X-Reasoner: Towards Generalizable Reasoning Across Modalities and Domains

1 Introduction

• Open-source reasoning research has focused mainly on text-only training and math or general-domain evaluation, leaving transfer beyond text and general domains uncertain.Source: §1 · p. 2
• Existing multimodal approaches often rely on task- or domain-specific multimodal datasets, which the authors say limits generalisability.Source: §1 · p. 2
• The paper asks whether general-domain text post-training can support cross-modal and cross-domain reasoning, citing compute efficiency and abundant, verifiable text data as practical advantages.Source: §1 · p. 2
• General-domain text-only post-training transfers reasoning to multimodal and specialized tasks, with X-REASONER and X-REASONER-MED outperforming prior systems across general and medical benchmarks.X-REASONER-MED is additionally trained on medical text and sets new records on numerous textual and multimodal medical tasks.Source: §1 · p. 2
• The study finds that combining SFT and RL yields the strongest gains, math-based data serves as a strong anchor, and domain-specific data remains important for specialization.Source: §1 · p. 2

2.3 Putting it all together: X-REASONER

• X-REASONER combines general-domain text-only long-CoT SFT with subsequent math-text RLVR, starting from an instruction-tuned VLM.The SFT stage uses OpenThoughts-114k; the RLVR stage uses Orz-math-57k.Source: Figure 2 · p. 3, §2.3 · p. 6

2.1 RQ1: Will Text-only SFT Improve Generalizable Reasoning?

• CoT SFT improves transfer across domains and modalities, outperforming non-CoT SFT and CoT prompting alone.SFT with general-domain OpenThoughts data also yields consistent gains across cross-domain, cross-modality, and combined settings.Source: §2.1 · pp. 3–4, Table 1 · p. 4
• Long-CoT SFT sometimes produces endless thinking, and forced-exiting reduces it while improving final task accuracy.Endless thinking occurs in approximately 17% of generations on MedQA and MMLU-Pro-Health; forced-exiting appends a stop token at a length threshold.Source: §2.1 · p. 4, Table 2 · p. 4

2.2 RQ2: Will Mathematical Text-only RL Improve Generalizable Reasoning?

• Math-trained RL supports gains on out-of-domain medical tasks, whereas MedQA RL yields the best medical performance.Math-trained models generate longer responses across benchmarks, while responses to math tasks are longer than those to medical tasks.Source: §2.2 · p. 5, Table 3 · p. 5
• SFT + RL achieves the best results across general-domain, cross-domain, cross-modality, and combined settings.Source: Table 4 · p. 6, §2.2 · p. 6

3.1 X-REASONER’s Cross-Modality Generalization

• Improvements on text-only benchmarks transfer to multimodal tasks across average, majority-vote, and pass@5 accuracy.Figure 3 compares X-REASONER with the baseline on text-only and multimodal benchmarks.Source: §3.1 · p. 7, Figure 3 · p. 7
• X-REASONER achieves state of the art on MMMU, MMMU-Pro, and MathVista, while remaining competitive on other multimodal benchmarks.The comparison is against state-of-the-art multimodal reasoning models; Table 5 reports greedy-decoding results.Source: §3.1 · p. 7, Table 5 · p. 8
• Qualitative examples show planning, verification, error correction, and step-by-step integration of visual information in X-REASONER’s reasoning.One example reviews all bar values in a plot, avoiding oversight errors described for the baseline.Source: §3.1 · pp. 7–8
• X-REASONER retains its advantage after text-solvable examples are removed, supporting genuine use of visual context.The ablation masks visual inputs to identify examples consistently solved through text alone before comparing performance.Source: §3.1 · p. 8, Table 6 · p. 8

3.2 X-REASONER in Medicine: A Cross-Modality and Cross-Domain Study

• X-REASONER surpasses baselines on textual and multimodal medical tasks, and matches or exceeds prior state of the art on MedXpertQA-MM and MMMU-Health.The prior state-of-the-art models on those multimodal tasks were explicitly trained on multimodal medical data.Source: §3.2 · p. 9, Figure 4 · p. 9
• Continued training with medical text produces X-REASONER-MED, which improves average and majority-vote accuracy over X-REASONER and sets new state of the art on text-only and multimodal medical tasks.X-REASONER sometimes has higher pass@n scores; the paper suggests X-REASONER-MED may have less room for further gains after targeted fine-tuning.Source: §3.2 · pp. 9–10

• Multimodal reinforcement-learning studies refine VLM reasoning, with some using text and multimodal data or alternating SFT and RL on multimodal inputs.The cited examples include VFT, MM-Eureka, Vision-R1, R1-V, and approaches by Zhou et al., Peng et al., and Deng et al.Source: §4 · p. 10
• For supervised fine-tuning, prior methods distill reasoning from captions or VLM-generated responses, while another approach links a text reasoning model to a vision encoder for visual grounding.Examples of caption- or response-based methods include MaMMOTH, Vision-R1, R1-OneVision, and Llava-CoT.Source: §4 · p. 10
• Medical reasoning research first focused on text, with later work exploring reinforcement-learning training on multimodal medical datasets.The passage cites Zhang et al., Huang et al., and Chen et al. for text-focused work, and Pan et al. and Lai et al. for multimodal RL.Source: §4 · p. 10

