X-Reasoner: Towards Generalizable Reasoning Across Modalities and Domains

TL;DR

This work investigates whether reasoning learned from general-domain text can generalize across modalities and domains. It introduces X-Reasoner, a 7B vision-language model post-trained with a two-stage, text-only recipe: supervised fine-tuning with distilled long chain-of-thoughts, followed by reinforcement learning with verifiable rewards on mathematical text. The results show strong cross-domain and cross-modality generalization, outperforming state-of-the-art models trained with in-domain multimodal data on general and medical benchmarks, and demonstrating further gains from domain-specific text data through X-Reasoner-Med. The findings suggest that textual supervision alone can induce robust, transferable reasoning patterns applicable to multimodal tasks and specialized domains, with practical implications for reducing data curation costs and enabling rapid deployment in fields like medicine.

Abstract

Recent proprietary models (e.g., o3) have begun to demonstrate strong multimodal reasoning capabilities. Yet, most existing open-source research concentrates on training text-only reasoning models, with evaluations limited to mainly mathematical and general-domain tasks. Therefore, it remains unclear how to effectively extend reasoning capabilities beyond text input and general domains. This paper explores a fundamental research question: Is reasoning generalizable across modalities and domains? Our findings support an affirmative answer: General-domain text-based post-training can enable such strong generalizable reasoning. Leveraging this finding, we introduce X-Reasoner, a vision-language model post-trained solely on general-domain text for generalizable reasoning, using a two-stage approach: an initial supervised fine-tuning phase with distilled long chain-of-thoughts, followed by reinforcement learning with verifiable rewards. Experiments show that X-Reasoner successfully transfers reasoning capabilities to both multimodal and out-of-domain settings, outperforming existing state-of-the-art models trained with in-domain and multimodal data across various general and medical benchmarks (Figure 1). Additionally, we find that X-Reasoner's performance in specialized domains can be further enhanced through continued training on domain-specific text-only data. Building upon this, we introduce X-Reasoner-Med, a medical-specialized variant that achieves new state of the art on numerous text-only and multimodal medical benchmarks.

Figures (4)

• Figure 2: (Left) Our recipe for generalizable reasoning: based on an instruction-tuned VLM, we first conduct SFT on general-domain text data with distilled long CoTs. This is followed by RLVR on mathematical textual questions. This resulting model, X-Reasoner, exhibits significantly enhanced reasoning capabilities across modalities and domains. (Right) Model performance on MMMU-Pro (multimodal task) steadily improves at each stage of our recipe.
• Figure 3: Comparing X-Reasoner and baseline on text-only benchmarks and multi-modal benchmarks. Despite being trained with general-domain text-only data, X-Reasoner can significantly improve multi-modal benchmarks, showing the generalization of X-Reasoner's reasoning.
• Figure 4: Comparing X-Reasoner, X-Reasoner-Med and baseline on text and multimodal medical benchmarks. X-Reasoner, trained with general-domain text-only data brings consistent improvement across medical tasks. X-Reasoner-Med, obtained by continued training of X-Reasoner on medical text data, further improves medical domain performance.
• Figure 5: RL training dynamics of X-Reasoner and X-Reasoner-Med