PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic Semantic Sensing for Resource-Adaptive AI-RAN

Zhongqin Wang, Xiaoqi Zhang, Nan Yang, Kai Wu, J. Andrew Zhang, Y. Jay Guo

arXiv:2610.07829v1cs.AIeess.SP

TL;DR

Existing semantic sensing generally fixes configurations and observation schedules before inference, limiting adaptation to evolving task-level evidence. Agentic SemS uses semantic belief and continuation utility to select feasible profiles and decide when to stop sensing. On Widar3.0, it reduces normalized cumulative sensing cost by 25.33% while achieving 85.79% Macro-F1 versus full-sequence High.

  • Problem

    Existing SemS typically fixes sensing configurations and observation procedures before inference, so evolving task-level information does not guide subsequent acquisition.

    Source: §I · p. 1

  • Method

    Agentic SemS uses evolving semantic belief and continuation utility to select communication-feasible profiles and support semantic early exit.

    Source: §I · p. 2

  • Results

    25.33% lower normalized cumulative sensing cost with 85.79% Macro-F1 versus full-sequence High on Widar3.0 with six emulated sensing profiles.

    Source: §I · p. 2

  • Takeaways & Limitations

    Adaptive profile selection and semantic early exit provide complementary control of sensing cost and observation duration in the evaluated setting.

    Source: §VI · p. 13

Abstract

from arXiv · show

Semantic sensing (SemS) acquires task-relevant information rather than reconstructing complete physical information. Existing SemS formulations typically operate open loop: sensing configurations and observation schedules are fixed before inference and cannot respond to evolving task-level evidence. We propose Agentic SemS, a closed-loop framework for AI-enabled radio access networks (AI-RANs) that controls sensing within a communication-feasible profile set. A profile-conditioned causal Transformer updates the semantic belief from streaming observations, while key-value caching enables efficient state updates across profile changes without repeatedly processing the complete history. A semantic utility network estimates the task-level benefit of acquiring the next observation block under each feasible profile after accounting for sensing cost. The resulting continuation utilities jointly support next-profile selection and semantic early exit, adapting sensing configuration and duration to evolving evidence. The expected semantic gain is further related to conditional mutual information, providing a value-of-information interpretation of continued online sensing. Experiments on Widar3.0 with six emulated sensing profiles show that, in comparison with full-sequence High, the resource-efficient Agentic setting reduces normalized cumulative sensing cost by 25.33% while achieving 85.79% Macro-F1. At the same utility checkpoint, semantic early exit provides a further 12.35% cost reduction over adaptive sensing without early exit, with a 0.97-percentage-point Macro-F1 decrease.

I. INTRODUCTION

Agentic SemS addresses open-loop sensing by using evolving task-level evidence to adapt sensing profiles and duration. It frames sensing as task-relevant acquisition within communication-feasible profiles.

I. INTRODUCTION

  • Open-loop configurations and observation schedules cannot respond to changing informativeness or inference uncertainty, potentially spending resources after reliable inference or stopping before sufficient evidence.Existing SemS acquires task-relevant information but generally determines sensing configurations and procedures before inference.Source: §I · p. 1
  • Agentic SemS uses evolving semantic belief to guide acquisition within communication-feasible profiles, with continuation utility supporting profile selection and semantic early exit.Expected semantic gain is linked to conditional mutual information as a value-of-information interpretation.Source: §I · p. 2
  • 25.33% lower normalized cumulative sensing cost with 85.79% Macro-F1 versus full-sequence High on Widar3.0 with six emulated profiles.The evaluation also compares profile adaptation and semantic early exit.Source: §I · p. 2

II. PRELIMINARIES AND MOTIVATION

  • The preliminaries frame SemS as task-oriented acquisition for downstream inference rather than complete physical-information reconstruction.The sensing-profile space is constrained by communication requirements supplied by the RAN.Source: A. Semantic Sensing Model · p. 2

A. Semantic Sensing Model

  • A profile-dependent sensing model maps sensing through the environment and channel to a task output without reconstructing task-irrelevant physical information.Profiles can specify subcarrier usage, sampling rate, signal quality, temporal configuration, and signal representation.Source: A. Semantic Sensing Model · p. 2

B. From Open-Loop SemS to Closed-Loop Agentic SemS

Agentic SemS closes the open-loop feedback gap by using semantic state to guide feasible sensing actions. The loop adapts profiles and stopping decisions as observations arrive.

