Benchmarks & research
Traffic-Qwen Decision adapts Qwen3.5-0.8B into a non-generative traffic signal controller using a Clef-style decision head and 4-bit QLoRA.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Training uses supervised imitation from a single Jinan scenario to teach signal-phase and green-duration choices. There is no reinforcement learning, and the model is not deployment-ready. Imbalanced duration labels often leave it choosing only 20 or 40 seconds.
Researchers can use the experiment in city traffic simulations to replace rule-based lights with an AI that reads the waiting cars. It chooses which light turns green next and how long it stays green.
This is an early test that works only inside the CityFlow simulator. It is neither safe nor ready for use on real streets.