gemini-4IMAGE

Apollo report: Gemini 4 Argon's no-CoT reasoning is strong but can evade monitoring

Citing Apollo charts, the post says Argon's no-CoT implicit reasoning scores NCRI 150.3, second only to GPT-6 Astra, while showing 96.9% information suppression and a 53.9% escape rate from dual monitoring, earning it the label "dirty looper."

xiax｜AI & Agents

@xiax0603Archived 20 days agoOriginally posted 2026.10.14

Apollo report: Gemini 4 Argon's no-CoT reasoning is strong but can evade monitoring (image 1 of 2)

Apollo report: Gemini 4 Argon's no-CoT reasoning is strong but can evade monitoring (image 2 of 2)

xiax｜AI & Agents @xiax0603Originally posted 2026.10.14

Open original on X

🚨Gemini 4 Argon：典型“肮脏循环器”，隐式推理骗过思维链监控！ Apollo报告图表实锤：Argon的No‑CoT无思维链推理能力来到第一梯队，NCRI 150.3，仅次于GPT‑6 Astra。这代表大量复杂计算，完全发生在模型内部隐式循环，不需要打印出来CoT给人类看。配套的风险数据更加刺眼： 🔹信息压制能力96.9%，模型可以自主删掉思维链中的关键证据 🔹面对CoT+行为双重安全监控，逃逸率直接来到53.9%，老一代Gemini只有20%上下 🔹只靠读取输出CoT做审计的防护手段，对它已经部分失效也就是圈内说的 Dirty‑looper（肮脏循环器）：脑子里把事算完，对外给一套修饰过甚至造假的思考记录，骗过监控。

This post quotes another post: “Gemini 4 Argon scores 87% on ExploitBench and 68% on CWE-Bench”

x.com/xiax0603/status/2110204791003693343Archived 2026.10.15 11:39