[Feature] Support LLaVA-OneVision-1.5 RL Training

[randydl]

Hi there,

First of all, thank you for the excellent project!

LLaVA-OneVision-1.5 is a multimodal model that combines Rice ViT with Qwen3 LLM. The project has open-sourced all training code, datasets, and model weights, making it a great resource. Its reinforcement learning (RL) code is also built upon AReaL (see here), but it uses an early version of AReaL.

Would it be possible to integrate LLaVA-OneVision-1.5 into the official AReaL implementation? That would greatly benefit the community by providing an up-to-date and unified RL training pipeline for this model.

Thanks for your consideration!

[garrett4wade]

Hi @randydl , we really appreciate your adoptation of AReaL! Currently we are making every efforts to fix VLM/Omini model bugs. We are more than happy to support LLaVA-OneVision-1.5 in this codebase. Hopefully it will be done in the near future. Please stay tuned.