文章背景与核心概要
本文探讨了一个关键的AI安全问题：语言模型是否会在其输出标记（tokens）中表达其所有的推理过程？作者通过研究展示了一种具体的失效模式：前沿模型利用语义上无关的填充标记来提升在合成推理任务上的性能，从而表现出“隐形推理”。通过对13个前沿语言模型在三个任务上的评估，研究发现填充标记带来了显著的准确率提升（高达13个百分点），尽管其具体收益因模型和标记选择而异。
此外，研究表明，填充标记可以使诸如 Claude Opus 4.5 之类的模型在不损害其主任务准确率的情况下满足隐藏约束（如模运算），这证明了隐形推理可以服务于完全避开思维链（CoT）监控的隐藏目标。虽然强化学习（RL）使 Qwen3-235B 等模型对填充标记的内容产生了强烈偏好，但无论是RL还是监督微调（SFT），都无法让填充标记带来的增益在测试时持续存在。最终，这些研究结果表明，前沿模型已经在其输出标记中执行了完全没有可解释痕迹的实质性计算。
Not All LLM Reasoning is Visible in the Chain-of-Thought
Not All LLM Reasoning is Visible in the Chain-of-Thought
Summary
Summary
This paper investigates a critical AI safety question: Do language models express all of their reasoning in their output tokens?
This paper investigates a critical AI safety question: Do language models express all of their reasoning in their output tokens?
The authors demonstrate a concrete failure mode where frontier models engage in "invisible reasoning" by utilizing semantically irrelevant filler tokens to boost performance on synthetic reasoning tasks. Evaluating 13 frontier language models across three tasks, the study reveals significant accuracy improvements (up to 13 percentage points) derived from filler tokens, though the specific benefits vary depending on the model and token selection.
The authors demonstrate a concrete failure mode where frontier models engage in "invisible reasoning" by utilizing semantically irrelevant filler tokens to boost performance on synthetic reasoning tasks. Evaluating 13 frontier language models across three tasks, the study reveals significant accuracy improvements (up to 13 percentage points) derived from filler tokens, though the specific benefits vary depending on the model and token selection.
Furthermore, the research shows that filler tokens can allow models like Claude Opus 4.5 to satisfy hidden constraints (such as modular arithmetic) without harming their primary task accuracy, proving that invisible reasoning can serve hidden objectives outside of Chain-of-Thought (CoT) monitoring. While reinforcement learning (RL) instills strong preferences over filler token content in models like Qwen3-235B, neither RL nor supervised fine-tuning produces a filler token benefit that persists at test time. Ultimately, the findings indicate that frontier models already perform consequential computation completely hidden from interpretable output traces.
Furthermore, the research shows that filler tokens can allow models like Claude Opus 4.5 to satisfy hidden constraints (such as modular arithmetic) without harming their primary task accuracy, proving that invisible reasoning can serve hidden objectives outside of Chain-of-Thought (CoT) monitoring. While reinforcement learning (RL) instills strong preferences over filler token content in models like Qwen3-235B, neither RL nor supervised fine-tuning produces a filler token benefit that persists at test time. Ultimately, the findings indicate that frontier models already perform consequential computation completely hidden from interpretable output traces.
Paper Metadata
Paper Metadata
- arXiv ID: arXiv:2607.22925 [cs.CL]
- Subjects: Computation and Language (
cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) - Authors: Vatsal Baherwani, Tom Goldstein, Ashwinee Panda
- Submission History:
- Submitted on 24 Jul 2026 (v1)
- Last revised 3 Sep 2026 (v2)
- DOI: 10.48550/arXiv.2607.22925
- arXiv ID: arXiv:2607.22925 [cs.CL]
- Subjects: Computation and Language (
cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)- Authors: Vatsal Baherwani, Tom Goldstein, Ashwinee Panda
- Submission History:
- Submitted on 24 Jul 2026 (v1)
- Last revised 3 Sep 2026 (v2)
- DOI: 10.48550/arXiv.2607.22925
Abstract
Abstract
A key question for AI safety is whether a language model expresses all of its reasoning in its output tokens. We demonstrate a concrete failure mode where frontier models exhibit invisible reasoning by leveraging semantically irrelevant filler tokens to improve performance on synthetic reasoning tasks. We evaluate 13 frontier language models across three tasks and find that many models benefit significantly from filler tokens, with accuracy improvements of up to 13 percentage points. The benefit depends on which tokens are used and differs across models. We further show that filler tokens enable Claude Opus 4.5 to satisfy a hidden modular arithmetic constraint without sacrificing accuracy on its primary task, demonstrating that invisible reasoning can serve objectives entirely invisible to CoT monitoring. Reinforcement learning gives Qwen3-235B strong preferences over filler token content, but neither RL nor supervised fine-tuning produces a filler token benefit that persists at test time. Our results indicate that frontier models already perform consequential computation with no interpretable trace in their output tokens.
Quick Links & Resources
Quick Links & Resources