B. From Open-Loop SemS to Closed-Loop Agentic SemS

  • Open-loop SemS fixes sensing configuration and procedure, whereas Agentic SemS uses semantic belief to adapt feasible profiles and terminate when further acquisition has low predicted value.This jointly adapts sensing configuration and duration to task requirements.Source: B. From Open-Loop SemS to Closed-Loop… · p. 3

1) Closed-Loop Semantic Sensing Task:

  • The closed-loop task updates task-level inference from observations, selects a feasible profile for the next block, and uses predicted value to adapt sensing duration.Each sensing decision affects subsequent observations.Source: 1) Closed-Loop Semantic Sensing Task · p. 3

2) Sensing Profiles and Feasible Actions:

  • Each sensing profile specifies an operating point through its observation, configuration, and resource attributes; feasible continuation profiles reflect hardware availability and reconfiguration constraints.The feasibility indicator captures which profiles are available from the current operating state.Source: 2) Sensing Profiles and Feasible Actions · p. 3

4) Joint Profile Selection and Semantic Early Exit:

  • Predicted utilities jointly determine whether sensing continues and which profile acquires the next block; each new block updates the semantic state before the next decision.If sensing stops, the current task prediction is returned.Source: 4) Joint Profile Selection and… · p. 3

B. Semantic Belief Network

The semantic belief network incrementally converts profile-conditioned streaming observations into a task-level state. Its CPI-based inputs and shared embedding structure accommodate differing sensing profiles.

B. Semantic Belief Network

  • The belief network uses each coherent processing interval as an observation unit, with the resulting state supplying task-level information for subsequent sensing decisions.The CPI provides temporal support for motion-sensitive features while preserving online inference resolution.Source: B. Semantic Belief Network · pp. 3–4

2) Signal Preprocessing and Observation Encoding:

3) Profile-Context Encoding:

  • A profile context compactly describes sensing configuration and resource requirements, allowing the network to interpret observations according to how they were acquired.The same profile encoder represents active profiles for belief updates and candidate profiles for utility evaluation.Source: 3) Profile-Context Encoding · p. 4

4) Semantic Belief Update and Prediction:

  • Causal attention forms each latent representation only from current and preceding profile-conditioned inputs; classification yields class-posterior beliefs, while regression can use a task-specific head with uncertainty.The latent representation and belief together form the semantic state.Source: 4) Semantic Belief Update and Prediction · p. 4

5) Weighted-Prefix Learning Objective:

The learning objective supervises intermediate prefixes as well as complete sequences, while a candidate-conditioned utility network evaluates feasible profiles from the current semantic evidence.

5) Weighted-Prefix Learning Objective:

  • Intermediate prefixes receive supervision at decision intervals after a minimum observed fraction, with later prefixes weighted more heavily and full sequences supervised separately.Online operation processes observations as they arrive and does not require the complete sequence length.Source: 5) Weighted-Prefix Learning Objective · pp. 4–5

C. Semantic Utility Network

  • The semantic utility network predicts the usefulness of continuing under each feasible profile from the current semantic state, supporting continued acquisition and profile choice.The network evaluates how useful each feasible profile is for the next observation block.Source: C. Semantic Utility Network · p. 5

1) Candidate-Conditioned Utility Input:

  • For each candidate, the utility input concatenates the latent representation, semantic belief, and profile embedding so evaluation reflects both semantic evidence and sensing characteristics.The profile embedding is produced by the shared profile encoder.Source: 1) Candidate-Conditioned Utility Input · p. 5

2) Semantic Utility Prediction:

The semantic utility network estimates the task-level benefit of another observation block under each candidate profile, accounting for sensing cost. Its supervised target combines task-loss reduction and normalized sensing-resource cost.

