Goal-Oriented Communication (GoC)

• Goal-oriented Communication (GoC) is a paradigm that prioritizes transmitting task-critical information over traditional bit-accurate reconstruction.
• It integrates communication, computation, and control through frameworks like GoT and Information Bottleneck to optimize decision-making under resource constraints.
• Applications span IoT, robotics, and 6G networks, demonstrating improved efficiency through co-designed sampling, scheduling, and actuation strategies.

Goal-oriented Communication (GoC) is a communication paradigm in which communication is not treated as an end in itself, but as a means to achieve a specific task or goal at the receiver. Across the recent literature, it is positioned against the Shannon-style objective of accurate bit-by-bit or symbol-by-symbol reconstruction and aligned instead with Weaver’s effectiveness problem: whether communicated information affects conduct in the desired way. In this view, the relevant question is not only whether information is fresh, accurate, or semantically meaningful, but whether it is useful for downstream inference, decision making, control, learning, or actuation under resource constraints (Getu et al., 2023, Li et al., 2023).

1. Conceptual foundations and scope

GoC is commonly described as a shift from transmitting information faithfully to transmitting information that is useful for accomplishing a task. In the tutorial and survey literature, it is discussed under the umbrella of goal-oriented semantic communication and associated most directly with the effectiveness level of communication, while some works distinguish it from semantic communication in the narrower sense by treating GoC as the task- and effectiveness-focused subset of semantic communication (Getu et al., 2023). A closely related distinction appears in the 6G-GOALS framework: semantic communication concerns transmitting the meaning of the data, pragmatic communication concerns the usefulness of that meaning in context, and GoC is the highest-level abstraction, asking whether the communication helps achieve the end goal (Strinati et al., 2024).

A recurrent theme is that the significance of information is not solely tied to its inherent meaning, but to the ultimate usage of that meaning for achieving particular goals. This leads to a design philosophy in which communication, computation, control, and learning are co-designed. In multi-agent systems, the corresponding question becomes what information is worth sending, given the agents’ shared objective, under partial observability, limited bandwidth, and decentralized decision-making (Li et al., 2023, Charalambous et al., 11 Aug 2025). In IoT-oriented work, the same idea is expressed as a shift from fidelity, accuracy, or reliability toward semantics extraction and goal accomplishment, especially when the receiver requires a feature, decision, or action rather than full source reconstruction (Zhang et al., 2022).

The literature also emphasizes that GoC is not reducible to a single application class. It appears in remote estimation and control of Markov processes, edge inference, sensor scheduling, networked control, multimedia transmission, robotic fault detection and recovery, UAV obstacle avoidance, medium access, 6G AI-native networking, interplanetary and non-terrestrial networking, and multi-agent coordination (Talli et al., 2024, Assaad et al., 2024, Merluzzi et al., 2023, Uysal, 2024). This suggests that GoC is best understood as a systems principle: communication decisions are judged by task relevance and application utility rather than by source fidelity alone.

2. Metrics, goal value, and formal objective functions

A central issue in GoC is the choice of performance metric. Much of the literature begins from the observation that classical metrics such as Age of Information (AoI), Value of Information (VoI), Age of Incorrect Information (AoII), Age of Synchronization (AoS), Mean Square Error (MSE), and Cost of Actuation Error each capture part of the usefulness problem, but do not provide a single direct measure of goal attainment (Li et al., 2023). AoI measures staleness; VoI adds nonlinear penalties to freshness; AoII and AoS capture mismatch duration or synchronization; MSE captures reconstruction error. These metrics remain indirect when the real objective is successful decision making and actuation (Li et al., 2023).

To address this, the Goal-oriented Tensor (GoT) was proposed as a tensor-based framework for significance-oriented communications and later specialized as a direct semantic-performance metric for decision making under semantic mismatch. In its general form, GoT is built on the tuple

$\langle X(t), \hat{X}(t), \Phi(t) \rangle,$

where $X(t)$ is the source semantic status, $\hat{X}(t)$ is the receiver’s estimated status, and $\Phi(t)$ is a context or weighting term. With a decision strategy $\pi(\hat{X}(t))$, the GoT value is computed as

$$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$

where $\mathcal{C}_1$ is the status inherent cost, $\mathcal{C}_2$ is the decision gain, and $\mathcal{C}_3$ is the action cost (Li et al., 2023). In the sampler–decision maker setting, the same structure appears as

$$\mathrm{GoT}^{\pi_A}(t)=\left[C_1(X_t,\Phi_t)-C_2(\pi_A(\hat{X}_t))\right]^+ + C_3(\pi_A(\hat{X}_t)),$$

explicitly tying cost to the true semantic state, context, estimate, and decision policy (Li et al., 2023).

