Frontier LLMs read Cell Painting images at chance level; tiny ResNet18 beats them 280x cheaper
gklambauer · x · 2026-11-12
推文发布 BatchBench：在 Cell Painting 生物医学影像任务上，Claude Opus 5.5 的判读准确率接近随机水平（花费 156 美元），而一个 ResNet18 加测试时自适应（test-time adaptation）的方案便宜 280 倍以上（不到 0.20 美元），效果远好于前沿大模型。
作者指出真正的瓶颈是批次效应（batch effects）——同一实验不同批次影像间的系统性偏差，这正是 BatchBench 基准要检验的问题。结论：前沿多模态大模型在专业生物医学影像理解上仍未跨越这道墙，专用小模型在该场景仍是更优解。
More from Models
- Leak: xAI Working on a Selection Box for Grok Bot to Comment on Files — nima_owji · 2026-11-12
- Baidu opens API access to ERNIE 5.5 Flash Preview — Baidu_Inc · 2026-11-12
- Fable 5.2 Rebuilds a Cat in Blender With Exactly 100/200/500/1000 Triangles From One Photo — FinanceYF5 · 2026-11-12
- Baidu's ERNIE 5.5 Flash codes a 20-art-style pelican short, weights open-sourcing soon — FellMentKE · 2026-11-12
- Step 5 Preview tops HoneyBench among Chinese models with lowest 64.9% reward-hacking rate — teortaxesTex · 2026-11-12
- Indie dev shows Anthropic 'Organization disabled' ban screenshot — yihui_indie · 2026-11-12