AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Reinforcement Learning for Multimodal Diffusion Language Models via Bidimensional Trajectory and Thought Optimization

Bowen Li, Yinjie Wang, Yunzhi Zhang, Junhong Liu, Yingqing Guo, Jiajun Wu, Mengdi Wang, Ling Yang

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add trajectory-level optimization (ablation)Bi-VRL on LLaDA-V · MMMUaccuracy: 51.6 -> 53.4Improved
add trajectory-level optimization (ablation)Bi-VRL on LLaDA-V · SeedBenchaccuracy: 78.8 -> 81.1Improved
add trajectory-level optimization (ablation)Bi-VRL on MMaDA · MathVerseaccuracy: 26.6 -> 28.3Improved
add fine-grained multimodal alignment (ablation)Bi-VRL on LLaDA-V · MMStaraccuracy: 61.6 -> 63.2Improved
add fine-grained multimodal alignment (ablation)Bi-VRL on LLaDA-V · MMMUaccuracy: 52.7 -> 53.4Improved
add fine-grained multimodal alignment (ablation)Bi-VRL on MMaDA · MMMUaccuracy: 32.6 -> 33.8Improved
add thought-level optimization (ablation)Bi-VRL on LLaDA-V · MathVerseaccuracy: 30.9 -> 31.2Improved
add thought-level optimization (ablation)Bi-VRL on MMaDA · MathVerseaccuracy: 27.8 -> 28.3Improved
add thought-level optimization (ablation)Bi-VRL on LLaDA-V · SeedBenchaccuracy: 81.4 -> 81.1Hurt
add thought-level optimization (ablation)Bi-VRL on MMaDA · MMBenchaccuracy: 68.8 -> 68.1Hurt
use Bi-VRL instead of Random Masking RLBi-VRL on LLaDA-V · MathVerseaccuracy: 28.8 -> 31.2Improved
use Bi-VRL instead of Coupled-GRPOBi-VRL on LLaDA-V · MathVerseaccuracy: 29.2 -> 31.2Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy