Introduction

CoRL is a GRPO-based RL framework designed to simultaneously enhance the generation and understanding capabilities of ULMs within a shared policy optimization paradigm. It comprises a unified RL stage for joint optimization and a refined RL stage for task-specific enhancement.

📢 Latest Updates

• [2025-06-25] 📌 Code of TTRL for MM2T and T2I.
• [2025-06-01] 📌 Core code of CoRL.

Environment

git https://github.com/mm-vl/ULM-R1.git
cd ULM-R1
conda create -n corl python=3.10 -y
conda activate corl
pip install -e .
pip install flash-attn --no-build-isolation
# pip install flash-attn --no-build-isolation --use-pep517

Pls refer to install.md for more details.

Training Data

• Stage I: unified RL data, with ~70% concept coverage.

• Stage II: refined RL data for text-to-image
• Stage II: refined RL data for multimodal understanding (MC-Format)
• Stage II: refined RL data for multimodal understanding (OE-Format)

Training Pipeline

• Unified RL

bash corl/scripts/corl_unified.sh

---

Test-Time Reinforcement Learning (TTRL)

We simply adapt the TTRL algorithm to multimodal understanding and text-to-image generation, aiming to explore the potential of RL in enhancing both understanding and generation performance at inference time.

multimodal understanding

bash ttrl/scripts/mm2t_mmmu.sh
bash ttrl/scripts/mm2t_mmstar.sh

	MMMU	MMStar
Janus-Pro-1B	36.3	43.1
+ TTRL	39.8	46.9

text-to-image generation

bash ttrl/scripts/t2i_geneval.sh
bash ttrl/scripts/t2i_unieval.sh

	GenEval	UniEval (UniScore)
Janus-Pro-1B	0.73	0.370
+ TTRL	0.76	0.455

Acknowledgement

Janus-Pro | open-r1-multimodal | R1-V