2) Semantic Utility Prediction:

  • The shared utility network predicts each candidate profile’s expected semantic improvement after accounting for sensing cost, guiding the next-block choice.A larger predicted utility indicates greater expected improvement; when sensing continues, the agent applies the highest-utility candidate.Source: 2) Semantic Utility Prediction · p. 5

3) Semantic Utility Objective:

  • The supervised utility target combines candidate-block task-loss reduction with normalized sensing-resource cost.For classification, the task loss is cross-entropy; the additional block and ground-truth target are used to construct utility targets, not during online sensing.Source: 3) Semantic Utility Objective · p. 5
  • Positive semantic gain means the additional observation reduces task loss, while negative gain indicates semantic degradation.The task loss is task-specific for tasks other than classification.Source: 3) Semantic Utility Objective · p. 5
  • The utility network is trained with mean-squared error to approximate the conditional target mean given semantic state and candidate context.The training objective uses decision–candidate samples.Source: 3) Semantic Utility Objective · p. 5

4) Profile Selection and Semantic Early Exit:

At each decision epoch, Agentic SemS updates candidate utilities to select a feasible next profile or stop sensing after early-exit eligibility. Under an ideal-posterior assumption, expected log-loss reduction from the next block equals its conditional mutual information with the task label.

4) Profile Selection and Semantic Early Exit:

  • The controller updates semantic state and candidate utilities after acquisition, then selects the feasible profile with the highest predicted utility when sensing continues.The selected profile adapts subsequent sensing-resource usage to the evolving semantic state.Source: 4) Profile Selection and Semantic… · p. 5
  • Semantic early exit becomes eligible after Nmin acquired CPIs, while Nmin and τ govern eligibility and the stopping rule.Before Nmin, the controller continues sensing while allowing profile adaptation at each decision epoch.Source: 4) Profile Selection and Semantic… · p. 5

D. VoI Interpretation of the Continuation Utility

  • Under the ideal-posterior assumption, the next block’s expected log-loss reduction equals the conditional mutual information between the task label and that block.An individual realized semantic gain may be negative even though expected gain is nonnegative.Source: D. VoI Interpretation of the… · pp. 5–6
  • The learned utility estimates expected task-loss reduction minus sensing cost, conditioned on semantic state and candidate context.Its conditional-mutual-information interpretation depends on the ideal-posterior assumption; labels and future observations are used only to construct training targets.Source: D. VoI Interpretation of the… · p. 6

IV. SYSTEM DESIGN OF AGENTIC SEMS

The system-design section describes a reference implementation of Agentic SemS, including network architectures, training, and online operation. Its sensing front end extracts CPI-level features and conditions profile representations on sensing configuration and available measurements.

IV. SYSTEM DESIGN OF AGENTIC SEMS

  • The reference implementation covers Agentic SemS network architectures, training, and online operation, and is evaluated on a sensing task with multiple profile configurations.Source: §IV · p. 6

1) CPI-Level Feature Extraction:

  • The CPI input includes CSI packet sequences and compact radio measurements, with the active profile determining which compact-radio inputs are available.Compact radio indicators can include received power, link quality, and interference measurements.Source: 1) CPI-Level Feature Extraction · p. 6
  • The sensing front end forms a 135-dimensional CPI observation from a 128-dimensional spectral feature and seven spectrum-derived statistics.CSI processing uses phase compensation and adaptive range–Doppler spectral processing; compact-radio measurements use mean-centering and a 128-point FFT.Source: 1) CPI-Level Feature Extraction · p. 6

2) Profile-Context Construction:

  • The profile context encodes configuration attributes, normalized per-CPI resource score, and available measurement branches into a 192-dimensional embedding.The profile encoder is shared by the belief and utility networks and is frozen for reuse by the utility network after joint training.Source: 2) Profile-Context Construction · p. 6

3) Network Architecture:

The belief network processes CPI observations with a causal Transformer, while a shared candidate-conditioned utility network estimates continuation utility. The implementation trains the belief model in two stages and retains a KV cache across profile transitions.

