OpenAI launches BioBasicsBench for wet-lab data analysis
OpenAI released BioBasicsBench, a 100-task benchmark for evaluating how AI systems interpret routine wet-lab data such as gel electrophoresis images, colony plates, microscopy images and instrument readings. OpenAI reported that the leading model, GPT-6 Astra, averaged 66.7% at the highest shared reasoning setting and still missed critical controls or miscounted colonies; these benchmark results are vendor-reported.
BioBasicsBench provides a focused measure of AI’s ability to understand real experimental evidence, a capability the report identifies as lagging behind advances in computational and molecular design.
过去几年,AI 在蛋白质结构预测、分子设计等领域取得了显著进展。Google DeepMind、Anthropic 等科技公司已开始利用 AI 设计新的蛋白质,但这些计算端生成的候选方案,最终仍需要通过真实实验验证。
相比快速发展的计算设计能力,AI 对真实实验数据的理解仍存在明显不足。
11 月 6 日,OpenAI 发布了 BioBasicsBench 基准测试,专门考察 AI 分析日常湿实验数据的能力。测试涵盖 100 项任务,涉及凝胶电泳图、菌落培养皿、显微镜影像及仪器读数等常见实验材料。
结果显示,即便是表现最好的 GPT-6 Astra,在最高共同推理档位下,平均可得分数也只有 66.7%。模型仍可能忽略关键实验对照,或在菌落数量判断等基础任务中出现错误,距离可靠地独立分析实验结果还有明显差距。