How to Teach Large Language Models to “See” Radar?

High‑resolution range profiles (HRRPs) are like fingerprints of aircraft: they reveal scattering points where radar waves bounce off the plane. These details are crucial for automatic target recognition (ATR). The catch? With only a few radar snapshots, traditional methods tend to overfit and stumble when the viewing angle changes.

Researchers tried to harness large language models (LLMs), famous for in‑context learning, to recognize HRRPs with just one example (HRRPLLM). It works, but accuracy drops when the aspect angle varies because the model only looks at the distribution of scattering points, ignoring how they shift with perspective.

The proposed fix is the Aspect‑Distributed Prototype (ADP). The model builds “prototypes” that capture not just scattering points but also their variation across different angles. Thus, even with few samples, it can identify an aircraft from many viewpoints.

Tests on both simulated and real aircraft radar data show ADP beating current benchmarks by a large margin, proving that LLMs can become more robust to geometric changes. A step forward for smarter radars, even with limited data.

The HRRP (High‑Resolution Range Profile) is a one‑dimensional radar response representation versus distance, obtained via a Fourier transform of an FMCW or chirp pulse signal. Peaks in the spectrum correspond to scattering centers (SC), points on the target where radar energy is reflected strongly. The position and intensity of these SC depend on the target's geometry and electromagnetic properties (conductivity, permittivity).\n\nIn ATR, traditional deep‑learning models learn a fixed mapping between HRRP and target class. With few examples, aspect‑dependent variance leads to overfitting: the model memorizes specific patterns from the training set without generalizing.\n\nThe ADP approach introduces aspect‑distributed prototypes. Formally, for each class c an ensemble of prototypes \(\{p_{c,\theta}\}\) is built, each associated with a different angle \(\theta\). During inference, the LLM receives the HRRP and a prompt describing the available prototype set. Through in‑context learning, the model compares the new HRRP with the nearest prototypes in a latent space, evaluating both SC peak matching and angular variation.\n\nComputationally, ADP requires a preprocessing step to extract SC (e.g., peak‑picking algorithm) and an angle encoding (e.g., sinusoidal embedding). Prototypes can be updated online via few‑shot meta‑learning, preserving adaptability to new targets with minimal examples.\n\nExperimental results show an average 12‑15% top‑1 accuracy gain over HRRPLLM and CNN/Transformer baselines, on both simulated (EM‑sim radar) and real (X‑band radar) datasets. This demonstrates that embedding geometric (aspect) information into LLM prompts is a promising strategy to overcome data‑scarcity limitations of traditional models.