments	• Technical optimization and deepening of application scenarios	/
[13,15]	• The basic principles of semantic knowledge bases, SC networks, and the three-layer architecture of multi-agent interaction	★ Knowledge base-generative model collaborative architecture, semantic-enhanced generation and transmission technology, knowledge base dynamic update and adaptation technology, and core architecture technology	/	• Deepening the integration of multi-modal knowledge bases and generative models, and optimizing the energy efficiency of SC networks	/
[16]	▲ The four classic theoretical paths of SC, and the new context-based SC theory	★ Classic SI technology, knowledge graph-related technologies, DL technology, information importance evaluation technology, and contextual SC technology	/	• The design and implementation directions of SC systems	/
[18,21]	• The development history and core basic theories of SC	▲ Semantic transmission optimization technology, DL support technology and intelligent task-oriented SC architecture	★ Core application scenarios of SC (major use cases driven by 6G)	• The potential of SC in supporting intelligent tasks	/
[19]	• Basic theory of vehicle networking	▲Communication protocol, wireless signal transmission optimization technology and communication security encryption technology	★ Intelligent networking automobile and vehicle networking	• Technology development and application landing	/
[20]	• Edge computing scene adaptation theory	▲The core technology of SC in edge scene and the fusion technology of computing network and SC	★ The application of task scenarios in internet of things (IoT)	• Scene expansion and technology integration	• Optimization technology of bandwidth and computing power resources
This work	▲ The basic concepts and characteristics of SC and ORA, as well as the development motivation for ORA	★ A comprehensive analysis of the key technologies for ORA in SC, covering end-to-end and semantic network multi-link ORA	▲ Intelligent unmanned devices, smart cities, etc.	• Technological innovation, model establishment	★ Theories, technologies, applications, and future research are all based on ORA, and a conceptual ORA framework is synthesized

Note: “★” denotes primary research focus, “▲” denotes secondary research focus, “•” indicates mentioned in the paper and “/” indicates not covered in the paper.

. To further clarify the positioning and unique contribution of this survey,

Table 2. Comparison of representative existing SC review papers.

Survey	SC	ORA	Optimization Techniques	AI/ML-Based Methods	Dynamic ORA	Application Scenarios	Future Research Directions
[10]	Yes	No	Partial	Yes	No	Partial	Yes
[11]	Yes	No	No	Yes	No	Yes	Yes
[12]	Yes	Partial	No	Yes	No	Yes	Yes
[13]	Yes	Partial	No	Yes	No	Yes	Yes
[14]	Yes	Partial	Partial	Yes	Partial	No	Yes
[15]	Yes	No	No	Yes	No	Partial	Yes
[16]	Yes	Partial	Partial	Yes	Partial	Partial	Yes
[17]	Yes	No	Partial	Yes	No	Yes	Yes
[18]	Yes	Partial	Partial	Yes	Partial	Yes	Yes
[19]	Yes	Partial	Partial	Yes	Partial	Yes	Yes
[20]	Yes	Partial	Partial	Partial	Partial	Yes	Yes
[21]	Yes	Partial	Partial	Partial	Partial	Yes	Yes
[22]	Yes	Partial	Partial	Yes	Partial	Yes	Yes
This work	Yes	Yes	Yes	Yes	Yes	Yes	Yes

Note: Yes: fully and systematically covered; partial: briefly mentioned without systematic review; no: not addressed in the survey.

provides a multi-dimensional comparison with the most relevant existing review papers. This paper presents a focused review of the latest research progress on ORA in SC. The main contributions of this paper are as follows:

• A chronological overview of SC development is presented to help readers clearly understand the historical evolution and research trajectory of SC.
• Key ORA technologies for SC are systematically summarized and reviewed, covering end-to-end ORA and semantic network multi-link ORA. These technologies include resource demand prediction, joint resource optimization, and dynamic adaptive resource adjustment. Based on this analysis, an integrative conceptual ORA framework is presented to unify the reviewed technologies and provide a foundation for future research.
• Emerging development trends in ORA for SC are thoroughly analyzed, enabling researchers to anticipate and align with future directions. These trends include the deep integration of AI technologies, semantic-centric networking, and advances in signal processing.
• A systematic analysis of ORA applications in representative SC scenarios is provided, focusing on domains such as intelligent unmanned equipment, smart city infrastructure, and intelligent transportation. This contributes to strengthening both theoretical and practical research into ORA for SC.

The remainder of this paper is structured as follows. Section 2 outlines the process of SC development and briefly describes the related concepts of SC and ORA. Section 3 summarizes and reviews the state-of-the-art technologies for ORA in SC. Section 4 describes the development trend in ORA for SC. Section 5 analyzes the application cases for ORA in SC. Section 6 proposes suggestions for future research directions. Finally, conclusions are drawn in Section 7. An overview of the paper’s structure is illustrated in Figure 3.

2. Overview of Semantic Communications and Optimal Resource Application

2.1. Development of Semantic Communications

The development of SC can be traced back to Charles Morris’s theory of semiotics proposed in 1938, which introduced the foundational concepts of grammar, semantics, and pragmatics [23]. Grammar addresses the formal structure and relationships between symbols without considering their meanings or usage contexts. Semantics focuses on the relationship between symbols and the objects they represent, while pragmatics examines the interaction between symbols and their users. This triadic framework provides a theoretical basis for SC. In 1948, Claude Shannon introduced information theory, which emphasized the grammatical level of information transmission, particularly encoding and channel capacity [24]. Although it did not explicitly incorporate semantics, it established the mathematical foundations necessary for information quantification and transmission. Building on Shannon’s work, Weaver proposed a three-level communication model in 1949, explicitly identifying “semantics” as a core issue in communication for the first time [25], thereby charting a direction for SC research. From 1949 to 1990, SC remained in a theoretical germination stage. In 1952, Bar-Hillel and Carnap formally defined SI from a logical perspective, proposing the quantification of meaning using logical probability and exploring semantic entropy based on Shannon’s entropy [26]. These early contributions laid important groundwork for SC.

From 1990 to 2010, SC development was a slow exploration period, which mainly focused on the construction of theoretical frameworks and initial attempts to apply basic technologies. During this period, the Semantic Web was proposed, which emphasized machine-understandable semantic data and pushed Internet resource retrieval from the grammatical level to the semantic level, laying an important foundation for SC [27]. As wireless communication technology develops from 5G to 6G, people generally believe that the technical potential of 5G for traditional grammatical communication has been exhausted, and there is an urgent need to improve communication efficiency through conceptual innovation. As a result, researchers began to pay attention to communication issues at the semantic level, but due to the technical conditions at the time, actual progress was relatively slow. Since 2010, fueled by advances in DL and artificial intelligence, SC has gained unprecedented attention. The period from 2010 to 2015 marked a critical transition from theoretical exploration to technical validation. During this stage, researchers challenged the traditional bit-centric model of Shannon theory and began emphasizing the transmission of meaning. The basic framework for SI theory was gradually established, while new approaches such as knowledge-driven communication [28] and structured semantic coding [29] were introduced to enhance SI compression and transmission. Nonetheless, due to the limited accessibility of DL tools and insufficient computational resources at the time, critical issues such as ORA and real-time semantic interaction remained largely unresolved at the theoretical level. From 2015 to 2020, SC entered a phase of rapid development. Breakthroughs in DL significantly enhanced the capabilities for SI extraction and transmission, facilitating the initial application of SC in domains such as intelligent interaction and the Internet of Things. Since 2020, SC has entered a stage of explosive growth. With the emergence of large-scale models and the acceleration of 6G research, SC has reached a new level of advancement. Current research focuses on cutting-edge areas including multi-modal semantic understanding, lightweight semantic coding, and joint semantic-physical layer design. These developments have enabled validation in practical scenarios such as the IoV and autonomous driving.

As next-generation communication technologies continue to evolve, SC is increasingly recognized as a key enabling technology, driving the field toward greater intelligence and efficiency. Correspondingly, ORA in SC has become a popular research topic, with the aim of conserving resources, reducing energy consumption, and improving transmission rate and quality—thus amplifying the benefits of SC. Concurrently, advancements in this area are contributing to the refinement of SC’s theoretical underpinnings and system architectures. The main research directions associated with the evolution of SC are illustrated in Figure 4.

2.2. Characteristics of Semantic Communications

The advancement of 6G mobile communication technology has broadened application scenarios and established SC as one of its key enabling technologies. While continual improvements in signal modulation techniques at the grammatical level have contributed to communication performance, they have also introduced increasing technical complexity and placed greater constraints on environmental and system resources. This highlights an urgent need to transform the mode of information generation and transmission to address these challenges effectively. SC, focused on SI, differs fundamentally from traditional communication, as outlined in

Table 3. Comparison between TC and SC.

