Context-Aware RL for Agentic and Multimodal LLMs

Open paper
Benchmark Model Rank Results
mathematical-reasoning-on-mathverseContextRL (trained from Qwen3-VL-8B)#6Accuracy: 66.4
visual-question-answering-vqa-on-blinkContextRL (trained from Qwen3-VL-8B)#14Accuracy: 66.6
vision-language-understanding-on-mmstarContextRL (trained from Qwen3-VL-8B)#25Accuracy: 74.8
multimodal-reasoning-on-mathvisionContextRL (trained from Qwen3-VL-8B)#30Accuracy: 52.0
multimodal-reasoning-on-mathvistaContextRL (trained from Qwen3-VL-8B)#32Accuracy: 79.8
multimodal-reasoning-on-mmmuContextRL (trained from Qwen3-VL-8B)#34Accuracy: 70.1
multimodal-reasoning-on-mmmu-proContextRL (trained from Qwen3-VL-8B)#34Accuracy: 57.5
code-generation-on-swe-bench-verifiedContextRL (trained from Klear-AgentForge-8B) (mini-SWE-agent)#83Accuracy: 30.2

Rank counts only results with a code link.