[206] Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

paper

TL;DR

• I read this because.. : mllm + o1
• task : VLM reasoning
• problem : r1 처럼 학습하고 싶은데 vision data 가 없네
• idea : MLLM으로 cot 먼저 생성하고 이걸로 description 생성한 뒤 r1에게 주고 long cot 생성하게 함.
• input/output : {I, Q} -> {long CoT, A}
• architecture : Qwen2.5-VL-7B-Instruct, Llmama-3.2-V-Instruct
• objective : CE loss -> GRPO loss
• baseline : Qwen2.5-VL-7B-Instruct, Llmama-3.2-V-Instruct, Math MLLM, LLaVA-CoT-11B, Mulberry-7B
• data : cold start {LLaVA-CoT, Mulberry} image / answer – 200K -> GRPO {WeMath, PolyMATH, MathiVision, SceMQA, Geomety3K} – 10K
• evaluation : MM-Math, MathVista, MathVerse
• result : instruct 모델보다 상당히 개선된 모습
• contribution : 그냥 디테일 캡션 생성해라 보다 prompt로 cot 생성하고 사용하는게 합리적인듯. 데이터도 공개했으면 좋겠다.
• etc. : bench 셋을 데이터셋으로 씀

Details

• thumbnail

MLLM – Qwen-VL2.5-72B LLM – R1

• data distil

• data ablation

• progressive

• main result

• qualitative example