Dimensions	TC	SC	Essential Differences
Transmission Object	Bits/symbols	Semantics/intention	Symbols → meaning
Core Objective	Error-free transmission	Accurate semantic understanding	Accuracy → efficacy
Coding Method	Syntactic coding	Semantic coding	Signal optimization → knowledge compression
Efficiency	Depends on channel coding	Depends on semantic compression and knowledge sharing	Physical layer → cognitive layer
Robustness	Resistant to physical layer noise	Tolerant to syntax errors, ensuring semantic correctness	Error correction code → common sense reasoning
Interaction Mode	End-to-end transmission	Knowledge collaboration	Channel → sharing

Note: This comparison highlights primary design orientations. Advanced TC can also realize task-aware and cross-layer ORA, but lacks explicit semantic-content awareness and semantic-utility-oriented optimization objectives.

.

In TC, the primary focus lies in the symbolic transmission of information. The goal is to ensure that the bit streams received by the receiver are identical to those transmitted by the sender, irrespective of the actual meaning of the transmitted symbols. In contrast, SC, as an emerging paradigm in the communications domain, seeks to transcend the limitations of TC by prioritizing not only the accuracy of data transmission but also the comprehension and exchange of SI. SC introduces several distinguishing features:

• Semantic perception and understanding: Emphasizes the extraction and transmission of meaningful semantic content. Analysis of semantic attributes enables communicating parties to better grasp the intrinsic meaning of messages, thereby enhancing both communication efficiency and accuracy.
• Cross-modal fusion: SC systems are capable of integrating information across multiple modalities—such as text, speech, image, and video—into a unified semantic representation. This enables more comprehensive and coherent communication in complex environments.
• Efficient use of resources: By leveraging semantic awareness, SC enables intelligent RA. Important semantic content can be prioritized in terms of RA to ensure its accurate delivery, while less critical information can be transmitted with reduced resource consumption, thereby improving overall communication efficiency.
• Task-driven and intelligent interaction: SC systems employ AI technologies to extract and transmit task-relevant SI, facilitating dynamic understanding and personalized responses.

In summary, SC introduces a novel communication paradigm that emphasizes the extraction, transmission, and utilization of the semantic content of information. By shifting the focus from syntactic accuracy to semantic relevance, SC offers a promising pathway to overcome the inherent limitations of TC systems. This paradigm shift not only enhances communication efficiency but also broadens the scope of practical applications across diverse intelligent and interactive scenarios. ORA is key in SC, enabling accurate semantic delivery and optimal resource use, boosting system performance and adaptability.

2.3. Overview of Optimal Resource Allocation

ORA in SC refers to the strategic application of methods and techniques that allocate bandwidth, power, time, and computational resources in a scientifically optimized manner, based on the characteristics of SI, communication demands, and system constraints [30]. The objective is to ensure the efficient processing, transmission, and interpretation of semantic content, thereby maximizing overall system performance under resource-limited conditions. Unlike traditional communication systems that transmit all data bit by bit, SC systems prioritize the transmission of core semantic content through preliminary semantic analysis at the sender side. This significantly reduces bandwidth consumption and alleviates processing burdens on the receiver. Upon reception, the receiver reconstructs the complete information using its contextual knowledge base [30]. In environments with restricted bandwidth or degraded channel quality, SC maintains communication efficacy by focusing on transmitting semantically essential information.

Table 4. Comparison of ORA in TC and SC.

Comparison Dimensions	ORA in TC	ORA in SC	Fundamental Differences and Core Ideas
Theoretical Basis	Shannon’s information theory	SI theory and utility theory	Paradigm shift from “accurate symbol transmission” to “effective meaning delivery”
Optimization Objectives	Maximizing channel capacity, minimizing bit error rate	Maximizing semantic utility, minimizing semantic distortion, optimizing task success rate	Shift of objectives from communication link performance to task completion efficacy
Consideration of multi-dimensional resources	Mainly communication resources: bandwidth, power, time slots, etc.	Communication-computation-storage converged resources: bandwidth, power, computing power, model complexity, cache	Expansion of resource scope
Scheduling and allocation	Bits/Data packets: all bits are theoretically equally important	Semantic units/tasks: differentiated processing based on their semantic importance	Resource allocation is directly related to the value of information content
Consider the significance of transmitting information	Being unaware of the information content is a transmission at the level of “grammar”	Resource allocation is based on the understanding of information semantics and value judgment	Semantic perception is introduced
Handling of “errors”	Zero tolerance: error-free recovery of each bit through channel coding	Flexibility/tolerance: allow distortion or loss of non-critical semantic units, and concentrate resources to protect key semantics	From “absolute reliability” to “effective reliability”
Typical measurement index	Throughput, bit error rate, signal-to-noise ratio	Semantic fidelity, task success rate, semantic utility, etc.	Evolution of evaluation criteria
Resource allocation strategy	Fixed proportion and average distribution	Dynamic allocation driven by cross-domain resource collaboration, task awareness and semantic importance	The leap in strategy complexity

compares ORA in SC with that in traditional Shannon-based communications.

In order to ensure the full use of this idea, a dedicated ORA model is essential, providing a mathematical and algorithmic framework for systematic RA. As illustrated in Figure 5, a typical ORA model for SC comprises four key components: optimization objectives, optimization variables, system constraints, and optimization algorithms.

In the resource optimization model for SC, a global objective function $L\left(x\right)$ is defined to achieve system-wide optimality. This function integrates both the utility gained from effective SI transmission and the cost incurred from resource consumption. Specifically, $L\left(x\right)$ represents the difference between the total utility function and the aggregated resource cost function across all resource entities within the system. The objective is to maximize net system utility under given constraints.

2.3.1. Importance of Optimal Resource Allocation in Semantic Communications

In SC systems, ORA plays a crucial role. Unlike traditional communication paradigms, SC is designed to transmit the core semantic content of information rather than raw bit-streams. As a result, efficient and intelligent RA becomes especially critical to ensure both the performance and resource efficiency of SC systems. By tailoring resource distribution—such as bandwidth, power, computing capacity, and transmission time—to the semantic relevance of information, ORA enhances the overall effectiveness of semantic transmission. This not only improves communication reliability and speed under constrained conditions but also significantly reduces redundant processing and resource waste [31,32]:

• Semantic-driven intelligent resource adaptation: Unlike conventional bit-rate maximization, SC solves constrained semantic utility maximization under resource scarcity. Multi-modal data exhibits non-linear value heterogeneity—marginal utility vanishes for redundant background but remains high for critical features. In the absence of ORA, mission-critical information is truncated while bandwidth is squandered on task-irrelevant data. Consequently, ORA guarantees maximum task completion probability under severe resource constraints.
• Energy efficiency optimization through semantics-integrated RA: ORA in SC also plays a crucial role in reducing system energy consumption. With the exponential growth of data volume, energy efficiency has become a critical concern in communication systems. Moreover, energy consumption can be directly incorporated as an optimization objective within the ORA model, contributing to the overall energy efficiency and sustainability of the communication system.
• Semantic-adaptive elastic resource architecture: The significance of ORA in SC is further manifested in enabling system scalability and flexibility. As communication demands evolve and emerging technologies are integrated, the system must maintain a high degree of adaptability. Semantic-aware resource optimization facilitates dynamic allocation and reconfiguration of resources in response to varying application scenarios and load conditions. This elasticity allows the system to efficiently accommodate diverse and complex requirements, ensuring robust support for future-oriented and heterogeneous communication environments.

In summary, ORA is essential in SC. By implementing scientifically grounded and intelligent resource management strategies, ORA enhances both the efficiency and stability of SC systems while simultaneously minimizing energy consumption. This not only improves overall communication performance but also supports the long-term sustainability and scalability of next-generation communication technologies.

2.3.2. Challenges of Optimal Resource Allocation in Semantic Communications

In SC systems, although ORA offers many benefits to the system, it still encounters several critical challenges [33,34]:

• High information volume: Semantic compression vs. resource constraints

  –

  Complexity of joint optimization: The joint optimization of semantic and channel coding spans a high-dimensional parameter space, rendering traditional optimization methods ineffective. Furthermore, wireless communication environments are inherently dynamic—channel conditions fluctuate over time and space. Therefore, ORA algorithms must rapidly adapt to these changes to ensure the reliable transmission of SI.

  –

  Multi-objective trade-offs: ORA must often address multiple, potentially conflicting goals simultaneously—such as maximizing transmission rates, minimizing latency, and ensuring communication reliability. Balancing these objectives within a unified optimization framework presents a significant technical challenge.

• Information abundance: Semantic heterogeneity and generalized RA

  –

  Resource diversity and application needs: Communication system resources are inherently limited. At the same time, SC applications vary widely in their demands. Designing allocation strategies that can generalize across diverse applications while maintaining efficiency is a core difficulty.

  –

  Dynamic semantic prioritization: The importance of semantic content can vary in real time, requiring the system to dynamically reassess and re-ORA. Additionally, the complex, context-dependent nature of semantics makes it challenging to formalize semantic importance through a unified, tractable mathematical model.

