Source-linked AI summary
Agentic RF Intelligence: Multi-Timescale 6G Sensing and Reasoning with On-Device Foundation Models
Jaron Fontaine, Jelle De Moerloose, Xander Vanparys, Anton Lambrecht, Eli De Poorter, Adnan Shahid
TL;DR
Wireless events can outlast neither the slow response of language models nor the constraints of passing raw RF signals directly to them. The paper proposes a dual-loop on-device system that pairs rapid RF perception with slower agentic reasoning; its prototype provides initial evidence that the fast loop can continue operating while the agent reasons and steers sensing.
Problem
Wireless events can last microseconds to milliseconds, while language models respond much more slowly and raw RF signals are too high-rate and low-level for direct LLM interpretation.
Source: §1 · p. 2Method
Results
The prototype provides initial evidence that learned perception can run in the fast loop and be invoked or steered by an agent without putting reasoning on the real-time path.
Source: §5 · p. 6Takeaways & Limitations
The measurements show a roughly 4–6 orders of magnitude latency gap between physical-layer inference and LLM tool interactions, with the fast loop remaining operational as the agent reasons.
Source: §4 · p. 5The benefits of the design remain unquantified against scripted or rule-based baselines and require repeated trials with ground-truth diagnoses and concurrent-load measurements.
Source: §4 · p. 5
Abstract
from arXiv · showhide
Future 6G systems share a central challenge with Physical AI: combining sensing, reasoning, and action on network-edge infrastructure in rapidly changing physical environments under strict latency, compute, and energy constraints. Addressing this challenge requires multi-timescale intelligence, combining fast perception that tracks wireless phenomena within milliseconds with slower reasoning that directs sensing and adapts network policies and resources over seconds. In this paper, we present a vision for Agentic RF Intelligence based on a dual-loop architecture that decouples fast, locally autonomous wireless sensing (e.g., detecting spectrum dynamics, interference, and signal sources) from slower agentic reasoning and orchestration. We provide an on-device prototype in which the full stack runs on a single NVIDIA Jetson Thor connected to a B200 Mini USRP. The fast loop processes IQ samples using signal-processing tools and wireless physical-layer foundation models (WPFM), while a local Large Language Model (LLM) asynchronously interprets RF events, invokes tools and steers sensing. Our experiments confirm the separation in timescales, as WPFMs operate at millisecond latency, while LLM interactions take seconds to minutes. The fast loop remains operational during agentic reasoning, providing initial evidence for the feasibility of this decoupled architecture. We conclude by outlining extensions toward memory-driven self-improvement, world models, network control, and multi-agent operation.
1 Introduction
The introduction argues that wireless agents need perception fast enough for transient RF events, alongside slower reasoning that chooses measurements and actions. It proposes connecting wireless physical-layer foundation models with agentic reasoning in a single on-device system.
1 Introduction
- The paper frames 6G autonomy as sensing environments, reasoning about them, and deciding what to do next, including which measurements to perform and tools to invoke.The introduction presents agentic AI as a framework for this behavior.Source: §1 · p. 1
- Wireless events can last microseconds to milliseconds, while language models respond in hundreds of milliseconds to seconds, so slow perception can miss events and raw signals overwhelm the reasoning layer.Prior systems expose higher-level information to agents, making it manageable but limiting what can be perceived.Source: §1 · p. 2
- The open challenge is to connect intelligent physical-layer perception, agentic decisions, and RF tools on-device under compute, memory, power, and timing constraints.The proposed system would let an agent select, configure, and combine sensing capabilities as needed.Source: §1 · p. 2
- The paper proposes a dual-loop vision and presents Agentic-USRP, a proof of concept on an NVIDIA Jetson Thor with a B200 Mini USRP, evaluating WPFMs and open-source LLMs.It also discusses extensions toward wireless world models, simulation, and multi-device and multi-agent systems.Source: §1 · p. 2
2 Agentic RF Framework: Dual-Loop Architecture and On-Device Challenges
The framework separates rapid, locally autonomous RF sensing from slower agent reasoning and orchestration, while allowing the agent to steer sensing through tools. Its on-device design must manage timescale mismatch, resource contention, representation limits, and action safety.
