SenseCFN: Dual Paradigms in Networking

• SenseCFN is a hybrid paradigm that fuses decision-aware semantic state synchronization in Compute-First Networking with cell-free integrated sensing and communication capabilities.
• In the compute-first variant, it leverages lightweight transformer encoding and a semantic deviation index to efficiently trigger state updates for optimized offloading decisions.
• In the cell-free ISAC variant, it jointly optimizes resource allocation and beamforming to enhance communication spectral efficiency while improving sensing accuracy.

SenseCFN denotes two distinct but leading-edge paradigms at the intersection of wireless networking, semantic state synchronization, and integrated sensing. The term appears in the context of (1) decision-aware semantic state synchronization for Compute-First Networking (CFN), and (2) cell-free integrated sensing and communication networks, fusing cell-free massive MIMO with Joint ISAC. Both lines share the goal of optimizing communication overhead and resource usage in future networks, but address fundamentally different layers and timescales.

1. Decision-Aware State Synchronization in Compute-First Networking

In Compute-First Networking, edge Access Points (APs) make per-task offloading decisions contingent on the perceived state of distributed Service Nodes (SNs). High-frequency state reporting burdens uplink channels; sparse reporting degrades decision accuracy under variable loads. Conventional approaches such as fixed periodic updates or Age-of-Information (AoI) metrics treat all state deviation symmetrically, optimizing for temporal freshness rather than relevance. SenseCFN reconceptualizes status synchronization as a decision-consistency problem: only state changes with potential to flip the offloading outcome warrant a network update (Qi et al., 3 Jan 2026).

1.1 Lightweight Semantic State Encoding

SNs collect a six-dimensional measurement vector at each slot:

• $\tilde{n}_{\rm idle}$: Idle CPU core count.
• $\tilde{q}_{\rm len}$: Normalized queue length.
• $w_{\rm head}$: Queue head-of-line wait time.
• $\delta_{\rm AoI}$: Age-of-Information of last update at AP.
• $\tilde{c}_{\rm last}$: Normalized most-recent finished task workload.
• $\tilde{A}_{\rm est}$: Exponential moving average of task inter-arrival times.

A lightweight Transformer encoder $E_\theta$ with two layers and four attention heads reduces this to a compact semantic vector $\mathbf{z}_t \in \mathbb{R}^{d_{\rm sem}}$, where $d_{\rm sem}=3$ achieves stable high performance and minimal overhead.

1.2 Semantic Deviation Index (SDI) and Update Triggering

Decision impact from semantic drift is quantified by the semantic deviation index: $$\mathrm{SDI}_t = \frac{ \| \mathbf{z}_t - \hat{\mathbf{z}}_t \|_2 }{ \| \hat{\mathbf{z}}_t \|_2 + \varepsilon }$$ where $\mathbf{z}_t$ is the current semantic state, $\hat{\mathbf{z}}_t$ is the AP-cached (possibly stale) semantic state, and $\varepsilon>0$ prevents division by zero. Large SDI signals a drift likely to alter the AP's decision boundary.

The SN's update policy, parameterized by a small MLP $\pi_{sn}$, takes $$[\mathbf{z}_t \oplus \mathrm{SDI}_t \oplus \mathrm{QoS}_t]$$ as input, where $\mathrm{QoS}_t$ encodes uplink congestion, and emits $p_{up}$, the update probability. Binary actions ($a^{sn}_t$) are sampled based on $p_{up}$. This learned, congestion-aware strategy subsumes both static and dynamic thresholding.

1.3 Offloading Policy at the Access Point

Upon each task arrival, the AP forms its input state vector $$[\hat{\mathbf{z}}_{t_k} \oplus \delta_{\rm AoI,t_k} \oplus \hat D_{down}(t_k) \oplus \mathbf x_{\rm loc}(t_k)]$$. A second MLP $\pi_{ap}$ outputs the probability $p_{loc}$ of local execution, balancing staleness, local resource load, and expected transmission delays. This model adaptively compensates for semantic staleness, prioritizing reliability under increasing AoI.

1.4 Centralized Training with Distributed Execution (CTDE)

Both the update and offloading policies are jointly trained in a centralized manner, but deployed in a fully distributed fashion, facilitating end-to-end optimization in a shared semantic space. Training alternates between hybrid label generation (using hardwired safety constraints and expert traces for policy supervision) and large-batch stochastic gradient descent minimizing a composite loss: $$\mathcal{L}_{\rm total} = \mathcal{L}_{\rm imit} + \mathcal{L}_{\rm sys} + \lambda_{\rm sem} \mathcal{L}_{\rm cons}$$ $\mathcal{L}_{\rm imit}$ enforces imitation of expert labels for update and offloading; $\mathcal{L}_{\rm sys}$ codifies task success, penalizes update rate, and latency; $\mathcal{L}_{\rm cons}$ ensures semantic manifold smoothness.

2. Cell-Free Integrated Sensing and Communication Networks

Independently, the term SenseCFN also refers to cell-free integrated sensing and communication networks, a paradigm that merges distributed (cell-free) massive MIMO architecture with integrated sensing and communication (ISAC) functionality (Galappaththige et al., 27 Feb 2025). These networks consist of geographically-distributed, low-power Access Points (APs) interconnected via fronthaul, jointly serving user equipment (UEs) and enabling multi-static radar-style sensing in shared time-frequency resources.