• Diverse application scenarios: Contextual complexity and adaptability

  –

  Computational overhead of resource algorithms: Effective ORA strategies typically require sophisticated algorithms, which may impose high computational burdens.

  –

  Security and privacy considerations: Ensuring the accurate and secure transmission of SI poses additional challenges. Resource allocation strategies must incorporate mechanisms for protecting sensitive information and user privacy, which adds further complexity to system design and protocol implementation.

In summary, although ORA in SC systems encounters multiple technical and practical challenges, constructing a robust, adaptable ORA model remains key to overcoming these obstacles and realizing the full potential of SC.

2.4. Terminology Definitions

The following is a list of terminology definitions:

• ORA: A specific subset of resource allocation. It refers to resource allocation obtained by solving well-defined optimization problems under given constraints to maximize/minimize predefined performance metrics.
• RA: The general process of assigning available communication, computing, spectrum and energy resources to multiple users, tasks or links; it is a broad general concept, which does not imply optimization objectives.
• Resource optimization: An action category that includes building objective functions, setting constraints and solving problems; it refers to the process of improving resource-usage performance, which may or may not output an optimal solution.
• Dynamic resource adjustment: A runtime behavioral mechanism. It denotes real-time updating of resource assignments in response to time-varying channels, task states or user demands; dynamic adjustment can implement either heuristic schemes or ORA solutions.
• SI: The carrier-independent abstract information reflecting facts, events or states extracted from original raw data, which is the object to be transmitted in SC.
• Semantic content: Concrete instantiated representation of semantic information (text tokens, feature vectors, image semantic features), which is the actual data payload transmitted over physical channels.
• Semantic meaning: Task-oriented interpretive connotation behind semantic information. It is the high-level understanding target for upper-layer tasks.
• Semantic utility: Quantitative metric representing how much benefit a piece of semantic information brings to the target task. It is the core optimization objective for ORA in SC.
• Semantic fidelity: Performance metric measuring the degree of matching between recovered semantic content at the receiver side and original SI at the transmitter side.

3. Key Technologies for Semantic Communications—Oriented Resource Optimization

3.1. Semantic-Aware Resource Demand Prediction Technology

In SC systems, accurate resource demand prediction is pivotal for optimizing system performance. This prediction is heavily dependent on effective semantic feature extraction and analysis. By accurately understanding the semantic content of the data being communicated, SC systems can dynamically allocate resources such as bandwidth, storage, and computing power. Resource demand prediction technology based on semantic understanding has broken through the limitations of traditional temporal and statistical methods through intent-aware and knowledge-driven semantic extraction. Its core innovation lies in dynamically correlating “demand semantics” with “system resource states”, thereby achieving a paradigm shift from “passive response” to “active anticipation”. In the future, with the advancement of multi-modal large models, semantic feature extraction will further evolve toward end-to-end adaptation and cross-domain generalization.

Semantic understanding refers to parsing the deep-level meaning in textual data through technologies such as natural language processing (NLP) and knowledge graphs, while resource demand prediction aims to forecast future resource usage or shortages based on historical data. The integrated technology, by extracting SI from demand-related texts and scenarios, enables more accurate dynamic resource planning.

Role of Resource Demand Prediction in the ORA Model: Resource demand prediction based on semantic understanding is integral to the overall operation of the ORA model in SC systems. It influences all four modules of the ORA model:

• Optimization variables: Accurate predictions allow for the precise initialization and dynamic adjustment of optimization variables.
• Optimization target weights: Prediction results inform the adaptive allocation of optimization target weights, ensuring that key resources are allocated efficiently.
• Constraint conditions: By predicting resource demand, the system can manage constraints flexibly and adaptively to meet real-time needs.
• Optimization algorithms: The predictions act as input features for resource optimization algorithms, enabling efficient resource utilization.

Resource demand prediction is a data-driven quantitative process that prioritizes numerical accuracy. Primarily aimed at system efficiency optimization, its core objective is to forecast future resource consumption.

Table 5. Research on resource demand prediction based on semantic understanding (Part 1).

Prediction Technology Type	Literature	Optimization Variables	Objective Function	Constraints	Resource Types	Optimization Method	Considered Scenario	Performance Metrics	Main Limitations
Based on semantic importance	[9]	Semantic unit transmission priority	Maximize semantic transmission utility	Bandwidth budget	Bandwidth, computing	LLM-driven semantic importance evaluation	Text and image SC	Semantic similarity, task success rate	Heavily depends on LLM reasoning accuracy
	[35]	ROI/RONI transmission bit-rate	Maximize effective semantic reconstruction quality	Channel capacity limit	Bandwidth	Semantic segmentation based ROI division	Image semantic transmission	mIoU, reconstruction quality	Performance relies on segmentation accuracy
Channel-adaptive joint source channel	[36]	Source channel coding parameters	Maximize semantic reception quality	Channel state, transmission power	Bandwidth, power	Channel- adaptive JSCC	Digital SC	Semantic fidelity	High implementation complexity
Based on key semantic extraction and enhancement	[37]	Implicit feature weight	Maximize semantic feature richness	Computing latency	Computing	RoBERTa + Bi-GRU attention pooling	Text semantic analysis	Semantic similarity	High computational overhead from pre-trained model
	[38]	Feature enhancement weight	Maximize segmentation precision	Edge node computation limit	Computing, storage	Lightweight real-time semantic segmentation	Edge SC	Segmentation accuracy	Poor adaptability for dynamic networks

and

Table 6. Research on resource demand prediction based on semantic understanding (Part 2).

Prediction Technology Type	Literature	Optimization Variables	Objective Function	Constraints	Resource Types	Optimization Method	Considered Scenario	Performance Metrics	Main Limitations
Based on key semantic extraction and enhancement	[39]	Multi-modal feature weight	Maximize feature discriminability	Computing resource budget	Computing	Multi-modal contrastive learning	Multi-modal SC	Feature representation quality	Complex model, high computation overhead
	[40]	Key semantic selection probability	Maximize key semantic retention rate	Latency constraint	Computing	Probabilistic graph + key semantic ranking	Text SC	Semantic extraction accuracy	Depends heavily on external knowledge base
Resource demand prediction based on task utility	[41]	Transmitted deep feature dimension	Minimize resource consumption under task completion	Task discrimination threshold	Bandwidth, computing	Task-oriented deep feature extraction	Image classification SC	Classification accuracy	Task-specific, poor generalization
	[42]	Image semantic compression ratio	Balance transmission overhead and reconstruction quality	Edge node resource limit	Bandwidth, storage	Blockchain knowledge- aided SC	Remote driving image transmission	Reconstruction quality	Depends on edge node distribution
	[43]	Compact semantic representation dimension	Minimize agent communication overhead	Inference latency	Computing, bandwidth	GPT-based semantic extraction	Multi-agent SC	Task decision accuracy	Semantic alignment ambiguity, high inference cost

summarize relevant technical research on resource demand prediction based on semantic understanding in SC in recent years. As mentioned above, studies propose innovative approaches to integrating semantic understanding with resource prediction, enhancing accuracy and performance through DL, joint coding, and pre-trained models. Innovatively using semantic understanding for resource prediction. It lays the foundation for resource prediction through effective SI extraction methods and shows application potential in multiple fields. However, its limitations are also obvious, with limitations on scenarios and data types, insufficient research on resource prediction technology, lack of unified performance evaluation indicators, and insufficient consideration of resource constraints in practical applications. These have limited the widespread application and in-depth development of related technologies, affecting resource prediction efficiency, accuracy, and practical application.

3.2. Optimal Resource Allocation Technology Based on Semantic Information Processing and Transmission

3.2.1. Semantic Information Processing Technology

SI processing is a knowledge-driven comprehension process that emphasizes logical consistency. The core objective of SI processing is to extract high-level SI (such as intentions, entity relationships, and knowledge concepts) from raw data, thereby enabling intelligent information understanding and maximizing semantic fidelity. The methodology primarily comprises three categories: one is structured extraction based on a knowledge graph. Mapping data to semantic space through ontology reasoning and relational embedding. The second combines semantic networks and end-to-end DL to generate semantic vectors by making full use of the context-coding ability of large language models, and efficient cooperative allocation of resources is realized in multi-link systems. The third is task-driven dynamic extraction, which adaptively focuses on key semantics according to specific scenario requirements to support flexible resource optimization in complex link environments.

JSCC offers unique advantages in efficiently encoding, transmitting, and decoding extracted SI [44]. Unlike the traditional approach that separates source and channel coding, JSCC enables better resource utilization and improved communication performance, especially in dynamic and complex environments. In SC, JSCC breaks the conventional separation by jointly optimizing both source and channel coding. It allows for unified RA based on semantic importance and real-time channel conditions, thereby enhancing resource efficiency and reducing redundant data transmission.

The implementation of dynamic JSCC requires real-time perception and analysis of both the communication environment and semantic content. In varying scenarios—such as day versus night or densely versus sparsely populated areas—the relative importance of SI and the state of the communication channel can differ significantly [45]. The system continuously monitors channel status and adapts accordingly. These adaptive strategies in SI processing help reduce the volume of transmitted data while maintaining communication fidelity.