2 Agentic RF Framework: Dual-Loop Architecture and On-Device Challenges
- The fast loop continuously processes high-rate RF observations, uses signal-processing tools and WPFMs to extract measurements and compact representations, and asynchronously reports relevant changes to the agent.It is intended to operate near the physical interface under latency, reliability, and power constraints.Source: §2 · p. 3
- The slow loop combines perceptual state with memory, external knowledge, and potentially an RF World Model to reason about wireless conditions and invoke sensing, analysis, or configuration tools.It remains outside the real-time path while selecting actions over longer horizons.Source: §2 · p. 3
- The loops close the perception-reasoning cycle: the agent steers what the fast loop observes, and those observations guide later agent decisions.The framework connects the agent to perception, expert tools, and wireless actions.Source: §2 · p. 3
- Wireless timescales, shared compute and power, and the tension between raw-data volume and lossy abstraction require local autonomy, resource prioritization, and physically grounded summaries.Excessive compression can hide rare events or distribution shifts.Source: §2 · p. 3
- Agent actions that configure wireless hardware or system parameters must be validated against physical, operational, and real-time constraints while retaining flexibility for unseen conditions.The framework also identifies delayed decisions whose validity may need checking.Source: §2 · pp. 3–4
3 An On-Device Agentic RF Prototype
Agentic-USRP implements a background fast loop and an asynchronous agent loop on an edge device, with tools for sensing, analysis, memory, and control. The prototype keeps radio operations independent of LLM reasoning latency, while identifying limits in safety and resource sharing.
3 An On-Device Agentic RF Prototype
- Agentic-USRP runs on an NVIDIA Jetson Thor coupled to a B200 Mini USRP, which supports real-time sensing over 56 MHz bandwidth.The proof of concept demonstrates an agent using on-device RF perception to answer higher-level sensing questions and adapt measurements.Source: §3 · p. 4
- The background fast-loop worker supports triggers, one-shot jobs, and fixed-period monitors, each returning results or aggregated reports for requested sensing tasks.Examples include an occupancy threshold trigger, signal classification, and a 60-second band observation.Source: §3 · p. 4
- A thread-safe, sequence-ordered event bus delivers fast-loop notifications asynchronously, so radio operation never waits for the LLM.Policies expire automatically, but the prototype does not re-check that a triggering signal is still present before the agent acts.Source: §3 · p. 4
- The local LLM accesses RF sensing, analysis, memory, and fast-loop control through a common interface, while the agent selects WPFM targets and tasks and the fast loop computes embeddings on-device.Raw embeddings may be retained in memory rather than exposed to the LLM.Source: §3 · p. 4
- The LLM submits validated policies and bounded tool calls rather than accessing the radio or generating executable code, but serialized radio tasks and unreserved resources can delay acquisitions.Agent decisions are asynchronous and do not interrupt ongoing sensing, but this does not guarantee resource isolation.Source: §3 · p. 4
- The evaluation benchmarks WPFM inference latency and throughput and assesses local LLM reasoning length, response latency, and generation throughput on the Jetson Thor.The prototype is evaluated over the air using the B200 Mini USRP.Source: §3 · p. 4
4 Preliminary Results and Insights
The prototype tests whether WPFM perception and local LLM reasoning can operate in separate fast and slow loops on-device. Measurements show millisecond-scale perception alongside much slower agent interactions, while several benefits remain unquantified.
