Small Language Model enabled Autonomous agent for Language-Conditioned Cognitive Radar

The paper, Small Language Model enabled Autonomous agent for Language-Conditioned Cognitive Radar, introduces a framework for an SLM-driven autonomous agent designed to serve as an intelligent controller for array signal processing tools in cognitive radar systems.

Modern radar systems require dynamic adaptation to changing conditions such as interference, clutter, and data availability. While traditional methods rely on fixed rules, lookup tables, or operator expertise, these designs do not generalize well to high-level task descriptions or dynamic environments. Cognitive radar offers a solution by formulating the system as an intelligent closed-loop system that learns from interactions with the environment and adapts its sensing and processing strategy accordingly.

The proposed solution is an SLM-driven autonomous agent, which functions as a user-facing controller for array signal processing tools. This framework leverages recent advancements in tool-augmented agents, specifically utilizing the ReAct (Reasoning and Acting) paradigm.

The system operates via a Language-Conditioned Cognitive Radar Processing Loop (Fig. 1). The process begins with a natural-language input, which the SLM agent parses to extract cues such as target direction, interference direction, and snapshot support. Based on these cues, the agent performs several critical steps:

1. Decision Making: Select an appropriate sequence of signal-processing methods from a predefined library (e of conventional and adaptive beamforming, subspace-based DOA estimation).

2. Tool Invocation: Configure parameters and invoke executable tools for numerical computation.

3. Execution and Feedback: The selected physics-based module outputs quantitative radar metrics (e.g., sidelobe levels, null depths, angular estimates). These metrics are returned to the agent for validation, explanation, and generating the final response, completing a closed-loop feedback cycle.

** 1. Physics Tools Module:** This is a standalone Python package built with NumPy and SciPy that provides a physics-informed foundation for numerical outputs. It includes three groups of tools:

• Beamforming Tools: Includes Delay-and-Sum (DAS), Minimum Variance Distortionless Response (MVDR), Minimum Power Distortionless Response (MPDR), and Linearly Constrained Minimum Variance (LCMV).

• DOA Estimation Tools: Includes Multiple Signal Classification (MUSIC), Estimation of Signal Parameters via Rotational Invariance Techniques (ESPRIT), and Orthogonal Matching Pursuit (OMP).

• An array simulator is also provided to generate synthetic data for testing.

** 2. SLM-based Autonomous Agent:** The reasoning engine utilizes Qwen3.5, a 4-bit quantized 9B parameter model served locally via Ollama. It follows the ReAct paradigm (observe-reason-act). Crucially, the agent never performs arithmetic directly; all numerical computations... are delegated to physics tools.

** 3. System Prompt Engineering:** The system prompt is the primary knowledge encoding mechanism, containing five parts:

• Defining the agent as an expert radar array signal processing agent.

• Embedding domain knowledge (e.g., If jammer or high-power interference to simulate ula with the jammer modeled).

• Specifying inviolable constraints (e.g., Rayleigh resolution limits).

• Requiring that every tool call be preceded by explicit reasoning.

The authors highlight four main contributions:

(i) Formulating language-conditioned cognitive radar as a human-in-the-loop refinement process.

(ii) Implementing a modular physics-grounded radar toolchain, ensuring that all numerical claims are derived from executable routines rather than model-generated arithmetic.

(iii) Designing mechanisms to enforce physical constraints, reduce hallucinated numerical reporting, and make each agent decision auditable.

(iv) Evaluating the system on an expanded natural-language benchmark.

The system was evaluated based on two perspectives: agentic decision-making (e.g., algorithm selection, valid tool calls) and physics-grounded radar performance (e.g., achieving requested sidelobe levels).

Ablation Study: Table 2 shows the results of an ablation study comparing the full SLM agent against baselines. The full SLM agent achieved an End-to-end success rate of 98.9%, demonstrating high reliability across the tasks.

Performance Results:

• Sidelobe Control: The agent successfully translates natural language into appropriate beampattern designs, such as selecting a Dolph-Chebyshev taper for a specified sidelobe level (Table 3).

• Adaptive Jamming Suppression: The system demonstrates the ability to select the correct adaptive beamformer (MVDR, MPDR, or LCMV) based on training data availability and successfully achieves deep null suppression at jammer locations (Table 4).

• DOA Estimation: Under constraints like limited snapshots, the agent does not default to a single algorithm. For instance, it selects ESPRIT when L is small relative to N, which is a stable choice when the sample covariance is poorly conditioned (Table 5).

The system's local deployment feasibility was also demonstrated, showing that the use of a quantized 9B-parameter SLM suggests a feasible path toward deployment on embedded GPU platforms.

As a diligent AI researcher, I have analyzed the methodological framework presented in this paper. The core strength of this work—the integration of a Small Language Model (SLM) with physics-grounded, executable tools—is not specific to cognitive radar; it represents a paradigm shift in how large language models interact with complex, constrained domains.

The improvements outlined below generalize the concept of language-conditioned cognitive control into three major architectural advancements that can be applied to any high-stakes AI system (e.g., automated financial analysis, molecular design, or autonomous vehicle path planning).

The paper introduces a module where all numerical claims are derived from an executable Python package (Physics Tools) rather than model-generated arithmetic. This is a critical fix for the fundamental weakness of LLMs: hallucination in quantitative reasoning.

Improvement: Implement Domain-Specific Deterministic Execution Layers (DEL), replacing the physics tools with specialized API wrappers and solvers relevant to the target domain (e.g, a chemical simulation engine, a financial risk model, or a CAD solver).

What the Improved System Can Do:

• Eliminate Epistemic Uncertainty: The AI system will never guess a result; it will calculate it. If asked to optimize the battery density for this material, the system calls a specific chemistry simulation tool, and the output is guaranteed to be physically valid.

• Guarantee Constraint Adherence: The DEL enforces hard constraints (e.g, K < N in radar; mass conservation in chemistry) that the LLM might fail to recall or apply.

The paper uses a highly structured System Prompt to encode domain knowledge and priors (e.g., If jammer to use MVDR). This transforms the LLM from a mere pattern matcher into an expert agent constrained by domain logic.

The paper’s SLMAgent.run loop includes a specific nudging mechanism to recover from premature conversation termination when the LLM fails to provide a final answer. This addresses workflow fragility in long, complex reasoning tasks.

Feature Original Cognitive Radar System Improved General AI System

:---:---:---

Logic Source (Why did it choose this?) Hardcoded rules in the system prompt. Dynamic, context-aware constraints (DKBI).

Output Validity (Is the math correct?) Verified by deterministic, executable code (DEL). Guaranteed correctness; no hallucination of results.

Execution Flow (Did it finish?) Handled by a nudging recursive loop. Self-corrective, state-aware pipeline that handles failure modes gracefully (RSAE).