Table 7. Research on SI processing technology based on ORA.

SI Processing Type	Literature	Optimization Variables	Objective Function	Constraints	Resource Types	Optimization Method	Considered Scenario	Performance Metrics	Main Limitations
General semantic extraction based on DL	[46]	Multi-head attention weight	Maximize semantic similarity	Transmission re-transmission limit	Computing	Extended context modeling	Text SC	BLEU, semantic similarity	Single processing object, poor real-time performance
	[47]	Multi-scale feature weights	Maximize task accuracy with low power	Power budget	Computing, power	Dual-branch multi-scale extractor	Image semantic transmission	Task accuracy, robustness	Large model memory overhead
	[48]	Vector quantization codebook size	Minimize transmitted data volume	Reconstruction distortion limit	Bandwidth, computing	Vector quantization generative model	Image SC	SSIM, LPIPS	Quantization accuracy loss, unstable training
	[49]	Semantic reconstruction parameters	Maximize target reconstruction quality	Receiver computation limit	Computing	Target-oriented semantic reconstruction	Vehicular network communication	Not reported	Heavy burden on receiving-side devices
Semantic compression based on efficient encoding	[50]	PCNN internal pulse parameters	Minimize computation and time consumption	Segmentation accuracy threshold	Computing, time	PA-PCNN semantic segmentation	IoT SC	Processing speed, segmentation accuracy	Low semantic segmentation accuracy
	[51]	Residual coding compression ratio	Maximize compression ratio	Semantic distortion threshold	Bandwidth, storage	Semantic- residual joint encoding	General semantic transmission	Compression ratio, runtime	Limited universality for semantic extraction
Content-aware and task-oriented semantic filtering	[52]	ROI compression quantization factor	Reduce transmitted data volume	Reconstruction quality lower bound	Bandwidth	ROI-RONI differential quantization	Image wireless semantic transmission	Compression ratio, mIoU	Strong dependency on ROI segmentation precision
	[53]	Mode switching threshold	Balance compression efficiency and reliability	Not reported	Bandwidth	Mask-based dual-mode adaptive transmission	Segmentation task image transmission	Compression rate, reconstruction quality	Sensitive to mode-switching threshold
	[54]	Codebook size	Minimize bandwidth occupation	Classification accuracy constraint	Bandwidth	Cognition- inspired semantic inference	Image classification	Classification accuracy, bandwidth saving ratio	Strong task-specific, weak generalization
	[55]	Adaptive semantic compression factor	Maximize task success probability	Not reported	Bandwidth, computing	Task-driven adaptive semantic compression	Task-oriented SC	Task success probability	Gap between model and real-world complexity
	[56]	Key-feature selection threshold	Minimize communication overhead	Semantic fidelity constraint	Bandwidth	Key feature screening	General semantic transmission	Semantic spectral efficiency	Difficulty in key feature definition and evaluation

summarizes the latest research on SI processing technologies based on ORA, which can be roughly classified into the following three categories: general semantic extraction based on DL, content-aware and task-oriented semantic filtering, and semantic compression based on efficient encoding.

The reviewed studies offer innovative approaches and effective technical solutions for SI processing, significantly advancing the development of SC. These works highlight the multi-dimensional advantages of SI processing, such as leveraging DL to extract semantic features, thereby improving processing accuracy and efficiency. They also introduce novel methods for semantic compression through semantic extraction and residual coding, which help preserve semantic integrity while reducing data volume. Furthermore, task-oriented strategies enable intelligent optimization of semantic compression, RA, and transmission processes. Techniques such as RL for semantic coding and joint processing frameworks further enhance communication efficiency.

These advances have facilitated the progress of SI processing across various scenarios and application domains. However, several challenges remain. DL models face high data costs and limited interpretability. Semantic extraction and residual coding lack universality and are data-sensitive, with accuracy still constrained. Task-oriented methods suffer from subjective definitions and poor adaptability. RL-based techniques converge slowly and face instability in dynamic environments. Joint processing is complex, requiring optimization due to multiple interacting factors.

3.2.2. Multi-Dimensional Joint Optimization Technology

In SC, effective resource co-allocation strategies must account for dynamic factors such as channel conditions and semantic significance. Under favorable channel conditions with minimal interference, more resources can be allocated to data streams carrying critical SI [57]. Given the reliable channel, power can be reduced without compromising signal quality, leading to energy savings. Conversely, in adverse or interference-prone multi-link environments, ensuring reliable transmission of vital semantic content may require dynamically increasing transmission power on vulnerable links or implementing semantic-aware link switching to enhance overall system robustness [58].

Coordinating energy and time resources across multiple semantic links is crucial for maintaining end-to-end system stability and optimizing performance. The proper allocation of energy directly impacts the operational status of distributed semantic devices and the quality of multi-hop signal transmission. Time resource management becomes particularly complex in multi-tasking SC environments with parallel links. Transmission schedules must be efficiently organized to prevent conflicts and resource contention, thereby enhancing overall network resource utilization.

Energy and time optimization also extends to energy-efficient operation [59]. During idle periods, SC devices can switch to low-power modes, conserving energy while maintaining readiness to quickly resume normal operation when communication is required. This ensures timely SI transmission.

Then, the coordination of storage and other computational resources is key to improving SC performance. Strategic use of local storage can reduce redundant transmissions by caching frequently used semantic data or intermediate processing results. This allows subsequent transmissions of related information to retrieve data locally, minimizing communication overhead. Additionally, storage can serve as a buffer for important semantic data, enabling rapid response to sudden communication demands.

Multi-modal data fusion technology plays a crucial role in SC. By integrating data from diverse sources and modalities, it enables the transmission and processing of richer and more comprehensive SI. This technology is a synergy-driven integration process that emphasizes informational complementarity. Its core objective is to achieve efficient and precise SI transmission and reconstruction through cross-modal semantic alignment and collaborative processing. Multi-modal data fusion in SC involves alignment, fusion, and semantic reasoning [60]. In recent years, deep learning has been widely used in multi-modal data fusion, significantly advancing the capabilities of SC systems. CNNs have demonstrated exceptional performance in extracting visual features from image data. Recurrent neural networks (RNNs) and LSTM networks are particularly effective in processing sequential data, including text and audio. Cross-modal transformation models also excel in aligning diverse data types, boosting SC’s applicability across domains.

By strategically coordinating the use of multi-dimensional resources alongside multi-modal data fusion, SC systems can ensure the accurate and efficient transmission of semantically processed information. This not only optimizes the overall performance of SC systems but also significantly improves resource utilization and operational efficiency.

Table 8. Multi-dimensional joint optimization technology.

Joint Optimization Type	Literature	Optimization Variables	Objective Function	Constraints	Resource Types	Optimization Method	Considered Scenario	Performance Metrics	Main Limitations
Combined optimization of communication and computing resources	[61]	Beamforming, sub-channel allocation, semantic bit number	Maximize SC QoS	Power, bandwidth limit	Bandwidth, power, computing	Adaptive ORA model	Semantic communication network	QoS	High computational complexity, lacks semantic measurement standard
	[62]	Resource unit allocation, semantic transmission mode	Adaptive resource mode matching	Latency constraint	Bandwidth, computing	Deep reinforcement learning	End-to-end SC	Transmission throughput	Unstable training, weak generalization ability
	[45]	Semantic compression, channel allocation, transmit power	Maximize multi-task QoE	Power budget	Bandwidth, power, computing	QoE-driven joint optimization	Multi-task semantic network	QoE	Subjective inaccuracy of QoE modeling
Combined optimization of communication and perception or control	[63]	Sub-channel, modulation, power, UAV position	Maximize overall network performance	UAV movement boundary, power limit	Bandwidth, power, spatial deployment resource	Mixed-integer nonlinear programming	UAV-assisted SC	Network throughput	Poor expansibility, over-idealized assumptions
	[64]	IRS phase coefficient, sub-channel allocation	Maximize secure semantic spectral efficiency	Interference threshold, power limit	Bandwidth, power, IRS hardware resource	Noise- enhanced hybrid DRL	IRS-aided secure SC	Secure spectral efficiency	Sensitive to channel estimation error, high hardware cost
Multi-modal cooperation and compression	[65]	Multi-modal collaborative feature weight	Maximize compression ratio under segmentation accuracy guarantee	Feature distortion constraint	Computing, bandwidth	Cooperative multi-modal SC	Autonomous driving perception semantic transmission	Compression ratio, segmentation accuracy	High computation cost, complex multi-modal synchronization
	[66]	Multi-level similarity threshold	Improve compression efficiency and task accuracy	Not reported	Computing, bandwidth	Graph-based multi-level similarity analysis	Multi-modal SC	Compression efficiency, task accuracy	Strongly relies on semantic map quality
	[67]	Cross-modal feature alignment weight	Improve multi-modal semantic understanding performance	Feature loss constraint	Computing	Contrastive learning cross-modal alignment	Multi-modal semantic understanding	Semantic understanding accuracy	Highly dependent on dataset quality

presents a summary of recent research on multi-dimensional joint optimization technologies in SC.

