September 18, 2026
Chinese AI telescope becomes a science case study: agent operates 10 telescopes
The Xingyu telescope was included in Stanford HAI's AI Index Report 2026 as a model case of AI applied to physics and astronomy. The StarWhisper system automates observations across 10 telescopes and uses LLMs for planning. By September 18, 2026, the agent had flagged 8 supernova candidates, with follow-up observations initiated for 2 of them.
Before StarWhisper was deployed, transient detection lagged behind TNS by about 1 day. After deployment, the delay for AT2024abqt, SN2024advj, and SN2025bl fell to a few hours. AT2025pk became the first transient discovered by the system.
Planning changed. In a controlled comparison, manually creating a plan took about 1.5 hours. With StarWhisper, a plan was assembled in under 1 minute, and the authors reported better target coverage and zero conflicts.
The agent runs the cycle itself. After receiving a researcher's request, StarWhisper builds a plan, calls telescope-control modules, delivers a result, and revises subsequent plans based on execution outcomes and expert feedback. It accounts for the scientific objective, equipment status, and observing conditions. For testing, the team built a simulator for equipment, sensors, environmental changes, and observation scenarios.
The published version of the system connects to Sitian Pathfinder and a Sitian prototype, while the NGSS network includes 10 amateur telescopes. The authors released the source code, prompts, and tools in the NGSS directory on GitHub. They published a user guide for the scientific workflow in Coze.
Open access. The current repository contains 17 skills. On Windows, run `git clone https://github.com/Yu-Yang-Li/StarWhisper.git`, `cd StarWhisper`, then `powershell -File .\skills\install.ps1`. On Linux and macOS, run `./skills/install.sh` after the first two commands.
StarWhisper is already available as an open scientific workflow with source code, prompts, tools, and 17 skills.
Source
