← 研踪 PaperTrailarXiv:2611.21799 · v2
Controllable Chains of Thought Can Reveal Less, Even Without Evasion Training
首次提交 2026-11-17 03:17
版本更新 2026-11-19 17:43
雷达观测 2026-11-20 19:58
摘要
Chain-of-thought (CoT) monitoring is a promising tool for detecting hidden goals, off-task behavior, and other failures in AI agents. One risk is that models could deliberately shape their reasoning to make it opaque to monitors. We stress-test this risk through persona-control training: supervised fine-tuning on mathematical reasoning traces rewritten in the voices of diverse personas. This induces a broader ability to control CoT, even though training contains no monitor-evasion examples. In covert side-task evaluations, traces in which a gpt-oss-20b model organism completes the hidden side task contain less monitorable evidence than the base model's. Averaged across evaluations and CoT lengths, weighted by side-task-success counts, the CoT monitorability metric g-mean falls from 0.94 for the base model to 0.68 for the model organism without a hiding instruction, and to 0.22 when explicitly instructed to hide. In intervention evaluations, explicit hiding often drives monitorability near zero. The effect extends to a Kimi-K2.6 model organism in agentic environments, though hidden objectives requiring extended reasoning or longer trajectories remain more monitorable than simple single-step side tasks. We further discover that targeted transparency prompts can substantially recover monitorability, sometimes surpassing the base model. These model organisms provide a testbed for studying monitorability failures and potential mitigations. Our code can be found on https://github.com/yongzx/cot-control-model-organism.
作者备注:22 pages, 17 figures. v2: Updated title and author notes, and added a citation in Related Work
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-19 17:43 · 最新观测版本
Controllable Chains of Thought Can Reveal Less, Even Without Evasion Training
Chain-of-thought (CoT) monitoring is a promising tool for detecting hidden goals, off-task behavior, and other failures in AI agents. One risk is that models could deliberately shape their reasoning to make it opaque to monitors. We stress-test this risk through persona-control training: supervised fine-tuning on mathematical reasoning traces rewritten in the voices of diverse personas. This induces a broader ability to control CoT, even though training contains no monitor-evasion examples. In covert side-task evaluations, traces in which a gpt-oss-20b model organism completes the hidden side task contain less monitorable evidence than the base model's. Averaged across evaluations and CoT lengths, weighted by side-task-success counts, the CoT monitorability metric g-mean falls from 0.94 for the base model to 0.68 for the model organism without a hiding instruction, and to 0.22 when explicitly instructed to hide. In intervention evaluations, explicit hiding often drives monitorability near zero. The effect extends to a Kimi-K2.6 model organism in agentic environments, though hidden objectives requiring extended reasoning or longer trajectories remain more monitorable than simple single-step side tasks. We further discover that targeted transparency prompts can substantially recover monitorability, sometimes surpassing the base model. These model organisms provide a testbed for studying monitorability failures and potential mitigations. Our code can be found on https://github.com/yongzx/cot-control-model-organism.
查看此版本解读 →查看此版本原文 ↗v1 · 2026-11-17 03:17
Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability
Chain-of-thought (CoT) monitoring is a promising tool for detecting hidden goals, off-task behavior, and other failures in AI agents. One risk is that models could deliberately shape their reasoning to make it opaque to monitors. We stress-test this risk through persona-control training: supervised fine-tuning on mathematical reasoning traces rewritten in the voices of diverse personas. This induces a broader ability to control CoT, even though training contains no monitor-evasion examples. In covert side-task evaluations, traces in which a gpt-oss-20b model organism completes the hidden side task contain less monitorable evidence than the base model's. Averaged across evaluations and CoT lengths, weighted by side-task-success counts, the CoT monitorability metric g-mean falls from 0.94 for the base model to 0.68 for the model organism without a hiding instruction, and to 0.22 when explicitly instructed to hide. In intervention evaluations, explicit hiding often drives monitorability near zero. The effect extends to a Kimi-K2.6 model organism in agentic environments, though hidden objectives requiring extended reasoning or longer trajectories remain more monitorable than simple single-step side tasks. We further discover that targeted transparency prompts can substantially recover monitorability, sometimes surpassing the base model. These model organisms provide a testbed for studying monitorability failures and potential mitigations. Our code can be found on https://github.com/yongzx/cot-control-model-organism.
查看此版本解读 →查看此版本原文 ↗