In summary, these studies explore multi-dimensional joint optimization strategies that employ intelligent algorithms to enable dynamic resource allocation (DRA) and mode selection in SC. These approaches aim to enhance resource utilization efficiency and overall system performance by incorporating semantic awareness and comprehension. The research addresses diverse application scenarios, including edge computing environments, SC networks, and multi-modal communication among unmanned systems. ORA is conducted based on both system state and semantic requirements, while also considering real-time responsiveness and security constraints.

Despite notable advancements, several limitations remain. Most existing studies rely heavily on intelligent algorithm-based model solving, which often incurs high training costs, slow convergence, and sensitivity to environmental variability. Furthermore, the generalizability of these solutions is limited, reducing their effectiveness in complex and dynamic real-world scenarios.

3.3. Dynamic Self-Adaptive Resource Adjustment Technology

The necessity of dynamic resource adjustment in SC lies in its ability to respond adaptively to complex, time-varying communication environments, thereby ensuring efficient and reliable transmission of SI. TC systems often rely on static RA strategies, which are insufficient to address the dynamic demands inherent in SC. In contrast, dynamic resource adjustment enables flexible optimization of bandwidth, power, and computational resources in real time, based on variations in channel conditions, the importance of semantic content, system workload, and multi-user competition.

Furthermore, SI often has hierarchical characteristics. Dynamic resource adjustment can exploit this by assigning higher priority to semantically important units and compressing less critical content, thus improving utilization efficiency without compromising essential meaning. Therefore, DRA is a fundamental enabler for achieving high energy efficiency, low latency, and reliable SC.

The core paradigm of SC shifts from ensuring reliable bit transmission to guaranteeing accurate semantic delivery and efficient task execution. This transformation has driven resource management to evolve from a static, passive allocation model to dynamic adaptive resource adjustment. By perceiving real-time dynamic changes in channel states, semantic content, and tasks, this technology aims to collaboratively optimize multi-dimensional resources ranging from the physical layer to the semantic layer, thereby achieving the global optimization of communication efficiency and task utility. Currently, research in this technology can be categorized into the following three directions based on its adaptive driving mechanisms and adjustment targets.

• Channel state-driven adaptive transmission technology
  Studies in this category extend the foundation of adaptive technology from traditional communication to SC. It primarily adjusts system parameters dynamically based on real-time channel state information (CSI) to ensure the robustness and transmission efficiency of SI under different channel conditions. Among these studies, a wireless image SC system based on CNNs is proposed in [68]. The core innovation of this system lies in the design of a semantic adaptive module, which can dynamically adjust the weight allocation of different semantic features according to CSI, which significantly improves the performance of the system in different channels. In [69], the paper proposes a task-oriented adaptive semantic reconstruction communication scheme. Efficient semantic compression is achieved through a compression mechanism based on semantic importance, and an adaptive semantic reconstruction network is designed to predict and repair lost SI. The main disadvantages are that the multi-layer adaptive mechanism may increase system complexity and computational overhead, and its performance improvement depends to a certain extent on the accurate estimation of channel states and semantic loss patterns. Furthermore, in [70], semantic importance awareness is integrated into the classical orthogonal frequency division multiplexing (OFDM) framework. By dynamically allocating subcarriers and power, the transmission of semantically high-importance information is prioritized, which embodies an adaptive strategy for safeguarding semantic quality under harsh channel conditions.
• Task and semantic-driven intelligent RA technology
  Studies in this category mark the deepening of adaptive technologies: their decision-making basis has shifted from physical-layer channel states to the semantic and task levels, enabling a transition from “ensuring connectivity” to “ensuring effectiveness.” A scalable coding framework based on semantic decomposition is proposed in [71]. By decomposing unstructured images into hierarchical semantic features to enhance the accuracy of semantic representation, and adopting a DRA strategy based on semantic knowledge bases and user intentions, this framework achieves differentiated coding for different semantic targets. The main disadvantages are that the system architecture is relatively complex, and its performance highly depends on the completeness of the semantic knowledge base and the accuracy of user intention recognition. In [72], a spatiotemporal importance-aware framework is proposed. This framework leverages deep learning models to extract the importance features of data in the spatial and temporal dimensions, and dynamically allocates resources according to the priority of these features to ensure the reliable transmission of high-importance features.
• Artificial intelligence-based cross-layer joint dynamic optimization technology
  Faced with complex optimization problems characterized by high dimensionality and non-convexity, DL and DRL have become key enabling technologies for realizing cross-layer, multi-dimensional joint dynamic adaptive decision-making. RL offers an adaptive solution for RA in SC systems by enabling an intelligent agent to learn optimal strategies through iterative interactions with a dynamic environment. The agent, typically a resource management controller, adjusts parameters like bandwidth or transmission power based on real-time network conditions and user demands [33]. Each action receives a reward reflecting its impact on performance metrics such as semantic accuracy, latency, or energy efficiency. Positive rewards reinforce actions that improve outcomes (e.g., higher semantic fidelity), while negative rewards penalize suboptimal decisions. Over time, the agent refines its policy to maximize cumulative rewards [33].

These studies highlight substantial advancements in the field of SC. From a technological innovation perspective, they introduce state-of-the-art methodologies that address the complex challenge of RA in SC environments, significantly enhancing both resource utilization efficiency and overall system performance. In terms of application scope, these studies demonstrate broad relevance across a variety of domains, including mobile edge computing and emerging 6G network infrastructures, thereby expanding the practical applicability and influence of SC technologies. A notable emphasis is placed on multi-factor joint optimization strategies, which incorporate factors such as personalized saliency and semantic importance. This enables precise and context-aware RA, facilitating more efficient and intelligent distribution of system resources across diverse scenarios and data modalities. However, these studies are not without limitations. Many methods lack generalizability, being tied to specific architectures or requiring extensive tuning. Advanced techniques like DRL face challenges such as high computational costs, slow convergence, and local optima, limiting their scalability in dynamic real-world environments.

While DRL-based ORA demonstrates strong adaptability within trained environments, its deployment in real-world SC networks faces critical generalization hurdles. Learned policies are often tightly coupled to specific channel distributions, traffic patterns, and topological configurations encountered during training. When confronted with out-of-distribution states–such as sudden node mobility, unanticipated interference profiles, or heterogeneous semantic tasks not represented in the training data–DRL agents may suffer significant performance degradation due to overfitting to narrow state spaces. Recent advances in generalized wireless protocol learning address this issue by abstracting the observation space into transferable semantic representations. In [73], an abstraction-based framework is proposed that enables learned MAC protocols to generalize across diverse deployment scenarios by operating on higher-level state features rather than raw channel measurements. Paper [74] introduces a feasible multi-agent reinforcement learning framework that explicitly constrains the action space to physically realizable protocols, thereby improving transferability to unseen network conditions while maintaining training stability. These works highlight that future ORA strategies in SC must move beyond environment-specific policy learning toward generalizable, abstraction-aware DRL architectures that can robustly handle the non-stationarity inherent in SC environments.

3.4. Conceptual Framework for Optimal Resource Allocation

ORA in SC needs to break away from the traditional modeling framework based on Shannon capacity to fully tap its performance potential and achieve the optimal overall network efficiency. This requires fundamentally establishing new RA principles based on semantic characteristics. Based on this, the objectives of ORA in SC are typically defined from two dimensions, as shown in Figure 6:

• The utilization efficiency of the resources themselves, such as semantic spectral efficiency and semantic energy efficiency.
• The semantic performance at the system level, which is commonly measured by semantic-based channel capacity, semantic similarity, and indicators oriented to specific service quality (e.g., recognition accuracy).

Current research focuses on diverse task scenarios, leading to differentiated optimization priorities and methods, as shown in

Table 9. Existing ORA schemes for SC.

References	Optimization Objective	Constraint Conditions	Scenario/Task	Method
[75]	Semantic spectral efficiency	Semantic similarity, semantic compression ratio	Text transmission	Exhaustive search method, Hungarian algorithm
[76,77]	Energy consumption	Computation, latency, bandwidth	Text transmission, machine translation	Iterative algorithms, RL
[78,79,80]	Semantic similarity	Bandwidth, latency, energy consumption, object detection accuracy	Text transmission, IoV, autonomous driving	RL, DL
[81]	Semantic transmission efficiency	Bandwidth, semantic compression ratio, transmission power	Image classification	DRL
[82]	Semantic message throughput	Bandwidth, background knowledge base matching	Heterogeneous networks	DNN
[83,84]	Semantic transmission rate	Semantic similarity, power, inter-user interference	Dense cellular scenarios	DNN
[85,86]	Semantic QoE	Semantic compression ratio, channel allocation, power, time slot allocation	Image transmission	Matching theory, DL, RL

Summary and analysis. (1) Optimization objectives: The lack of a unified cross-modal semantic utility function. (2) Constraints: Distributed semantic resource coordination mechanisms in multi-agent competitive scenarios have not been fully explored. (3) Solution methods: The absence of a transferable and general semantic RA model. (4) Application scenarios: Insufficient online adaptive capability in dynamic and open environments.

