`max_tokens` and `temperature` not used in payload

[dandansamax]

Hi, and congratulations on having the paper accepted to NeurIPS 2025 — great work and exciting results!

I'm trying to reproduce your work. While reading the code, I noticed that call_openai_model_with_tools() takes max_tokens and temperature as arguments, but they are not forwarded into the payload.

DeepVideoDiscovery/dvd/utils.py

Line 82 in 64414b2

max_tokens: int = 4096,

As a result, the model uses defaults. Did the experiments in the paper rely on default decoding parameters?

Another question about the released caption files. I downloaded the LVBench and it works great, but for VideoMME caption, I noticed that the matrix field is not provided so I could not load them into nanovectorDB.

Thanks for open-sourcing the code! I apperaciate it if you can help with my problems.