3) Network Architecture:

  • A causal Transformer processes observation, active-profile, and positional embeddings to produce a semantic belief at each CPI.The reference architecture uses four self-attention layers, six heads, a 192-dimensional model, and a 768-dimensional feed-forward layer.Source: 3) Network Architecture · pp. 6–7

4) Two-Stage Weighted-Prefix Training:

  • Two-stage weighted-prefix training first learns fixed-profile representations, then improves robustness using mixed-profile trajectories with feasible adjacent transitions.Stage 1 uses 70% weight on complete sequences; Stage 2 uses 50%, with 30% fixed-profile and 70% mixed-profile trajectories.Source: 4) Two-Stage Weighted-Prefix Training · p. 7

5) KV-Cached Inference:

  • Streaming inference retains a KV cache for at most 128 CPI-level tokens, and the cache persists across profile transitions.Cache capacity is an implementation parameter that can be increased for longer sensing horizons, subject to memory.Source: 5) KV-Cached Inference · p. 7

1) Candidate-Conditioned Utility Input:

  • For each feasible candidate profile, the utility input concatenates the belief network’s latent representation and semantic belief with the candidate-profile embedding.The candidate-conditioned input has dimension 384 + C.Source: 1) Candidate-Conditioned Utility Input · p. 7

2) Utility Network Architecture:

  • A lightweight multilayer perceptron with hidden dimensions 256 and 128 predicts continuation utility using one network shared across candidate profiles.The first linear layer is followed by LayerNorm and GELU; the second hidden layer also uses GELU.Source: 2) Utility Network Architecture · p. 7

3) Utility-Target Generation:

  • Utility targets are generated at epochs spaced by four CPIs using the next four-CPI block, with the implementation-level λcost absorbing the common block-cost factor.Cumulative sensing cost is still calculated at the CPI level, and each λcost value requires regenerated targets and a separately trained utility checkpoint.Source: 3) Utility-Target Generation · p. 7

4) Utility Network Training:

This section specifies utility-network training and online profile/exit control, then describes the Widar3.0 data, emulated profiles, and normalized sensing-resource costs.

4) Utility Network Training:

  • The utility network trains for 30 epochs with record-level sampling and candidate-level examples, using MSE, AdamW, and validation-selected minimum-MSE checkpoints for each λcost.Each epoch samples 2,000 training records; each record contributes decision epochs with a complete next observation block and feasible candidates.Source: 4) Utility Network Training · p. 7

5) Online Profile Adaptation and Semantic Early Exit:

  • After each acquired observation block, the controller updates semantic state and utilities, selects the highest-utility continuation profile, or exits when maximum utility does not exceed τ after Nmin.The reference transition rule permits the current profile and adjacent lower- and higher-resource profiles; sensing also stops at environment-imposed availability or CPI-budget limits.Source: 5) Online Profile Adaptation and… · p. 7

V. EXPERIMENTAL EVALUATION

  • The evaluation tests the reference implementation on six-class WiFi gesture recognition using Widar3.0.The experimental section introduces the dataset, profile realizations, controller configuration, and evaluation protocol.Source: §V · p. 7

1) Dataset and Partitioning:

  • Widar3.0 contains 165,724 records from 16 subjects, split by class into 116,006 training, 16,572 validation, and 33,146 test records, with all profiles of each source record kept together.The validation split selects network checkpoints and controller parameters; the test split is reserved for final evaluation.Source: 1) Dataset and Partitioning · p. 8

2) Emulated Sensing Profiles:

  • Six emulated profiles vary feature type, subcarrier count, sampling rate, and SNR; higher configurations retain richer information at greater measurement and processing demand.Lightweight uses RSSI, while the other five profiles use CSI; each source recording contributes one receive-antenna stream across all six profiles.Source: 2) Emulated Sensing Profiles · p. 8, Figure 2 · p. 8

3) Normalized Sensing-Resource Cost:

  • Normalized per-CPI cost is a relative resource index computed from fixed resource weights, not a calibrated measure of physical energy consumption.The reference weights normalize [1, 1, 0.75, 0.5, 1]ᵀ; Table I reports costs for the ordered profiles, with unit cost for High.Source: 3) Normalized Sensing-Resource Cost · p. 8, Table I · p. 8

4) Controller Parameter Selection:

The evaluation compares fixed and adaptive sensing controls using recognition and resource-use metrics, with controller parameters selected on validation data. Fixed profiles trade higher Macro-F1 for greater sensing cost.