.

This study conceptualizes an ORA workflow for task-oriented SC systems operating under resource-constrained conditions, focusing on the fundamental contradiction between massive data volume, diverse content types, sophisticated application scenarios, and limited communication resources. Taking a task-oriented SC system as our research object, this study aims to enhance information transmission capability under resource-constrained conditions. Deconstructed from the “sender–receiver” model, and based on the communication channel under the condition of a poor network, the situation awareness mode of “demand feedforward SI feedback” is designed, and the ORA model is applied in the transmission process, as shown in Figure 7.

Based on task criticality and QoS requirements, this paper establishes an ORA model that incorporates available resources (e.g., bandwidth, computational resources) to maximize overall system utility, with application to UAV video or image collection and transmission. Specifically, the front-end unmanned equipment performs adaptive multi-scale compression on SI units (including task-relevant information, task-irrelevant information, and background information) based on the target requirements of the receiving end and the currently available resources. This model is task-demand-oriented, conducting in-depth analysis of the semantic features and priorities of different data acquisition tasks. By integrating existing resources, it establishes an optimization objective function and incorporates compression algorithms (dimensional compression, JPEG compression) and optimization algorithms (reinforcement learning, game theory, and particle swarm optimization). This enables the model to consistently provide ORA solutions in complex and dynamic environments while ensuring the timely and accurate transmission of critical information. Additionally, it rationally allocates bandwidth and other available communication resources to meet the demands of diverse tasks and optimize transmission performance. The objective of the proposed ORA model is to maximize the aggregated utility across all tasks under resource constraints. The optimization decision-making process is shown in

Table 10. Optimization decision-making process.

Step	Name	Core Content	Key Formulas/Technologies
Step 1	Multidimensional Resource Modeling	Construction of Unified Resource State Space	Resource vector $r_i\left(t\right)={\left[B_i,P_i,E_i,C_i,T_i\right]}^T$
Step 2	Multidimensional Constraint Analysis	Modeling of Resource Coupling and Competition Relationships	Communication constraints, energy constraints, latency constraints and coupling relationships
Step 3	Semantic Value Driven Objective Construction	Maximization of Semantic Utility	$max{\sum }_{t=0}^T{\sum }_{i=1}^{N}U(V_i,d_i,T_i)+\delta J\left(r_i(t+1)\right)$

.

Building on the definition given in Table 10, we detail the three-phase solving procedure for our conceptual ORA framework as follows:

Step 1: Construct the problem space for resource optimization, including communication bandwidth (Hz), transmit power (W), battery energy (J), and time resource (time slot).

Step 2: Establish constraint conditions for the aforementioned resources.

Step 3: Construct the optimization objective function to maximize the semantic utility.

$$max{\sum }_{i=1}^N[V_i\left(t\right)·\left(1-e^{-{\lambda }_d{d}_i\left(t\right)}\right)·e^{-{\lambda }_T{T}_i\left(t\right)}]+\delta {\sum }_{i=1}^{N}J\left(r_i(t+1)\right)$$

where $d_i\left(t\right)$ denotes the transmission data volume, $T_i\left(t\right)$ represents the total latency, ${\lambda }_d$ and ${\lambda }_T$ are attenuation coefficients, and the resource transition function is given by $r_i(t+1)=\mathrm{\Phi }\left(r_i\left(t\right),a_i\left(t\right),e\left(t\right)\right)$.

The following terms are introduced:

• Semantic value weighting $V_i\left(t\right)$. This term represents the intrinsic importance or task relevance of the SI associated with task i at time t. It acts as a priority weight, ensuring that resources are preferentially allocated to high-value tasks. $V_i\left(t\right)$ is typically determined by semantic importance evaluation models.
• Diminishing marginal return of data volume ($(1-e^{-{\lambda }_d{d}_i\left(t\right)}$). The term models the saturation effect of transmitted data: as the data volume $d_i\left(t\right)$ increases, the marginal semantic utility gained from additional bits diminishes.
• Exponential delay penalty ($e^{-{\lambda }_T{T}_i\left(t\right)}$). The total latency $T_i\left(t\right)$ incurred by task i is penalized exponentially, reflecting the time-sensitive nature of semantic tasks.
• Future resource state value $\delta J\left(r_i(t+1)\right)$. This term transforms the static optimization into a sequential decision-making problem, enabling the model to account for the long-term impact of current allocation decisions.

However, in practical complex network environments (such as UAV swarm collaborative reconnaissance, IoT sensor networks, and IoV), SC systems inevitably operate in a multi-agent context. Multiple agents simultaneously act as senders or processing nodes of SI, forming a dynamic semantic perception and transmission network.

Deploying SC systems in multi-agent network environments poses a series of core challenges. First, the competition among multiple agents for shared communication and computing resources requires ORA models to shift from single-user optimization to multi-agent collaborative game theory that balances overall system efficiency, fairness, and group task completion. Second, the redundancy of perceived information among agents necessitates the development of cross-agent semantic perception and fusion technologies to achieve redundancy reduction and complementary enhancement, thereby alleviating network load. Meanwhile, group collaborative tasks require the exchange of high-level semantic intentions and states, which raises the requirements for low latency and high reliability in the transmission of critical commands. Additionally, the dynamic changes in network topology demand that ORA strategies possess adaptive and distributed decision-making capabilities. Finally, to avoid the signaling overhead and single-point failure issues of centralized optimization, it is essential to research scalable distributed learning and lightweight algorithms.

To address these challenges, systematic expansion is required at the critical technology level. This includes developing a distributed ORA model that leverages game theory or multi-agent RL, enabling each agent to jointly optimize semantic compression, channel selection, and power control in dynamic environments. By constructing a group knowledge graph or shared semantic foundation model, collaborative semantic perception and compression can be achieved—allowing each agent to transmit only differentiated information, which is then fused and reconstructed by the receiver. Additionally, it is necessary to design a hierarchical communication architecture of “raw data–semantic features–task intentions” and provide differentiated QoS guarantees for the information flow at each layer, thereby systematically improving the overall performance and robustness of multi-agent SC networks.

4. Development Trends in Optimal Resource Allocation for Semantic Communications

The development trend of ORA in SC is primarily driven by advancements in semantic understanding, intelligent resource scheduling, and signal processing.

4.1. Deep Integration with Artificial Intelligence Technologies

The rapid advancement of artificial intelligence (AI), particularly in DL, is significantly enhancing the semantic extraction and representation capabilities of SC systems. A prime example is NLP, which enables intelligent semantic perception. Modern NLP models go beyond lexical-level analysis, capturing contextual nuances and latent logical relationships within textual data [87]. These capabilities allow systems to accurately classify user queries and determine the necessary resources for each type. This semantic-aware RA ensures that resources are matched to actual task requirements, reducing waste and enhancing system responsiveness and QoE.

Artificial Intelligence of Things (AIoT) devices play a central role in SC by both generating and processing SI [88]. In smart homes, intelligent speakers convert voice commands into semantic representations using speech recognition technologies, enabling them to interpret user intents. This SI can be exchanged locally or transmitted across cloud platforms and other devices. In distributed AIoT environments, ORA models must exhibit high adaptability. These settings involve vast numbers of widely distributed devices with heterogeneous resource requirements and application scenarios. Traditional static ORA strategies often fall short. To meet these challenges, next-generation ORA models must dynamically account for multiple factors, including device location, information type, and real-time responsiveness [89]. By continuously adjusting ORA in response to device demands and network dynamics, the system can ensure optimal resource utilization, enabling efficient operation across the entire AIoT ecosystem and supporting scalable SC deployment.

Edge computing further enhances SC, especially in localized semantic processing. In intelligent transportation systems (ITS), roadside sensors and cameras continuously gather massive volumes of data—including vehicle speed, position, and traffic flow. Transmitting all raw data to the cloud would result in excessive latency and network load. Edge computing mitigates this by processing data at or near the source. Equipped with computer vision and analytics capabilities, edge nodes can perform real-time tasks such as vehicle type recognition, license plate detection, traffic signal assessment, and congestion analysis [90]. Only high-level SI—e.g., accident alerts, congestion hotspots, or traffic patterns—is sent to cloud platforms or control centers. This approach not only reduces bandwidth usage and processing delays but also enhances real-time responsiveness, providing timely and actionable insights for traffic management and decision-making.

4.2. Adaptation to Future Networks

To meet the ultra-high-speed and low-latency demands of 6G networks, ORA models must demonstrate exceptional efficiency and real-time performance. One such scenario is the satellite–terrestrial integrated network, which exemplifies significant resource heterogeneity across network layers [91]. Satellite networks provide broad coverage, offering seamless global connectivity. However, they face challenges such as large transmission delays, limited bandwidth, and power resources. In contrast, terrestrial networks offer low latency and high bandwidth but have restricted coverage. Effective multi-level resource collaboration is essential, leveraging the complementary strengths of both network types.