A second important line of formalization uses the Information Bottleneck (IB) principle. For edge learning, the IB objective

$X(t)$0

defines a compressed representation $X(t)$1 that preserves information relevant to the target $X(t)$2 while minimizing representation complexity (Pezone et al., 2022). In the tutorial literature, related IB formulations, robust IB, distributed IB, graph IB, and task-oriented rate-distortion are presented as formal tools for task-aware compression and decision-quality preservation (Getu et al., 2023).

A third family of metrics defines goal value and goal-effectiveness directly at the application layer. In the 6G coexistence setting, the goal value $X(t)$3 quantifies whether the application goal is being achieved, and the long-term goal-effectiveness is defined as

$X(t)$4

with a corresponding long-term goal cost

$X(t)$5

This formulation makes explicit that communication reliability and goal-effectiveness are not straightforwardly linked, because higher reliability can improve inference quality while worsening timeliness through lower finite-blocklength rate (Merluzzi et al., 2023).

3. Decision-theoretic formulations and scheduling policies

Many GoC models cast communication itself as a sequential decision variable. In the GoT-based perception–actuation loop, a sampler observes the semantic state $X(t)$6 and context $X(t)$7, while a decision maker acts on the estimate $X(t)$8. Their joint objective is to minimize the long-term average cost

$X(t)$9

leading to an infinite-horizon Decentralized Partially Observable Markov Decision Process (Dec-POMDP). Under the unichain assumption, an RVI-Brute-Force-Search algorithm is claimed to achieve the optimal joint deterministic policies $\hat{X}(t)$0 (Li et al., 2023).

Remote control of finite-state Markov processes provides another core formulation. In the push-based architecture, encoder and decoder form a two-agent Dec-POMDP with costly zero-delay communication; in the pull-based architecture, the decoder decides both whether to request an update and which control action to apply (Talli et al., 2024). The objective is to maximize discounted reward net of communication cost,

$\hat{X}(t)$1

showing directly that communication is valuable only insofar as it improves task reward enough to justify its price (Talli et al., 2024). In a related pull-based monitoring and control setting, the optimal GoC policy for a Markov process is described by a scheduling function

$\hat{X}(t)$2

so that the update interval depends on the last observed state and the task rather than on a fixed clock (Mason et al., 15 Jul 2025).

Goal-oriented sensor reporting yields comparable structures in IoT monitoring. For a non-linear dynamic system monitored by sensors and queried by clients, the edge node decides whether to poll one sensor or poll none, with the sole purpose of minimizing the expected MSE of future query responses. The problem is modeled as a POMDP, and a DRL scheduler uses a compact state consisting of the trace of the prior covariance and the elapsed time since each client’s last query. Its reward is explicitly tied to query-response MSE and to a no-query operating mode, so that the scheduler may choose that no communication is the best communication choice (Raghuwanshi et al., 2024).

The same decision-theoretic logic extends to medium access. Goal-oriented Multiple Access (GoMA) formulates the shared-channel problem for multiple intelligent nodes that each observe a local value of information. The expected reward combines the value of successful transmissions and a fixed transmission cost, and the resulting optimization is non-convex and may admit multiple Nash equilibria (Chiariotti et al., 26 Aug 2025). Best responses are threshold policies in which a node transmits only when its observed value exceeds an interference- and cost-dependent threshold, and the paper proposes both the Local Iterated Best Response Access protocol and the Bandit-based Emergent Threshold Adaptation algorithm for distributed optimization and learning (Chiariotti et al., 26 Aug 2025). A plausible implication is that, in GoC, medium access control is no longer separable from application value.

4. Semantic representations, compression, and generative reconstruction

A large branch of GoC research concerns the representation of task-relevant information. In IoT data compression, the central question is not how to reconstruct the original signal well, but how to enable the task well. This alters preprocessing, quantization, clustering, source coding, and feature extraction, because the compression rule must preserve utility rather than raw distortion (Zhang et al., 2022). The same perspective appears in edge learning, where IB-based encoding identifies the information in $\hat{X}(t)$3 that is relevant to the task variable $\hat{X}(t)$4, and stochastic optimization adapts communication and computation resources to satisfy energy, delay, and accuracy constraints (Pezone et al., 2022).

