UCSD Hao AI Lab· X：UCSD Hao AI Lab (@haoailab)·· 3 小时前AI 评分32
GPT6-Astra 与 Jev 集群抗贿赂证人测试
AI 导读
在 100 个更难的案例中，被收买的证人数量超过诚实证人，GPT6-Astra 加 Jev 集群得分 93，GPT6-Astra 单独为 59，Jev 集群单独为 42，规则式领导者加 Jev 集群仅 0–1。结论是面对有组织的谎言，需要一个会追问"谁在说话"的推理型领导者。
正文
4️⃣ But when the witnesses are bribed… 💰
In 100 harder cases, paid witnesses outnumber honest ones. Here is how each team performs:
• 🏅 GPT6-Astra + Jev swarm: 93
• 🥈 GPT6-Astra alone: 59
• 🥉 Jev swarm alone: 42
• ❓ Rule-based leader + Jev swarm: 0–1
Takeaway: against organized lies, you need a reasoning leader who asks who's talking.
来源：UCSD Hao AI Lab · x.com