In emergency communication scenarios, such as in remote areas where terrestrial networks are unavailable, satellite networks become the primary communication medium. However, due to the limited satellite resources, careful planning of bandwidth and power is essential. Resource reservation can be utilized to allocate bandwidth and power for key tasks, ensuring communication quality during emergency situations [91]. Efficient coordination between satellite and terrestrial networks requires the establishment of a unified resource management platform to optimize network performance and reliability.

Vehicular networks serve as another quintessential SC scenario, characterized by inherent complexity and dynamism. These networks integrate various communication modalities, such as vehicle-to-vehicle (V2V), vehicle-to-infrastructure (V2I), and vehicle-to-pedestrian (V2P) interactions, forming a sophisticated communication ecosystem. Vehicles need to interact not only with surrounding vehicles to exchange real-time information such as speed and location for safe driving and traffic flow management, but also with roadside infrastructure like traffic lights and cameras to access current traffic rules and road conditions. The dynamic nature of vehicle networks primarily stems from frequent changes in network topology, as vehicles constantly alter their positions, impacting communication links and transmission quality.

Dynamic ORA is critical for ensuring efficient energy usage and maintaining communication quality. This involves three core operational dimensions: ORA, computational task processing, and real-time environmental adaptation. The first step is to precisely define resource requirements, including bandwidth, power, and cache space, each of which contributes differently to service quality and performance. Based on this, optimization objectives are formulated, such as minimizing power consumption, maximizing throughput, and reducing latency. The selection of an appropriate allocation algorithm is crucial. It must not only adapt effectively to the current environment but also respond quickly to potential conflicts and competition for resources. If an algorithm is deemed suitable, RA proceeds, and resource usage is continuously monitored to ensure QoS [92]. A simple flowchart of dynamic ORA is shown in Figure 8.

4.3. Signal Processing Technologies

Under the developmental trajectory of ORA in SC, signal processing technologies have experienced substantial advancements, playing a crucial role in enhancing system performance and efficiency. Semantic encoding and decoding technologies, powered by deep learning, enable the automatic extraction, encoding, and joint source-channel decoding of SI, improving data handling and transmission accuracy [93]. Multi-modal semantic signal processing leverages feature fusion and cross-modal conversion to integrate and utilize information from different modalities, enhancing the system’s ability to process diverse types of data. Additionally, intelligent channel estimation and equalization techniques, driven by machine learning, allow for precise modeling of channel conditions, leading to more accurate predictions and improved communication quality [94]. These advanced signal processing technologies enable dynamic power adjustment and the use of low-complexity algorithms based on semantic importance and channel conditions, optimizing resource utilization and improving energy efficiency. This progression significantly supports the comprehensive development of SC systems, ensuring better performance, adaptability, and scalability.

Intelligent channel estimation and equalization technology supports semantic signal transmission in complex channel environments. Channel estimation technology is based on machine learning, such as linear regression and auto-encoder algorithms in deep learning. With the help of received signals and known training data, channel state information can be estimated more accurately, channel changes can be tracked in real time, and an accurate basis can be provided for ORA. Adaptive equalization technology automatically adjusts the amplitude and phase of the signal based on channel estimation results, effectively reducing the impact of inter-symbol interference and signal fading on semantic signal transmission. It ensures the accuracy and integrity of the signal during transmission and guarantees SI transmission reliability. Interference management and elimination technology is an important part of ensuring SC quality. Semantic-aware interference detection technology analyzes the received signal at the semantic level, combines SI with signal characteristics, and uses machine learning or signal processing algorithms to accurately determine the type and intensity of interference, laying the foundation for subsequent interference elimination and resource adjustment.

Energy-efficient signal processing technology is committed to reducing system energy consumption while ensuring SC performance. Power-adaptive adjustment technology dynamically adjusts the signal transmission power according to SI and channel conditions, and promotes the development of SC in a green and sustainable direction.

4.4. Summarize and Analyze

While future technological advancements will enhance the capabilities of SC, it still faces a series of challenges in practical deployment.

• Limitations in model interpretability
  Although DL models perform excellently in semantic extraction, their “black-box” nature makes the decision-making process difficult to interpret. In high-risk SC scenarios, the lack of interpretability may raise doubts about system reliability. When deviations occur in semantic understanding, it is hard to trace the root cause—hindering system debugging, trust establishment, and compliance.
• High computational costs
  The training and inference of complex semantic perception models (e.g., large-scale pre-trained language models) require massive computational resources and energy consumption. Deploying such models on resource-constrained edge devices or large-scale AIoT nodes may fail to achieve real-time responses due to limitations in computing power, storage, and battery capacity. This conflicts with the goals of “lightweight” and “high efficiency” pursued by SC. As a result, the system can only operate on some high-performance nodes in practice, limiting its universality and scalability.
• Challenges of data scarcity and domain adaptation
  The performance of ML and DL models heavily relies on large-scale, high-quality training data. In many professional fields, semantic data may face issues such as scarcity, high annotation costs, or privacy sensitivity—leading to insufficient model training or poor generalization ability. Additionally, semantic understanding models are often trained for general domains. When confronted with professional terminology, specific contexts, or cultural differences, the accuracy of their semantic representation may decline significantly, affecting the precision of ORA strategies.
• Generalization and robustness in dynamic environments
  SC environments are highly dynamic and heterogeneous. Although the system can adjust based on real-time data, models may exhibit fragility when facing new semantic scenarios outside the distribution of training data, extreme network conditions, or adversarial inputs. This can result in semantic parsing errors or improper RA. Further research is needed to ensure that models maintain robustness in complex and changing environments.

In summary, while the deep integration of future technologies into SC has brought revolutionary improvements to semantic perception and ORA, limitations in areas such as model interpretability, computational costs, data dependence, and environmental adaptability remain obstacles that must be overcome for its widespread, reliable, and trustworthy deployment. Future research should, while improving performance, focus on developing SC frameworks that are more lightweight, interpretable, and adaptable to small datasets and dynamic environments—ultimately realizing the true intelligence and universality of SC.

5. Case Studies on Optimal Resource Allocation for Semantic Communication

5.1. Intelligent Unmanned Equipment

Intelligent unmanned equipment, represented by unmanned aerial vehicles (UAVs), relies on real-time perception data interaction to support autonomous reconnaissance, monitoring, and collaborative decision-making. SC allows unmanned platforms to transmit task-critical information while filtering redundant raw data, which effectively reduces the consumption of onboard communication and computing resources. Benefiting from flexible deployment, UAV-borne SC faces highly dynamic topologies and time-varying air--ground channels, where ORA is essential to guarantee reliable execution of semantic-driven tasks.

In [95], the optimization objectives summarized in the survey include maximizing network throughput, minimizing energy consumption, and improving task completion reliability for UAV-assisted services. The allocated resources cover transmission power, spectrum sub-channels, UAV flight trajectories, and edge computing resources. According to the survey, mainstream ORA algorithms involve convex optimization, heuristic solutions, and deep reinforcement learning. The considered scenario focuses on general UAV-enabled wireless networks supporting diverse sensing and data-relay missions. Typical performance metrics surveyed are energy efficiency, throughput, task success rate, and latency. In [33], the discussed ORA objectives target energy-saving, coverage enhancement, and quality-of-service guarantee for UAV-assisted communication systems evolving from 5G toward 6G. The managed resource set consists of spectrum resources, transmit power, UAV three-dimensional deployment positions, and offloading computing resources. The reviewed ORA approaches include model-driven mathematical optimization and data-driven intelligent learning methods. The considered application scope covers UAV-aided ground coverage, emergency communication, and multi-UAV collaborative sensing scenarios. Commonly adopted evaluation metrics include coverage rate, energy consumption, spectral efficiency, and end-to-end latency. It highlights open challenges of ORA under high-mobility UAV conditions, such as fast channel variation and limited onboard payload. Despite these advances, several practical challenges persist:

• Energy constraints—Battery-powered UAVs face trade-offs between transmission power and flight endurance.
• Real-time CSI acquisition—Accurate channel state information is difficult to obtain in high-mobility aerial links, leading to sub-optimal or unstable allocation decisions.
• Security vulnerabilities—UAVs operating in adversarial environments risk jamming and spoofing attacks that corrupt semantic knowledge alignment.
• Scalability—Swarm scenarios require distributed ORA algorithms to avoid centralized control bottlenecks, yet current game-theoretic or DRL-based methods suffer from high computational overhead and slow convergence.

5.2. Smart Cities

Smart city systems integrate massive heterogeneous edge sensing devices, including roadside sensors, urban monitoring cameras, and edge computing terminals, generating large-scale multi-modal perception data. Traditional bit-centric RA cannot distinguish task importance and semantic value, resulting in low resource utilization in complex urban environments. Semantic-aware ORA enables differentiated scheduling according to semantic task priority, which is crucial for intelligent urban perception and edge service delivery.