Generative and latent-variable models provide a different implementation path. GO-COM introduces a task-driven semantic pipeline in which a Vector Quantized Variational Autoencoder (VQ-VAE) compresses media data into discrete semantic representations at the transmitter, while the Goal-Oriented Semantic Variational Autoencoder (GOS-VAE) regenerates task-appropriate content at the receiver (Chao et al., 25 Feb 2025). Instead of targeting pixel-wise reconstruction, the receiver is evaluated by a pre-defined task-incentivized model, and imitation learning is used to measure whether regenerated samples preserve the behaviorally relevant semantics required for the downstream goal (Chao et al., 25 Feb 2025).

Diffusion-based GoC pushes the same principle into generative media transmission. Diff-GO$\hat{X}(t)$5 restricts forward diffusion to a pre-sampled finite Noise Bank shared between transmitter and receiver, so that only a semantic condition and a compact noise index need to be transmitted. For an NB size of 1000, the paper states that noise transmission requires only 10 bits, compared with about 384 KB for a traditional DDPM noised latent payload and around 4 KB for Diff-GO+ (Wanninayaka et al., 2024). The same work reports LPIPS $\hat{X}(t)$6 and FID $\hat{X}(t)$7 on Cityscapes, together with faster convergence than conventional diffusion-based GO-COM (Wanninayaka et al., 2024).

LaMI-GO takes a related approach in quantized latent space by combining a latent diffusion model with a VQGAN codebook. The transmitter sends masked latent codeword indices and a textual semantic description, and the receiver uses latent mixture integration to preserve known latent structure while regenerating only missing parts (Wijesinghe et al., 2024). On Cityscapes, the reported bandwidth is 8.50 KB for LaMI-GO at masking probability $\hat{X}(t)$8 and 9.39 KB at $\hat{X}(t)$9, with corresponding FID values 27.23 and 22.72, outperforming the compared GOCOM baselines in the reported settings (Wijesinghe et al., 2024).

Computation–communication co-design also appears in inference offloading. The recursive early-exit framework for edge inference treats GoC as a joint communication-computation optimization problem: at each step, the system decides whether to classify locally, compute another layer, or offload intermediate features to a server. It defines goal effectiveness as task success under a latency deadline and uses Q-learning to optimize early exits, computation splitting, and offloading under wireless constraints (Pomponi et al., 2024). This suggests that, in modern GoC systems, the “message” may be an intermediate representation or a task outcome rather than an explicit source reconstruction.

5. Closed-loop control, robotics, and networked systems

GoC is especially prominent in closed-loop systems, where communication affects future state evolution through control. In RIS-assisted networked control, a set of sensors observes dynamical processes and sends state estimates over an unreliable channel assisted by a reconfigurable intelligent surface. The joint objective is to optimize the controller policy and RIS phase policy so as to minimize a finite-horizon LQR regulation cost, not to maximize SNR or sum rate. The optimal control is certainty-equivalent, while the RIS phases are optimized through a reduced problem and an approximate semi-definite relaxation technique (Assaad et al., 2024). The relation to GoC is explicit: a channel configuration is “good” only if it improves regulation performance.

UAV obstacle avoidance under integrated sensing and communication offers a more tightly coupled example. In the GOSC framework, a base station uses a Kalman filter to predict UAV position, a Mahalanobis distance-based dynamic window approach to generate command-and-control signals under uncertainty, and an effectiveness-aware DQN to determine whether sensing and C&C should be transmitted based on their value of information (Liu et al., 2 Mar 2026). Sensing VoI is quantified by the reduction in uncertainty entropy of position estimation; C&C VoI is measured by contribution to navigation improvement (Liu et al., 2 Mar 2026). The reported outcome is the same 100% task success rate as a conventional always-transmit ISAC framework, with 92.4% fewer transmitted sensing and C&C signals and 85.5% fewer transmission time slots (Liu et al., 2 Mar 2026).