4) Controller Parameter Selection:

  • Controller parameters are selected on validation data, with Nmin = 8 CPIs for exit-enabled settings and fixed before test evaluation.Validation selects (λcost, τ) = (0.005, −0.20) for the high-F1 setting and (0.005, 0) for the resource-efficient reference setting.Source: 4) Controller Parameter Selection · p. 9

B. Evaluation Metrics

  • Macro-F1 measures recognition; P, R, and C summarize average profile cost, observed-sequence fraction, and cumulative sensing cost, respectively, with 1−P, 1−Robs, and 1−C denoting reductions versus full-sequence High.The metrics distinguish profile adaptation, observation duration, and their combined resource impact.Source: B. Evaluation Metrics · p. 9

C. Evaluation Protocol and Baselines

  • The protocol separates profile adaptation from early exit: fixed High with early exit varies duration, while adaptive profiles without exit vary the profile sequence.The complete controller combines both mechanisms; all exit-enabled methods use Nmin = 8 CPIs and thresholds selected on validation data.Source: C. Evaluation Protocol and Baselines · p. 9

1) Fixed-Profile Recognition and Cost:

  • Macro-F1 rises from 51.02% for Lightweight to 90.70% for High, while normalized cumulative cost rises from 0.236 to 1.000.The six full-sequence fixed-profile baselines use the same final causal Transformer checkpoint and differ only by profile.Source: 1) Fixed-Profile Recognition and Cost · p. 9
  • The fixed-profile recognition–cost hierarchy provides multiple operating points for adapting sensing-resource use to the evolving semantic state.Each full-sequence fixed-profile baseline uses one profile throughout the complete sequence.Source: 1) Fixed-Profile Recognition and Cost · pp. 9–10

2) Adaptive Profile Selection:

On Widar3.0, adaptive profile selection and semantic early exit reduce sensing cost, while recognition and semantic gain vary with gesture class and accumulated evidence.

2) Adaptive Profile Selection:

  • Adaptive profiles without early exit achieve 86.76% Macro-F1 at 14.81% lower cost than full-sequence High, while observing the complete sequence.The normalized cumulative sensing cost is 0.852; most resource reduction in the high-F1 Agentic setting comes from profile adaptation.Source: 2) Adaptive Profile Selection · p. 10

3) Semantic Early Exit:

  • 12.35% further relative cost reduction comes from early exit over adaptive sensing without it, with Macro-F1 decreasing 0.97 percentage points.At the same λcost = 0.005 checkpoint, cost changes from 0.852 to 0.747 and Macro-F1 from 86.76% to 85.79%; versus full-sequence High, the complete controller reduces cost by 25.33% at 85.79% Macro-F1.Source: 3) Semantic Early Exit · p. 10

4) Per-Gesture Results:

  • Across gesture classes, cost reductions versus full-sequence High range from 12.36% for Clap to 33.72% for Sweep, while per-class F1 gaps range from 2.46 to 7.76 percentage points.Ground-truth labels are used only to compute per-class statistics and remain unavailable to the online controller.Source: Table III · p. 10, 4) Per-Gesture Results · p. 10

1) Belief-Network Comparison:

  • The KV-cached causal Transformer reaches 91.29% Macro-F1, near MobileViT-XXS at 91.93%, while supporting causal prefix processing and incremental inference.The comparison uses the same 49,718-record Widar3.0 subset with a 70/30 train/test split; the GRU reaches 77.40%.Source: 1) Belief-Network Comparison · p. 10

