Nature Medicine那篇论文把模型考倒了，但作者直接把考卷开源了。后来GPT-5.5 Pro重新考，分数涨了10%！
Nature Medicine论文对GPT-5、Gemini 2.5 Pro等前沿模型进行了医学视觉推理压力测试，通过移除图像、替换图像、调整问题格式来检验模型是否真正理解。结果显示，在标准基准上表现优异的模型在条件变化时出现推理不稳、幻觉视觉细节等问题。作者将测试框架开源，随后有人用GPT-5.5 Pro重新运行，得分从之前最好的67-70%提升至79%，约10个百分点的进步。这解决了论文发表时模型已过时的问题，使评估能随模型更新而保持时效性。
Open-sourcing test suites for peer-reviewed papers…
Open-sourcing test suites for peer-reviewed papers is the cleanest fix for the “you’re testing old models” problem.
A Nature Medicine paper put frontier models through medical visual reasoning - X-rays and scans, with questions that should force real understanding. Then they stress-tested them: remove the image, swap it, tweak the question format. The point was simple - are these models genuinely seeing what’s there, or just pattern-matching and taking shortcuts?
The result was brutal. Even GPT-5 and Gemini 2.5 Pro buckled when conditions changed. They looked strong on standard benchmarks, then fell apart when things got messy - shaky reasoning, hallucinated visual details, and poor judgement about when to say “I don’t know”.
Here’s the important bit: the authors open-sourced the framework. Someone promptly reran it on newer models. GPT-5.5 Pro hit 79% on the base test, up from roughly 67 - 70% for the previous best - about a 10-point jump in just a few months.
That’s why this matters. By the time a paper lands, the models have already moved on. Open tests keep the evidence current as systems improve - not a six-month-old snapshot that’s stale on arrival.
We need more initiatives like this: