K-Dense 发布 K-Bench 预印本：9 个模型在 178 项真实科研任务上的基准测试

摘要

K-Dense 的 K-Bench 预印本已上线 arXiv，在 178 项真实科研任务上测试了 9 个模型，均为开箱即用、不带智能体框架或技能。论文梳理了模型最常见的失败模式及其对研究的影响。K-Dense Web 新增 Amazon S3 与 Google Cloud Storage 数据连接器，可直接把自有存储桶中的文件导入研究会话。

摘要由模型根据原文撰写，以原文为准。

// August 27, 2026

Which model performs best on scientific applications? Our K-Bench preprint is now live on arXiv with answers, and the recording of this week’s live BYOK walk-through is up on YouTube.

>Benchmarking Model Performance on Real Scientific Requests

This week our preprint for K-Bench was released on arXiv. We tested 9 models (out-of-the-box, no agentic harness or skills) on 178 real tasks and measured their strengths and weaknesses. Check out the full paper to read about the most common failure modes, and what it means for your research.