2) Profile-Dependent Recognition:

  • Recognition generally improves from Lightweight to High, although some gesture classes retain strong recognition under lower-resource profiles.The complete-sequence MobileViT-XXS confusion matrices across six profiles illustrate class-separability changes with sensing information.Source: 2) Profile-Dependent Recognition · p. 10, Figure 5 · p. 11

4) Recognition versus Observed Prefix:

  • Macro-F1 rises from 39.68% at a 24.03% average observed fraction to 90.70% at 100%, with improvements diminishing near the complete sequence.The intermediate observed-fraction/Macro-F1 pairs are 43.94%/63.66%, 64.09%/78.12%, and 83.46%/86.85%.Source: Figure 7 · p. 11, 4) Recognition versus Observed Prefix · p. 11

3) Semantic-Gain Analysis:

  • Across 29,941 candidate evaluations, 62.77% of semantic gains are positive, 36.66% negative, and 0.57% numerically zero.The overall mean and median gains are 0.115 and 0.036. A profile’s gain varies with the current semantic state and acquired observation; negative realized gains do not contradict nonnegative expected VoI.Source: 3) Semantic-Gain Analysis · p. 10, Figure 6 · p. 11

F. Parameter Selection and Early Exit

Agentic SemS couples adaptive profile selection with semantic early exit, adjusting sensing resources as semantic evidence evolves. The section selects a minimum observation budget and reports validation and test results across stopping thresholds, fixed profiles, initialization choices, and profile transitions.

1) Role of the Minimum Observation Budget:

  • At Nmin = 4, 97.50% of validation records exit early and Macro-F1 is 22.22%; Nmin = 8 achieves 85.92% Macro-F1 at normalized cumulative cost 0.7487.Results at 8 and 12 CPIs are identical; 16 CPIs changes Macro-F1 by only 0.02 percentage points and cost by less than 0.001, so Nmin = 8 is selected.Source: 1) Role of the Minimum Observation… · pp. 11–12

G. Fixed-Profile Early-Exit Baselines

  • Under fixed High, semantic early exit reduces cumulative cost by 11.49% with a 0.54-percentage-point Macro-F1 decrease; under fixed Medium-High, the reduction is 13.14% with a 0.82-point decrease.Because the profile stays fixed, the cost reduction is attributable to shorter sensing duration.Source: Table VI · p. 12, G. Fixed-Profile Early-Exit Baselines · p. 12

1) Initial-Profile Sensitivity:

  • Starting from Lightweight yields cumulative sensing cost 0.661 and Macro-F1 80.09%, while starting from High yields cost 0.800 and Macro-F1 88.85%.The random initialization used in the main reference experiments lies between these cases, at 85.79% Macro-F1 and C = 0.747.Source: Table VII · p. 12, 1) Initial-Profile Sensitivity · p. 12

1) Role of the Minimum Observation Budget:

  • At λcost = 0.005, τ = 0 gives the resource-efficient validation reference, lowering normalized cumulative cost from 0.851 to 0.749; the highest validation Macro-F1 is 86.74% at τ = −0.20.The search evaluates six utility checkpoints and nine τ values per checkpoint.Source: 2) Validation-Set Search over λcost and · p. 12

2) Profile Transitions:

  • Across 304,371 continuation decisions, the controller retains its profile 54.45% of the time, moves higher-resource 30.82%, and moves lower-resource 14.72%.Both upward and downward transitions occur; High remains unchanged in 90.12% of its continuation decisions.Source: Table VIII · p. 12, 2) Profile Transitions · p. 12

VI. CONCLUSION

  • 85.79% Macro-F1 accompanies a 25.33% reduction in normalized cumulative sensing cost versus full-sequence High; early exit adds a 12.35% cost reduction over adaptive sensing without exit, with a 0.97-percentage-point Macro-F1 decrease.The resource-efficient Agentic setting combines adaptive profile selection and semantic early exit.Source: §VI · p. 13

Related papers

Loading 2610.07829v1…