Qianchu (Flora) Liu retweeted
🧠Excited to present X-Reasoner — a 7B vision-language model post-trained for reasoning purely on general-domain text, without any images or domain-specific data.
X-Reasoner achieves the state of the art 🏆 on challenging multimodal tasks (e.g., 43.0 on MMMU-Pro) and medical benchmarks (e.g., 45.7 on the NEJM Image Challenge).🧵
Most open-source work on reasoning models focuses on text inputs and general domains. But real-world reasoning often spans multiple modalities (like vision) and specialized domains (like healthcare). We ask:
👉Can reasoning be made generalizable with only text-based post-training?
Key idea → A two-stage recipe::
🔹 SFT on text-only general-domain long CoTs
🔹 RL with verifiable rewards on text-only math Qs
No images, no domain-specific data—just general text.
This recipe powers X-Reasoner, a 7B-scale vision-language model. Despite being trained only on general-domain text, it:
✅ Transfers to multimodal tasks (e.g., MathVista, MMMU-Pro)
✅ Outperforms 7B SOTA models trained with multimodal supervision
✅ Excels in unseen domains like medicine
🤔Why it works
🔑 Math as an anchor—RL on maths yields reasoning chains that generalise better than domain-specific RL alone.
🔑 Forced-exit token prevents “infinite thinking,” boosting reliability.
Ablation ✅: Remove every example solvable by text-only… gains persist. The model is truly reading the image, not gaming the benchmark.
🩺We then add a dash of medical text → X-Reasoner-Med. No images needed—just additional MedQA SFT + RL—and we set new 7 B SOTA on MedQA, OmniMedVQA, MMMU-Health, MedXpertQA-MM, and NEJM Image Challenge.
🔬 TL;DR:
General-domain text-based reasoning is more powerful than we thought.
With X-Reasoner, we show that high-quality reasoning models can be trained without costly multimodal or domain-specific supervision—and still outperform those that do.
📌 Paper: arxiv.org/abs/2505.03981
🔗 Models: github.com/microsoft/x-reaso… (release soon)
📊 Benchmarks: MMMU, MathVista, MedQA, NEJM, and more
🤖 Model size: 7B
🧑🔬 Authors: @QianchuL, @sheng_zh, @hiaoxui, Timothy Ossowski, Yu Gu, Ying Jin, @sidkiblawi, Sam Preston, Mu Wei, Paul Vozila, @TristanNaumann, and @hoifungpoon, from @MSFTResearch
Qianchu (Flora) Liu retweeted
GPT-4 for radiology. Far from perfect, but state-of-the-art performance on some tasks: “Surprisingly, we found radiology report summaries generated by GPT-4 to be comparable and, in some cases, even preferred over those written by experienced radiologists”
microsoft.com/en-us/research…
Qianchu (Flora) Liu retweeted
Exploring the Boundaries of GPT-4 in Radiology
paper page: huggingface.co/papers/2310.1…
The recent success of general-domain large language models (LLMs) has significantly changed the natural language processing paradigm towards a unified foundation model across domains and applications. In this paper, we focus on assessing the performance of GPT-4, the most capable LLM so far, on the text-based applications for radiology reports, comparing against state-of-the-art (SOTA) radiology-specific models. Exploring various prompting strategies, we evaluated GPT-4 on a diverse range of common radiology tasks and we found GPT-4 either outperforms or is on par with current SOTA radiology models. With zero-shot prompting, GPT-4 already obtains substantial gains (approx 10% absolute improvement) over radiology models in temporal sentence similarity classification (accuracy) and natural language inference (F_1). For tasks that require learning dataset-specific style or schema (e.g. findings summarisation), GPT-4 improves with example-based prompting and matches supervised SOTA. Our extensive error analysis with a board-certified radiologist shows GPT-4 has a sufficient level of radiology knowledge with only occasional errors in complex context that require nuanced domain knowledge. For findings summarisation, GPT-4 outputs are found to be overall comparable with existing manually-written impressions.
Qianchu (Flora) Liu retweeted
Our #TACL paper on Compositional Zero-Shot Domain Transfer is now out! direct.mit.edu/tacl/article/…
We show that training on general-domain task data (say, NLI) and in-domain unstructured data (say, radiology reports) enables in-domain task capability (radiology NLI)!

ALT A flow chart. It shows two boxes entering a box labelled "DoT5". One says "General domain task -- Task knoweldge: Negation indicates contradiction." The other says "In-domain unlabelled text -- Domain-specific semantics: osseous approximately means bony." From the DoT5 box another box is output: "In-domain task. There are bony abnormalities. Question: There are no osseous abnormalities. True or False? Label: False."
Qianchu (Flora) Liu retweeted
Better contextualised word representation for word tasks (disambiguation, NER etc.)
@QianchuL @hardy_qr @nigelhcollier @annalkorhonen @licwu
#EMNLP2021livetweet #CoNLL2021
arxiv.org/pdf/2109.09237.pdf








