Code of Separate Tokens implementation

[ruwwww]

Thank you for sharing this excellent work and making the code available

While reading the paper. I noticed that Section 3.3 describes two fusion strategies:

• Merged Tokens (channel‑wise sum, Eq. 7–8)
• Separate Tokens (sequence‑wise concatenation, Eq. 9–10), which preserves modality‑specific information at the cost of increased computation.

However, in the released repository I only see the merged‑tokens implementation. Could you clarify:

• Was the separate tokens variant implemented internally for the experiments, but not released here?
• If so, do you plan to share that code, or could you provide guidance on how to adapt the current implementation to support it?