• Compute limits and the lack of recent open-source base VLMs constrained empirical validation of claims favoring pretrained base checkpoints over instruction-tuned models.Source: §5 · p. 10
• The study is limited to 7B models from the Qwen-VL series and does not evaluate other publicly available VLM backbones.Source: §5 · p. 10
• Evaluation focuses on mathematical and multiple-choice questions in general and medical domains, leaving open-ended generation, interactive dialogue, and instruction-following untested.Source: §5 · p. 10

6 Conclusion

• X-REASONER’s text-only SFT+RL generalizes across modalities and domains; X-REASONER-MED, trained on medical text, achieves new state of the art on medical benchmarks.The authors leave continued in-domain multimodal SFT/RL for future work.Source: §6 · p. 11

A GRPO

• GRPO estimates each response’s token-level advantage relative to group rewards and uses a clipped surrogate objective with a KL-divergence penalty.For each query, the old policy samples a group of outputs; the clip ratio and KL coefficient are hyperparameters.Source: A GRPO · p. 16

B.1 Supervised Fine-Tuning

• SFT uses AdamW at a 1 × 10−5 learning rate, trains each dataset for four epochs, and truncates sequences to 4096 tokens.Training uses batch size 4 per device across 32 GPUs.Source: §B.1 · p. 16

B.2 Reinforcement Learning

• RL samples eight responses per example, with a 4096-token maximum and temperature 1.0; actor updates use a global batch size of 128.Training uses KL coefficient 1.0 × 10−2 and clip ratio 0.2, and stops when training-set reward converges.Source: §B.2 · p. 16

C Prompt Templates

• The training prompt asks the model to place its thoughts inside <think> tags and its final answer inside <answer> tags.The prompt supplies the question after these instructions.Source: C Prompt Templates · p. 17
• For evaluation multiple-choice questions, the prompt requests reasoning followed by just one listed option inside <answer> tags.The template includes the question and options.Source: C Prompt Templates · p. 17
• For mathematical evaluation, the prompt specifies that reasoning and the final answer go inside <think> and <answer> tags, respectively.The final answer is formatted as an option letter or number, depending on the question.Source: C Prompt Templates · p. 17

D Evaluation Tasks

• Evaluation spans general-domain text-only and multimodal tasks, plus specialized-domain text-only and multimodal tasks.The paper focuses on medicine as its representative specialized domain, citing its practical importance and intensive research activity.Source: D Evaluation Tasks · pp. 17–18

E Previous SOTA Results

• Table 8 presents previous state-of-the-art results, with an asterisk marking results replicated by the authors.The supplied caption does not provide the table’s values or identify a winner.Source: Table 8 · p. 18

F Comparison of RL and SFT on MedQA data

• Pure RL performs worse than SFT, while SFT + RL performs best on MedQA data.Table 9 covers in-distribution text, out-of-distribution text, and cross-modality settings.Source: F Comparison of RL and SFT on MedQA data · p. 18, Table 9 · p. 18

G Training Dynamics

• Figure 5 compares RL training dynamics for X-REASONER on Orz and X-REASONER-MED on MedQA.The plotted measures include training reward, policy-gradient loss, mean advantage, response length, clip ratio, and validation accuracy.Source: G Training Dynamics · p. 18, Figure 5 · p. 19
• Training signals include reward, policy-gradient loss, and mean advantage for both models.The passage also identifies response clip ratio as a training measure.Source: G Training Dynamics · p. 18
• Response length and validation accuracy are also tracked during training.Source: G Training Dynamics · p. 18

H Qualitative Analysis

• For a text-only question, X-REASONER produces more complex reasoning traces with planning and verification than the baseline.Table 10 identifies the compared example as text-only.Source: H Qualitative Analysis · p. 19, Table 10 · p. 20
• X-REASONER integrates visual evidence with medical knowledge to identify PML in a histology image, unlike the baseline’s herpes simplex diagnosis.The medical multimodal example is identified as such in Table 12.Source: H Qualitative Analysis · pp. 20, 22, Table 12 · p. 22
• In the chart example, X-REASONER reviews bar values and corrects its reasoning, while the baseline misses half the values.Table 11 presents this as a multimodal question comparison.Source: H Qualitative Analysis · p. 20, Table 11 · p. 21
• In its bar-count reasoning, X-REASONER revises its initial assessment after checking which values fall below 40.The worked comparison reports three qualifying bars, including coins in 2016, 2017, and 2018.Source: H Qualitative Analysis · p. 21