2.1 System and Signal Model

A canonical SenseCFN deploys $M$ downlink and $N$ uplink APs, each with $L$ antennas, to simultaneously serve $K$ single-antenna users and $T$ point-targets (sensing). All APs are connected to a central processing unit (CPU) via constrained fronthaul. The downlink transmit waveform at time $t$,

$$\mathbf{x}(t) = [\mathbf{x}_1^T(t),\dots,\mathbf{x}_M^T(t)]^T$$

aggregates both communication and sensing components per AP: $$\mathbf{x}_m(t) = \sum_{k=1}^K \mathbf{w}_{mk} q_k(t) + \sum_{t'=1}^T \mathbf{s}_{mt'}(t)$$ subject to per-AP power constraints.

Communication channels are modeled as i.i.d. Rayleigh fading, while the radar/sensing link to each target models round-trip delay, Doppler, and angle-of-arrival (AoA).

2.2 Key Performance Metrics

Communication performance employs sum spectral efficiency: $$\mathrm{SE}_{\rm com} = \sum_{k=1}^K \log_2(1+\mathrm{SINR}_k)$$ where $\mathrm{SINR}_k$ explicitly accounts for both inter-user interference and ISAC waveform interference.

Sensing performance is characterized by multi-static SINR, total sensing SE, Cramér–Rao lower bound (CRLB) for parameter estimation, and classical radar metrics ($P_D$, $P_{FA}$, RCS, $\Delta r$, $\Delta \theta$).

3. Joint Resource Allocation and Beamforming

The design challenge in SenseCFN is the joint optimization of communication and sensing—balancing beamformers $\{\mathbf{w}_k\}$ and sensing vectors $\{\mathbf{s}_t\}$ to maximize a weighted sum of communication spectral efficiency and inverse CRLB for sensing accuracy, under power and fronthaul constraints: $$\begin{aligned} & \max_{\{\mathbf{w}_k\},\{\mathbf{s}_t\}} && \alpha \sum_k \log_2(1+\mathrm{SINR}_k) - (1-\alpha) \sum_t \mathrm{CRLB}_t \ & \text{s.t.} && \sum_k \|\mathbf{w}_{mk}\|^2 + \sum_t \|\mathbf{s}_{mt}\|^2 \le P_m^{\max} \ &&& C_{\rm fronthaul}(\{\mathbf{w},\mathbf{s}\}) \le \tilde C_m \end{aligned}$$ $\alpha \in [0,1]$ controls the SE-sensing trade-off.

Convex and non-convex strategies include successive convex approximation (SCA), semidefinite relaxation (SDR), alternating optimization (AO), block coordinate descent, and, for scalable scenarios, deep reinforcement learning (DRL) and Riemannian-manifold optimization.

4. Algorithms and Representative Results

Model-based optimization (AO–SCA) converges to local optima in 10–20 iterations, while learning-based (e.g., DRL) approaches frame the beamforming allocation as an episodic reward-maximizing problem at the CPU actor. Riemannian geometry-based optimizers exploit the structure of beamforming matrices for improved constraint satisfaction.

In a representative scenario (9 APs, 4 UEs, 3 targets, 8 antennas/AP, 10 MHz), SenseCFN achieves:

• Communication SE: 40 bps/Hz (no sensing), 38 bps/Hz at $\alpha=0.7$
• Detection probability $P_D$ (for $P_{FA}=10^{-4}$): 0.85 (vs. 0.6 for co-located ISAC)
• 50% reduction in CRLB delay error as APs are increased from 4 to 12

Macro-diversity from distributed APs demonstrably improves both coverage and sensing resolution (Galappaththige et al., 27 Feb 2025).

5. Technical Challenges and Research Directions

SenseCFN deployments face several tightly-coupled challenges:

• Sub-nanosecond synchronization is required across distributed APs for joint beamforming and precision sensing.
• Multi-target interference and echo association complicate signal separation, especially with dense deployments and limited fronthaul.
• Fronthaul bottlenecks, both in capacity and latency, necessitate on-device pre-processing and edge fusion mechanisms.
• Near-field effects demand models going beyond the planar-wave approximations, particularly as aperture sizes and deployment density increase.

Emerging trends include the adoption of reconfigurable intelligent surfaces (RIS), THz-band operation, fluid/holographic MIMO, and ML-driven adaptive beamforming for low-overhead, environment-agnostic CSI tracking.

6. Comparative Summary and Open Problems

SenseCFN Variant	Core Functionality	Layer/Focus	Key Innovations	Principal Bottlenecks
CFN Decision-Aware (Qi et al., 3 Jan 2026)	Semantic state sync for offloading decisions	Edge computation; protocol	Semantic encoding, SDI, CTDE	Staleness-robustness, scaling, label calibration
Cell-Free ISAC (Galappaththige et al., 27 Feb 2025)	Distributed joint comm-sensing via APs	Physical, radio resource	Joint beamforming, multi-static diversity	Synchronization, fronthaul, multi-target

CFN-centric SenseCFN delivers up to 99.6% task success with 70–96% reduction in update frequency over AoI/content-aware baselines, while cell-free ISAC–centric SenseCFN achieves dual gains in coverage and sensing under joint resource control. Limitations in both variants relate to scaling, transferability to multi-node or general workload/topology settings, and practical adaptation to dynamic environments.

• "Decision-Aware Semantic State Synchronization in Compute-First Networking" (Qi et al., 3 Jan 2026)
• "Cell-Free Integrated Sensing and Communication: Principles, Advances, and Future Directions" (Galappaththige et al., 27 Feb 2025)