In recent studies, practical ORA implementations have been widely explored for smart city semantic edge networks. In [30], a task-offloading-enabled semantic-aware ORA framework is proposed for urban edge networks. The optimization objective is to maximize semantic task execution efficiency under latency and energy constraints. The managed resources include wireless bandwidth, transmit power, and edge computing offloading resources. This work adopts an iterative, convex optimization-based RA algorithm. The scenario targets multi-task semantic perception and edge computing in urban heterogeneous networks. Evaluation metrics include task completion latency, semantic accuracy, and system energy consumption. Experimental results show that the semantic-aware ORA scheme achieves lower latency and higher task accuracy compared with conventional generic RA methods. In [89], an edge AIGC-enabled ORA strategy is designed for intelligent urban edge services. The optimization goal is to balance model caching efficiency and semantic transmission quality under limited edge storage and bandwidth resources. The optimized resources involve edge cache space, communication bandwidth, and computing scheduling resources. A model-aware dynamic resource scheduling algorithm is developed for urban edge AIGC semantic services. The considered scenario covers urban intelligent sensing, content generation, and real-time semantic interaction. System throughput and cache hit ratio are adopted as key metrics. The proposed method improves the stability and efficiency of urban edge semantic service delivery. Key practical obstacles include:

• Heterogeneous device constraints—Urban IoT sensors vary widely in computation, storage, and energy, making uniform ORA strategies ineffective.
• Data privacy and silos—Municipal agencies often cannot share raw data due to privacy regulations, complicating the construction of shared knowledge bases essential for SC.
• Dynamic load imbalance—Peak-hour communication demands in dense urban areas can overwhelm edge nodes, necessitating rapid semantic-aware load balancing that current frameworks do not fully support.
• Lack of standardized semantic metrics—Without agreed-upon measures of semantic utility, cross-system ORA comparisons and interoperability remain difficult.

5.3. Smart Vehicle and Internet of Vehicles

The Internet of Vehicles (IoV) requires ultra-low-latency and high-reliability semantic interaction for autonomous driving, including road environment perception, obstacle recognition, and vehicle-to-everything (V2X) collaborative decision-making. Different from traditional bit transmission, IoV SC focuses on transmitting decision-related core semantic features, which imposes new task-oriented requirements on dynamic ORA.

Several representative ORA implementations have been proposed for vehicle SC scenarios. In [69], an adaptive semantic reconstruction and resource scheduling framework is developed for task-oriented vehicular SC. The optimization objective is to maximize semantic reconstruction accuracy while satisfying stringent IoV latency constraints. The allocated resources include transmission bandwidth and computing resources for semantic encoding and reconstruction. A task-adaptive dynamic resource adjustment strategy is designed based on network real-time status. The scenario focuses on real-time driving perception semantic transmission in dynamic vehicle networking environments. Evaluation metrics include semantic similarity, object detection accuracy, and end-to-end latency. Simulation results demonstrate that the semantic-aware ORA method outperforms static allocation in high-mobility vehicular scenarios. In [80], a multi-objective dynamic ORA algorithm is proposed for IoV systems. The optimization targets include minimizing transmission latency and improving communication reliability under high-speed vehicle movement. The optimized resources include channel bandwidth and transmit power. A multi-objective joint optimization strategy is adopted to adapt to time-varying IoV network topology. The considered scenario covers high-dynamic V2X and vehicle-to-infrastructure semantic transmission. The performance metrics are network throughput, transmission success rate, and latency. The proposed dynamic allocation scheme effectively enhances the stability of vehicular SC under fast channel variations. However, real-world deployment of ORA in IoV faces critical hurdles:

• Ultra-low latency requirements—The automatic driving decision requires an end-to-end delay not higher than milliseconds, leaving minimal margin for complex ORA computations.
• High mobility and frequent handovers—Rapid topology changes due to vehicle speed make CSI prediction and resource reservation unreliable.
• Spectrum scarcity and interference—Dense vehicular platooning scenarios create severe co-channel interference, yet semantic-aware interference management remains largely at the theoretical stage.

6. Future Research Direction

Deepen the theoretical foundation: Current SI measurement lacks a mature framework. Future work should develop accurate, universal models to quantify semantic value and importance, along with evaluation metrics. This will provide a scientific basis for optimal ORA in SC. Defining the semantic channel capacity in complex environments such as multi-user scenarios and fading channels, and establishing a unified and computable semantic value evaluation model to support the theoretical optimization of ORA, represent critical areas of investigation.

Technological innovation: One of the core challenges in SC is the accurate extraction and transmission of SI. Advancing semantic understanding and processing enables precise identification of critical information for ORA. AI-driven techniques like deep learning and reinforcement learning further improve adaptability in dynamic SC environments. Designing lightweight semantic encoders for resource-constrained IoT devices and leveraging deep reinforcement learning to realize semantic-aware DRA that adapts to time-varying semantic importance are key innovative directions.

Model establishment: Performance evaluation of ORA in SC requires a comprehensive framework incorporating both traditional metrics (spectrum and energy efficiency) and semantic-specific indicators (accuracy, delay, transmission quality). Research should emphasize dynamic resource allocation under constrained conditions to optimize real-time communication efficiency and transmission quality. Under multiple constraints such as spectrum and energy, the core model construction task involves jointly optimizing semantic accuracy, latency, and reliability, as well as designing an online learning mechanism that can adaptively allocate resources based on dynamic semantic content and channel states.

Expansion of application scenarios: The ORA for SC will face increasingly diverse application scenarios and heightened performance demands. Future research should explore how to customize ORA models to suit different use cases, such as healthcare, transportation, and smart cities. Additionally, innovative applications in sectors like finance, education, and entertainment should be examined, expanding the versatility and impact of SC technologies.

Communication security and privacy: With the growing volume and complexity of data in SC systems, ensuring the security and privacy of communications becomes paramount. Future research should focus on methods to protect sensitive information during transmission, addressing challenges related to data confidentiality, integrity, and user privacy. Preventing SI from being tampered with or misunderstood during transmission, and embedding privacy-preserving mechanisms in semantic encoding and processing to avoid the leakage of sensitive information are security and privacy issues that must be addressed.

In summary, the future of ORA in SC presents rich opportunities for research across multiple domains, from theoretical advancements to practical applications. Addressing these challenges will pave the way for more efficient, secure, and adaptive systems in a variety of industries.

7. Conclusions

This paper conducts an in-depth investigation into ORA for SC, briefly analyzes TC systems and SC systems, and clarifies the differences between them. The historical development and distinctive characteristics of SC are systematically reviewed. Furthermore, this paper explains the importance of ORA in SC and discusses the challenges encountered in this domain.

This paper reviews a large body of literature related to ORA in SC, deeply explores the key technologies of ORA for SC, and summarizes the current status of related research in recent years. This includes resource demand prediction technology based on semantic understanding, which provides important preliminary support for accurate RA. Resource optimization technology is based on SI processing and transmission. By removing data redundancy, the data transmission volume is reduced, and resource utilization efficiency is improved. Through the collaborative optimization of multiple resources, the performance of the SC system is effectively improved. Dynamic adaptive resource adjustment technology realizes dynamic ORA through real-time monitoring and feedback mechanisms, combined with resource adjustment algorithms such as game theory and reinforcement learning, to adapt to complex and changing communication environments. A multi-dimensional task-oriented ORA framework is conceptualized to illustrate how transmission performance can be optimized by integrating available resources to meet diverse task needs, and an exemplary algorithmic flow for realizing this framework is also presented.

A comprehensive outlook on the development trend of ORA for SC is given. With the development of edge computing and artificial intelligence in the IoT, RA will become more intelligent and efficient through deep integration with AI technology, intelligent semantic perception, automated resource management, and optimization. Adaptation to future networks, including RA under ultra-high-speed and low-latency requirements and resource coordination in air–ground integrated networks, points out the direction for the development of SC in future networks. Signal processing technology reduces system energy consumption while ensuring SC performance.

Through analysis of application cases in multiple fields, such as intelligent unmanned equipment, smart cities, and smart vehicles, the effectiveness and practicality of ORA for SC are further verified. During tasks performed by unmanned equipment, communication stability and efficiency are assured. In the construction of smart cities, efficient RA such as intelligent traffic control and security monitoring is realized. In smart vehicles, Internet of Vehicles communications and autonomous driving systems provide assistance with the technical development of smart vehicles.

In summary, with the increase in the amount of information, the richness of information content, and the increase in application scenarios, the resources involved in communication systems have also increased. ORA for SC is the key to solving these practical problems. It is also the core driving force behind promoting SC from theory to widespread application. It focuses on the entire communication process, from information collection to transmission, reception, and processing. It rationally allocates resources in all aspects, which significantly improves communication reliability and efficiency, and provides a solid guarantee for efficient and interactive future communication systems.