Скачать или смотреть OpenMMReasoner: Open Recipe for Multimodal LMs

In this AI Research Roundup episode, Alex discusses the paper:
'OpenMMReasoner: Pushing the Frontiers for Multimodal Reasoning with an Open and General Recipe'
OpenMMReasoner introduces a fully open, two-stage training pipeline—supervised fine-tuning followed by reinforcement learning—for large multimodal reasoning models. The authors focus on transparent, reproducible data curation, starting from public QA sources and distilling teacher-generated reasoning traces into a high-quality 583k example general reasoning set, then expanding it to 874k examples with cross-domain image and text math data. Key design choices include careful teacher model selection, preserving diverse reasoning traces without heavy filtering, and enforcing a unified step-by-step output format. The reinforcement learning stage then refines the model on a cleaned, multimodal dataset spanning multiple benchmarks to further boost reasoning performance.
Paper URL: https://arxiv.org/abs/2511.16334

#AI #MachineLearning #DeepLearning #MultimodalReasoning #LargeLanguageModels #ReinforcementLearning #ComputerVision

Resources:
GitHub: https://github.com/EvolvingLMMs-Lab/O...
Hugging Face model: https://huggingface.co/OpenMMReasoner...
Hugging Face model 2: https://huggingface.co/OpenMMReasoner...