SaaS Metrics
GitHub Issue
Question about RL for Unified Models
HunyuanImage3 | Unified AR + diffusion | Text → Image | ✅
-- | -- | -- | --
Bagel | Unified AR + diffusion | Text → Image | ✅
Thanks for the great job!
Can this RL framework support these two models on Image+Text → Text, that is, the image understanding task?
View Raw Thread
Developer & User Discourse
Wang-Xiaodong1899 • Jun 11, 2026
> Thanks for your interest!
>
> Yes, the framework already supports the image understanding task (Image + Text → Text). We currently ship Qwen2.5-VL with full RL support — see the `unirl/models/qwen_vl/` pipeline and the example recipe `examples/ar/qwen_vl_grpo_geo3k_mc_4x8.yaml` (GRPO on Geometry3K, with SGLang and LoRA variants as well; we have tested the training and the reward curve increases well).
Thanks a lot. Qwen2.5-VL is ok. How about the RL for image understanding task on Bagel?
>
> Yes, the framework already supports the image understanding task (Image + Text → Text). We currently ship Qwen2.5-VL with full RL support — see the `unirl/models/qwen_vl/` pipeline and the example recipe `examples/ar/qwen_vl_grpo_geo3k_mc_4x8.yaml` (GRPO on Geometry3K, with SGLang and LoRA variants as well; we have tested the training and the reward curve increases well).
Thanks a lot. Qwen2.5-VL is ok. How about the RL for image understanding task on Bagel?
yanghlll • Jun 11, 2026
> Thanks for your interest!
>
> Yes, the framework already supports the image understanding task (Image + Text → Text). We currently ship Qwen2.5-VL with full RL support — see the `unirl/models/qwen_vl/` pipeline and the example recipe `examples/ar/qwen_vl_grpo_geo3k_mc_4x8.yaml` (GRPO on Geometry3K, with SGLang and LoRA variants as well; we have tested the training and the reward curve increases well).
Hi， great work. Can this RL framework support think mode, like UniGRPO.
>
> Yes, the framework already supports the image understanding task (Image + Text → Text). We currently ship Qwen2.5-VL with full RL support — see the `unirl/models/qwen_vl/` pipeline and the example recipe `examples/ar/qwen_vl_grpo_geo3k_mc_4x8.yaml` (GRPO on Geometry3K, with SGLang and LoRA variants as well; we have tested the training and the reward curve increases well).
Hi， great work. Can this RL framework support think mode, like UniGRPO.
Yes, the framework already supports the image understanding task (Image + Text → Text). We currently ship Qwen2.5-VL with full RL support — see the `unirl/models/qwen_vl/` pipeline and the example recipe `examples/ar/qwen_vl_grpo_geo3k_mc_4x8.yaml` (GRPO on Geometry3K, with SGLang and LoRA variants as well; we have tested the training and the reward curve increases well).