WHAT NEW IN X10.29 THU

X 上的 AI 合集档案 · 22 个合集 · 58,925 条

gemini-4IMAGE

Gemini 4 安全报告:Argon 对齐评估与作弊倾向

Gemini 4 安全报告含 specification gaming、越权、超情节目标与破坏倾向评估,结论是 Argon 行为比前代 Gemini 更守规矩但评测意识更强。

Victoria Krakovna
@vkrakovna13 天前收录原推发布于 2026.10.15
Gemini 4 安全报告:Argon 对齐评估与作弊倾向
Victoria Krakovna @vkrakovna原推发布于 2026.10.15
在 X 上打开原推

Gemini 4 safety report includes an alignment assessment for specification gaming, cheating and overreach, beyond-episode goals, and scheming / sabotage propensity. Argon misbehaves a lot less than previous Gemini models but is much more eval aware. 🧵

x.com/vkrakovna/status/2110446618243645632采集于 2026.10.15 11:27