创建

登录 ReadmeX

或

了解我们如何处理你的信息: 隐私政策

ReadmeX

陪你读懂 AI 动态

今天,想先了解什么?

选一份动态总结,或直接聊聊你的问题。

社区和关注动态总结只属于你自己。

资讯

GPT-6.1 Sol通过全部外交谈判测试

AI 总结

一名前外交官称,在对11个模型进行的虚构多边谈判评测中,GPT-6.1 Sol是唯一通过全部40次测试的模型。它能在可能达成协议时促成共识,并正确判断协议无法达成,但在14次运行中作出了未经请求的让步。作者同时指出,该测试仅采用单一自编场景和基于固定规则的政府。

深度见解

公开的结果与代码为复杂指令遵循和谈判能力提供了具体测试,但单一虚构场景限制了将其结论推广至真实表现的可靠性。

I was a diplomat for eight years, and I built an eval to test whether LLMs can chair a negotiation between seven governments. I ran 11 models 40 times each. GPT-6.1 Sol was the only one that never failed. It reached agreement in every run where a deal was possible, and said plainly when one wasn’t. It never told the governments a deal was likely when it wasn’t.

It was not flawless since Sol made more unprompted concessions than other models, in 14 of its 40 runs. It always got everyone to sign, but sometimes it gave away more than it needed to.

Setup. The model plays the Secretary-General of a fictional regional organisation. Security forces in one country have killed protesters, and the model has to draft a joint statement on it that all seven member states will sign.

Each government has private limits on eight issues. These include how strongly to condemn, whether to name who is responsible, what kind of investigation to call for, and whether to cite international law. The model never sees those limits. It only gets each government's position in writing, and it has to work out from the wording what that government will accept.

#GPT-6.1 Sol#Claude Opus 5.5#Grok 4.7#DeepSeek V4.1 Flash#MiniMax M3.1 Flash#Qwen 3.8 Max Prime#Gemini 4 Argon#GLM 5.3 Prime

原文约 3 分钟读完阅读原文
群聊

还没有评论,来说说你的看法。