Source-linked AI summary
Small Language Model enabled Autonomous agent for Language-Conditioned Cognitive Radar
Minhaj Uddin Ahmad, Zakia Zaman, Shunqiao Sun, Mizanur Rahman
TL;DR
Radar processing pipelines often rely on fixed rules that generalize poorly to high-level task descriptions and dynamic environments. This paper develops an SLM-driven, physics-grounded agent that selects and executes array-processing tools, demonstrating reliable cognitive control when supported by radar-specific prompting and tool execution.
Problem
Fixed-rule radar-processing pipelines generalize poorly to high-level task descriptions and dynamic environments, motivating more adaptable strategy selection.
Method
The paper develops an SLM-driven agent that maps natural-language commands to validated array signal-processing actions through radar-specific prompting and physics-grounded executable tools.
Results
Across six task categories and 180 trials, ablations show that radar-specific prompting and tool execution support reliable algorithm selection and numerical reporting.
Takeaways & Limitations
A small locally hosted language model can serve as a cognitive controller for radar signal-processing tools when guided by physical priors and executable computation.
Takeaways & Limitations
Evaluation is limited to synthetic ULA data and a fixed tool suite, without demonstrating robustness on hardware radar data, other array geometries, or complex propagation environments.
Abstract
from arXiv · showhide
Modern radar systems require adapting their processing strategies in response to changing interference, clutter, and data availability. This paper introduces a framework for a small language model (SLM)-driven autonomous agent designed for language-conditioned cognitive radar, functioning as an intelligent controller for a suite of array signal processing tools. Given a natural-language command, the agent extracts radar-operation-related cues, selects an appropriate sequence of signal-processing methods, configures parameters, and invokes executable tools for numerical computation. Experiments with a synthetic uniform linear array (ULA) radar demonstrate that, given a natural-language command, the agent performs meaningful algorithm selection across diverse scenarios for sidelobe control, jammer suppression, multiple-null beamforming, coherent-source handling, and low-snapshot direction-of-arrival (DOA) estimation. Ablation results show that radar-specific prompting and physics-grounded tool execution are both required for reliable decisions and hallucination-free numerical results.
1. INTRODUCTION
The paper frames language-conditioned cognitive radar as an SLM-driven autonomous agent that maps natural-language commands to array signal-processing actions. It combines a physics-grounded executable toolchain with radar-specific mechanisms to support reliable, constraint-aware numerical reporting.
- Motivation: Array-radar performance depends on selecting signal-processing strategies suited to operating conditions, but conventional designs generalize poorly to high-level task descriptions and dynamic environments.Examples include sidelobe-controlled beamforming for cluttered scenes and adaptive beamforming when needed.
- Framework: The proposed framework uses an SLM-driven autonomous agent as an intelligent user-facing controller for array signal-processing tools in language-conditioned cognitive radar.The agent is motivated by cognitive radar’s closed-loop adaptation paradigm and advances in language-model tool augmentation.
- Contributions: The paper formulates language-conditioned cognitive radar as a tool-augmented SLM-agent problem that maps natural-language commands to array signal-processing actions.This formulation is listed as the first main contribution.
- Contributions: A modular physics-grounded radar toolchain derives numerical claims from executable signal-processing routines rather than model-generated arithmetic.The toolchain is intended to support numerical reliability and reduce hallucinated reporting.
- Contributions: Radar-specific prompting, parameter validation, state management, and transcript logging enforce physical constraints and reduce hallucinated numerical reporting.These mechanisms are identified as part of the paper’s main contributions.
2. METHOD
The method uses a modular SLM-based autonomous agent to implement a language-conditioned cognitive-radar processing loop. It parses natural-language input into radar-relevant cues and selects processing actions from a predefined array-signal-processing library.
- 2. METHOD: The SLM-based autonomous agent is designed with a modular architecture for cognitive radar.Figure 1 presents the language-conditioned cognitive-radar processing loop.
- 2. METHOD: The loop begins by parsing input into structured cues including target and interference directions, clutter, snapshot support, and resolution constraints.These cues represent the radar-operation information used to guide subsequent processing decisions.
- 2. METHOD: Using the extracted cues, the SLM agent selects a processing action from a predefined library of array signal-processing methods.The architecture therefore connects language-conditioned cue extraction with radar algorithm selection.
2.1. Physics Tools
The Physics Tools module provides a physics-informed Python foundation for language-conditioned radar processing, combining array simulation, beamforming, and DOA estimation. Its algorithms support scenario-dependent weighting, interference suppression, and source-direction estimation while reducing hallucinations in numerical outputs.
- Module organization: The standalone NumPy/SciPy package groups tools into an array simulator, beamforming methods, and DOA estimation methods.These functions provide the physics-informed foundation for the SLM’s numerical outputs and help reduce hallucinations.
- Array simulation: The array simulator generates narrowband ULA snapshots from a standard signal model and serves as a surrogate for actual radar hardware.Windowing options include uniform, Dolph-Chebyshev, and Taylor weights for scenario- and requirement-dependent sidelobe control.
- Beamforming: Beamforming tools include DAS, MVDR, MPDR, and LCMV, with MVDR and MPDR minimizing interference-plus-noise power while preserving unit target-direction gain.DAS applies uniform or tapered steering-vector weights; MVDR and MPDR differ primarily in their covariance-matrix usage.