Robotic fault detection and recovery extends GoC to the full communication–computation–control loop. In that setting, the semantic representation for fault detection is a 3D scene graph, fault detection is based on spatial relationship changes, recovery motions are generated by a fine-tuned small LLM with LoRA and knowledge distillation, and a lightweight digital twin reconstructs only task-relevant object contours when fine-grained control refinement is required (Chen et al., 26 Jan 2026). Extensive simulations report up to 82.6% reduction in FDR time and up to 76% improvement in task success rate relative to state-of-the-art frameworks based on vision LLMs and LLMs (Chen et al., 26 Jan 2026).

At larger system scale, the 6G-GOALS approach positions GoC as a key paradigm for AI-native 6G networks. Its three pillars are AI-enhanced semantic data representation, foundational AI reasoning and causal semantic data representation for contextual relevance and goal-oriented effectiveness, and sustainability enabled by more efficient wireless services (Strinati et al., 2024). The proposed O-RAN-based architecture includes a Semantic Engine, Semantic RIC, RAN Semantic Plane, Application Plane, UE and edge components with AI capabilities, and a Knowledge Base (Strinati et al., 2024). In space networking, GoC is proposed as a natural design principle for interplanetary and non-terrestrial networks, where very long and variable delays, intermittent connectivity, and store-and-forward operation make the utility of information at delivery time more important than transparent end-to-end packet transfer (Uysal, 2024). The space vision is organized around three pillars: goal-oriented sampling and multi-user scheduling, goal-oriented grant-free access, and DTN-based flow control using metadata such as Age and Lifetime (Uysal, 2024).

6. Coexistence, security, and unresolved issues

Because GoC departs from classical communication design, it raises specific coexistence and security problems. In the 6G coexistence study, a goal-oriented edge inference system shares spectrum with a legacy eMBB video upload. The objective is to maximize the legacy user’s long-term average rate subject to a target goal-effectiveness for the inference service, using Lyapunov optimization over the legacy transmit power and the goal-oriented link’s target packet error rate (Merluzzi et al., 2023). A key analytical point is that communication reliability and goal-effectiveness are not straightforwardly linked: higher reliability can improve inference confidence while harming deadline satisfaction through lower finite-blocklength rate (Merluzzi et al., 2023). This result is often presented as a correction to the misconception that better PHY reliability automatically implies better task performance.

A different challenge is security leakage through timing. In pull-based remote monitoring and control of Markov processes, adaptive GoC scheduling creates a timing side channel even if message contents are protected. The eavesdropper observes transmission times, knows the scheduling rule, and uses forward-backward hidden Markov model inference to estimate the hidden state (Mason et al., 15 Jul 2025). The paper reports that a naive goal-oriented scheduler allows the eavesdropper to correctly guess the system state about 60% of the time, while heuristic defenses can halve the leakage with a marginal reduction of the benefits of goal-oriented approaches (Mason et al., 15 Jul 2025). The proposed countermeasures are ADE, which alternates between GoC and periodic scheduling depending on leakage thresholds, and PDE, which compresses the entropy of the scheduling policy so that multiple states map to the same waiting time (Mason et al., 15 Jul 2025).

Beyond these concrete problems, the literature repeatedly identifies foundational open issues. The tutorial and survey paper highlights the lack of a commonly accepted definition of semantics or semantic information, the lack of a unified semantic performance metric, semantic mismatch due to different knowledge bases, limited interpretability of deep-learning-based semantic coding, non-asymptotic analysis difficulties, real-time latency constraints, compatibility with existing BitCom infrastructure, coexistence with classical communication, and privacy and security vulnerabilities (Getu et al., 2023). The GoT framework points to heterogeneous goals, goal-oriented PHY-layer techniques, and perception–communication–computation–control co-design as major directions, including the possible need for higher-order GoTs to represent the full closed loop (Li et al., 2023). The 6G-GOALS work adds causal representation learning, domain adaptation for semantic mismatches, backward compatibility with legacy nodes, and sustainability as first-class concerns (Strinati et al., 2024). In multi-agent settings, scalability, safety, reliability, interpretability, and richer temporal and causal semantic models remain central open challenges (Charalambous et al., 11 Aug 2025).

Taken together, these lines of work define GoC not as a single algorithmic family but as a general reorientation of communication engineering. Its unifying principle is that communication should be optimized around the utility of information for a specified goal, with the consequence that sampling, representation, coding, scheduling, access, inference, and control are treated as parts of one task-driven system rather than as isolated layers.