Event ·Multi-source

GPT-5.6 Sol 在数学难题上成功，却在 ARC-AGI-3 表现差？调整 API 设置后性能三倍提升

GPT-5.6 Sol 成功解决了数学中的开放问题，但在 ARC-AGI-3 2D 拼图基准测试中表现挣扎。调查发现，问题出在 API 的 harness 不允许模型记住之前学到的内容。启用两个 API 设置后，分数提高了 3 倍，同时输出 token 减少了 6 倍。

Current read

OpenAI 发现 GPT-5.6 Sol 在 ARC-AGI-3 上表现差是因为 API 设置没开对。改两个参数分数涨三倍，token 省六倍，学起来很实用。

11 sources68° AI heatLast updated 2026/7/29 23:57:54

Evidence

11 sources

Primary sourceOpenAI

2026/7/29 23:57:54View source

Related source 1IT之家

2026/7/31 07:35:42View source

Related source 2歸藏(guizang.ai)

2026/7/29 04:11:22View source

Related source 3Greg Brockman

2026/7/29 21:31:16View source

Related source 4@OpenAIDevs

2026/7/29 21:33:09View source

Related source 5Augment Code

2026/7/29 22:31:00View source

Related source 6Sam Altman

2026/7/30 00:41:57View source

Related source 7小互

2026/7/30 04:12:00View source

Related source 8Decoder

2026/7/30 07:40:11View source

Related source 9宝玉

2026/7/28 22:10:39View source

Related source 10Simon Willison

2026/7/30 01:34:49View 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