Large Language Models Simply Achieve Explainable and Training-Free One-shot HRRP ATR, IEEE_SPL_2025

<div align="right"> <a href="#english">English</a> | <a href="#chinese">中文</a> </div>

English

This repository provides the codes for HRRPLLM, IEEE_SPL_2025.

Diverging from conventional methods requiring extensive task-specific training or fine-tuning, our approach converts one-dimensional HRRP signals into textual scattering center representations. These are then directly processed by an LLM, which performs target recognition via few-shot in-context learning, effectively leveraging its vast pre-existing knowledge without any parameter update. As the first work to utilize general-purpose LLMs directly to HRRP target recognition, our simple but effective approach generates competitive results among current HRRP ATR baselines. This opens new avenues for the domain of few-shot radar target recognition.

Platform :pushpin:

We wrote, ran and tested our scripts on PyCharm IDE in a Conda environment, which we recommend for reproduction. We also recommend running the code on Linux (Our testing was done on Ubuntu 20.04).

Dependencies :wrench:

You actually don't need PyTorch, Tensorflow and etc. because our HRRPLLM is training-free and based on API calling.

Set up the environment with the requirements.txt.

A DEMO Toy Example

An interactive, client-side demonstration of the HRRPLLM prompt structure and simulated reasoning process is available:

This demo allows you to:

View the fixed contextual information and few-shot exemplars provided to the LLM.
Input or modify the scattering center data for a test sample.
See a simulated LLM prediction and a generated rationale based on simple heuristics.

Important: This demo runs entirely in your browser using JavaScript. It does not make actual calls to any Large Language Model API. The "LLM" responses are simulated based on a simplified comparison of the input scattering centers to the predefined prototypes. It serves to illustrate the concept and the type of information an LLM would process.

How to Run the DEMO Locally

Clone this repository.
Open the index.html file in your web browser.

中文

本仓库提供了 HRRPLLM 的代码，该工作已发表于 IEEE SPL 2025。

与传统方法需要大量任务特定的训练或微调不同，我们的方法将一维 HRRP 信号转换为文本散射中心表示。然后由 LLM 直接处理这些表示，通过少样本上下文学习进行目标识别，有效利用其庞大的预存知识而无需任何参数更新。作为首个将通用 LLM 直接应用于 HRRP 目标识别的工作，我们简单而有效的方法在当前 HRRP ATR 基线方法中取得了有竞争力的结果。这为少样本雷达目标识别领域开辟了新的途径。

平台 :pushpin:

我们在 PyCharm IDE 的 Conda 环境中编写、运行和测试了脚本，我们推荐您使用此环境进行复现。我们还建议在 Linux 上运行代码（我们的测试在 Ubuntu 20.04 上完成）。

依赖 :wrench:

实际上您不需要 PyTorch、Tensorflow 等，因为我们的 HRRPLLM 是无训练的，基于 API 调用。

使用 requirements.txt 设置环境。

DEMO 玩具示例

我们提供了一个交互式的客户端演示，展示 HRRPLLM 的提示结构和模拟推理过程：

该演示允许您：

查看提供给 LLM 的固定上下文信息和少样本示例。
输入或修改测试样本的散射中心数据。
查看基于简单启发式的模拟 LLM 预测和生成的推理依据。

重要提示： 此演示完全在您的浏览器中使用 JavaScript 运行。它不会实际调用任何大型语言模型 API。"LLM" 响应是基于输入散射中心与预定义原型的简化比较进行模拟的。它用于说明概念和 LLM 将处理的信息类型。

如何在本地运行 DEMO

克隆本仓库。
在您的网页浏览器中打开 index.html 文件。

Citation / 引用

If you find our work useful in your research, please consider citing:

如果您在研究中觉得我们的工作有用，请考虑引用：

@ARTICLE{11122886,
  author={Chen, Lingfeng and Hu, Panhe and Pan, Zhiliang and Liu, Qi and Zhang, Shuanghui and Liu, Zhen},
  journal={IEEE Signal Processing Letters}, 
  title={Large Language Models Can Achieve Explainable and Training-Free One-Shot HRRP ATR}, 
  year={2025},
  volume={32},
  number={},
  pages={3395-3399},
  keywords={Indexes;Target recognition;Scattering;Radar;Training;Large language models;Frequency-domain analysis;Data mining;Frequency modulation;Transforms;High-resolution range profile;automatic target recognition;large language models;in-context learning},
  doi={10.1109/LSP.2025.3598220}}

Acknowledgements :small_red_triangle:

This project is released under :page_facing_up: the MIT license.

本项目基于 :page_facing_up: MIT 许可证发布。

HRRPLLM

Install / Use

README