Event ·Official first-hand

千图假说：代码推理中的任务条件关系物化

DeepSeek-V4-Flash模型在SWE-bench Verified基准测试中，通过两层索引分离全局路由与局部实体焦点，实现了零预构建实体关系边条件下的95.6%成功率。研究提出了一种仅实体的外部接口，在推理过程中实现任务条件关系物化。基础、单层和双层条件分别达到92.1%、94.2%和95.6%的成功率。

Current read

DeepSeek团队提出千图假说，解决大型代码库推理问题，无需预构建关系图就能达到95.6%成功率。

5 sources73° AI heatLast updated 2026/8/27 04:33:48

Evidence

5 sources

Primary sourcearXiv: DeepSeek

2026/8/27 04:33:48View source

Related source 1AI Will

2026/8/26 08:51:35View source

Related source 2Geek

2026/8/26 14:14:10View source

Related source 3Decoder

2026/8/26 14:40:09View source

Related source 4elvis

2026/8/27 19:24:02View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story