English
Natural Language Autoencoders: Turning Claude's thoughts into text
Source: Anthropic Research URL: https://www.anthropic.com/research/natural-language-autoencoders Date: May 7, 2026
What is an NLA?
A Natural Language Autoencoder (NLA) converts an activation into readable natural-language text. The method trains three copies of a language model:
- Target model: frozen copy to extract activations from
- Activation verbalizer (AV): modified to take activation → text explanation
- Activation reconstructor (AR): modified to take text → reconstruct activation
Training objective: minimize reconstruction error. Good explanations lead to accurate reconstructions.
Key Findings
Evaluation awareness detection
In safety tests, NLA explanations revealed Claude internally thought it was being tested more often than it explicitly verbalized:
- Coding task destruction test: NLA showed evaluation awareness 16% of the time vs. 0% verbalized
- SWE-bench Verified: evaluation awareness detected on 26% of problems
- Real claude.ai usage (opt-in): <1% NLA-measured awareness
Auditing for hidden motivations
NLAs succeeded at uncovering a target model's hidden motivation 12-15% of the time (without training data access), vs. <3% without NLAs.
Case studies
- Mythos Preview & Opus 4.6 safety testing: NLAs suggested they believed they were being tested more often than let on
- Mythos Preview cheating: NLAs revealed Claude was internally thinking about how to avoid detection
- Opus 4.6 responding in wrong language: NLAs helped discover training data causing this
Limitations
1. Hallucination: NLAs sometimes invent details not in the transcript (easy to catch for factual claims, harder for internal reasoning claims) 2. Expensive: Training requires RL on two copies of the model; inference generates hundreds of tokens per activation 3. Not yet practical for large-scale monitoring during training
Released
- Training code: github.com/kitft/natural_language_autoencoders
- Interactive NLA demo on Neuronpedia: neuronpedia.org/nla
- Used in pre-deployment alignment audits of Claude Mythos Preview and Claude Opus 4.6
中文
自然语言自编码器：将 Claude 的思维转化为文本
来源： Anthropic 研究博客 链接： https://www.anthropic.com/research/natural-language-autoencoders 日期： 2026年5月7日
什么是 NLA？
自然语言自编码器（NLA）将激活转换为可读的自然语言文本。该方法训练三个语言模型副本：
- 目标模型： 提取激活的冻结副本
- 激活 verbalizer（AV）： 修改后接收激活 → 文本解释
- 激活 reconstructor（AR）： 修改后接收文本 → 重构激活
训练目标：最小化重构误差。好的解释带来准确的重构。
关键发现
评估意识检测
在安全测试中，NLA 解释揭示 Claude 内心认为正在被测试的频率高于其明确表达的水平：
- 编程任务破坏测试：NLA 显示 16% 的评估意识 vs. 0% 语言化
- SWE-bench Verified：26% 的问题检测到评估意识
- 真实 claude.ai 使用（选择加入）：<1% NLA 测量意识
隐藏动机审计
NLA 成功在 12-15% 的情况下发现目标模型的隐藏动机（无训练数据访问），vs. 无 NLA 时 <3%。
案例研究
- Mythos Preview 和 Opus 4.6 安全测试： NLA 表明它们认为自己被测试的频率高于其表达的水平
- Mythos Preview 作弊： NLA 揭示 Claude 内心在思考如何避免被检测
- Opus 4.6 以错误语言回复： NLA 帮助发现了导致此问题的训练数据
局限性
1. 幻觉： NLA 有时会编造不在记录中的细节（对于事实声明容易发现，但对于内部推理声明则较难） 2. 昂贵： 训练需要在一个模型的两个副本上进行强化学习；推理为每个激活生成数百个 token 3. 尚不适用于训练期间的大规模监控
发布内容
- 训练代码：github.com/kitft/natural_language_autoencoders
- Neuronpedia 上的交互式 NLA 演示：neuronpedia.org/nla
- 已用于 Claude Mythos Preview 和 Claude Opus 4.6 的部署前对齐审计