- DOA estimation: DOA tools include MUSIC, ESPRIT, and OMP, using subspace orthogonality or rotational invariance to estimate source directions.MUSIC identifies directions through steering-vector orthogonality with the noise subspace, while ESPRIT uses two overlapping subarrays without grid search.
2.2. SLM based Autonomous Agent
The autonomous agent uses a locally deployed, 4-bit quantized Qwen3.5 model and a ReAct-style tool-calling loop to control radar signal-processing workflows. Radar-specific prompting supplies algorithm-selection knowledge and physical constraints, while tool execution and persistent run context support reliable numerical processing.
- Model deployment: Qwen3.5 is deployed as a locally served, 4-bit quantized 9B-parameter SLM suitable for embedded radar platforms.The local Ollama deployment also supports offline operation and future model upgrades without changing the radar signal-processing pipeline.
- Agent loop: The agent follows a ReAct-style observe-reason-act loop that streams full conversation history and invokes tools until producing a turn without tool calls.The loop accepts reasoning, provider-schema tool calls, or both, then treats a tool-free turn as the final human-readable answer.
- Radar-specific prompting: The system prompt supplies radar expertise, algorithm operating conditions, selection rules, and physical constraints because the general-purpose SLM lacks radar-specific fine-tuning.Examples include snapshot requirements for MUSIC, jammer modeling, MVDR versus MPDR selection, Rayleigh resolution, source capacity, and MUSIC rank conditions.
- Execution constraints: Every tool call must be preceded by explicit reasoning, and final numerical results must come only from tool outputs.The loop inserts a nudge message when the model produces an empty answer, reinforcing non-empty final responses.
- Tool integration: A persistent RunContext preserves numerical results across stateless tool calls, enabling sequential simulate-then-process-then-analyze workflows.Subsequent handlers can reuse and update values stored during the same invocation.
3. EXPERIMENTAL EVALUATION
The evaluation tests both the agent’s decision-making reliability and the physics-grounded radar performance of its selected tools. Experiments use paraphrased natural-language tasks spanning six radar-processing categories, ablations, scenario-specific beamforming and DOA tests, and runtime measurements.
- Evaluation objectives: The evaluation measures tool or tool-sequence selection, valid tool calls, physically valid parameters, and avoidance of hallucinated numerical results.It also examines whether selected tools achieve requested sidelobe levels, jammer suppression, and DOA-estimation performance.
- Benchmark design: The benchmark covers six radar-processing categories using paraphrased commands, repeated trials, and expert-defined acceptable tool or tool-sequence labels.This design tests natural-language robustness rather than memorization and accounts for stochastic variation in SLM responses.
- Ablation study: The ablation compares the full agent with agents lacking radar-specific prompting or tool execution and with a fixed keyword-matching controller.Without domain-specific prompting, the SLM may confuse similar methods such as MVDR and MPDR; without tools, it cannot reliably produce numerical radar values.
- Sidelobe control: For a 30◦ scan direction and sidelobes below −30 dB, the agent selected delay-and-sum beamforming with a Dolph-Chebyshev taper.The taper is described as providing the narrowest mainlobe for a prescribed equiripple sidelobe level.
- Adaptive beamforming and DOA estimation: The experiments evaluate adaptive beamforming under high-power jamming and DOA estimation under limited snapshots, closely spaced sources, and coherent-source conditions.The DOA discussion distinguishes MUSIC, ESPRIT, and OMP according to covariance reliability, snapshot stability, and sparse-recovery suitability.
- Runtime: Runtime measurements include SLM generation, tool-call parsing, physics execution, and final response generation, with local model generation identified as the main bottleneck.The experiments use an Intel i9-14900F CPU, 64GB of RAM, a quantized 9B-parameter SLM, and lightweight NumPy/SciPy processing.
N L SNR True DOAs Selected tool Estimated DOAs / RMSE · 4. CONCLUSION AND FUTURE DIRECTION
The paper presents an SLM agent that reliably controls radar array signal-processing tools when guided by a carefully engineered system prompt. Its evaluation remains limited to synthetic ULA data and a fixed radar-tool suite.
- 4. CONCLUSION AND FUTURE DIRECTION: The system uses an SLM agent for radar array signal processing.The agent is designed as a controller for radar signal-processing operations.
- 4. CONCLUSION AND FUTURE DIRECTION: Qwen 3.5:9b serves as a locally hosted cognitive controller for a suite of signal-processing tools.The paper characterizes the model as small and locally hosted.
- 4. CONCLUSION AND FUTURE DIRECTION: Reliable control depends on a carefully engineered system prompt.The prompt encodes algorithm-selection priors, physical sanity checks, and behavioral-discipline constraints.
- 4. CONCLUSION AND FUTURE DIRECTION: The system prompt encodes algorithm-selection priors for the agent.These priors are part of the prompt conditions associated with reliable control.
- 4. CONCLUSION AND FUTURE DIRECTION: The system prompt includes physical sanity checks and behavioral-discipline constraints.Both elements are identified as components of the engineered prompt.
- 4. CONCLUSION AND FUTURE DIRECTION: The current evaluation uses synthetic ULA data and a fixed radar-tool suite.The paper identifies these evaluation choices as remaining limitations.