Hi @jmwang0117,

Thank you for sharing your work on HE-Drive; it’s very impressive! I noticed that the code repository currently doesn’t include the VLM-guided scoring section, specifically the component where Llama 3.2V is used to dynamically adjust scoring weights based on scene context (e.g., adapting to “aggressive” or “conservative” driving styles).

From reading the paper, I understand that this module uses a Vision-Question-Answering (VQA) approach with multi-round dialogues to adjust scoring weights dynamically. I’d like to ask:

1. Will the VLM adjustments be added as an interactive or periodic update feature within the rule-based scoring framework?
2. Could you share any insights on integrating Llama 3.2V with the rule-based scorer, particularly to balance interpretability with adaptivity in different driving scenarios?

Looking forward to any updates you can share on this feature :)