WPFM benchmarking
- The live USRP-to-WPFM pipeline sustains up to 56 Msps only in isolation, while data movement, buffering, and quantization remain end-to-end bottlenecks.A compact 0.82M-parameter CNN has 0.42 ms single-window latency; 2.91M-parameter transformer encoders take 1.36 ms for interference classification.Source: §4 · p. 5
On-device LLM evaluation
- The LLM evaluation spans spectrum observation, signal investigation with additional tool selection, and asynchronous event investigation followed by on-device measurements.Each scenario was tested with and without explicit reasoning.Source: §4 · p. 5
- Across five local LLMs, all completed asynchronous event investigation, while reasoning-enabled Nemotron 3.5 30B-A3B (lightning) had the lowest mean latency at 37.0 s and highest throughput at 37.8 tokens/s.Without reasoning, qwen3.5:9b was fastest at 18.9 s; disabling reasoning reduced latency for every model, and gemma4:12b completed only Scenario C with reasoning enabled.Source: §4 · p. 5
Timescale separation
- Physical-layer inference takes sub-milliseconds to milliseconds, whereas LLM tool interactions take seconds to minutes—a gap of roughly 4–6 orders of magnitude.The agent is therefore suited to operator-timescale investigation rather than real-time radio control, while the fast loop continues processing as the agent reasons and steers perception.Source: §4 · p. 5
Evaluation gaps
- The design’s benefits remain unquantified pending scripted and rule-based comparisons, repeated trials with ground-truth diagnoses, and concurrent-load latency, missed-event, memory, and power measurements.The validity of queued events before action also remains to be checked.Source: §4 · p. 5
Open questions
- Whether WPFMs offer on-device benefits over task-specific models for agent-selected tasks or continual learning from agent-labeled observations remains to be assessed.The passage distinguishes these possible benefits from prior accuracy and generalization gains.Source: §4 · p. 6
5 Conclusions & Future Work
The paper presents Agentic RF Intelligence as a dual-loop framework separating locally autonomous physical-layer perception from slower on-device reasoning. Agentic-USRP provides initial evidence for this premise and motivates extensions spanning self-improvement, world models, control, and multi-agent systems.
Conclusion
- Agentic RF Intelligence separates fast, locally autonomous physical-layer perception from slower on-device reasoning.Agentic-USRP provides initial evidence that learned perception can run in the fast loop and be invoked or steered by an agent without putting reasoning on the real-time path.Source: §5 · p. 6
- Asynchronous execution lets the fast loop remain operational while the agent reasons and adapts subsequent sensing, highlighting the importance of end-to-end pipeline design beyond model inference.Source: §5 · p. 6
Future Work
- Future directions include memory-driven recursive self-improvement, RF-world-model counterfactual simulation, validated and traceable network actions, and multi-agent operation across sensing nodes.The proposed self-improvement loop would use accumulated observations and conclusions as labeled cases to fine-tune the WPFM and refine sensing strategies.Source: §5 · p. 6
A Appendix: Representative Agentic-USRP Execution Trace
The representative trace illustrates an agent iteratively selecting RF sensing operations from prior observations while the fast loop performs background monitoring. Its interference explanation is a hypothesis, not a verified diagnosis.
Trace interaction
- The agent begins with a spectrum survey, investigates a suspicious emission with RF tools, then delegates background monitoring to the fast loop.The trace is presented as an example of interaction, not a verified diagnosis.Source: A Appendix: Representative… · pp. 7–8
Survey and investigation
- The survey reports activity across seven bands, and anomaly detection finds an emission centered at 2487.5 MHz with about 14.2 dB SNR and 18.3 kHz reported bandwidth.Source: A Appendix: Representative… · p. 7
- The agent zooms in and requests modulation classification; the result reports an amplitude-varying envelope and a 64QAM top hypothesis.The characterization reports about 3.7 MHz occupied bandwidth and 11.6 dB SNR.Source: A Appendix: Representative… · p. 8
Background monitoring
- After 20 seconds of monitoring at a 2-second acquisition cadence, the report gives 6.2% mean occupancy, a 40.4 dB peak at 2426 MHz, and multiple Wi-Fi signals around 2442–2448 MHz.The cadence yields ten snapshots, so signals appearing between acquisitions may be missed.Source: A Appendix: Representative… · p. 8
Interpretation and caveats
- The observed activity is consistent with legitimate BLE and Wi-Fi traffic, but the proposed congestion-or-contention explanation is unverified because the anomaly lay outside the monitored band and link-loss data were not correlated.The trace was manually trimmed, including low-level logs and full reasoning traces.Source: A Appendix: Representative… · pp. 7–8