Xinzhiyuan · WeChatMay 15, 2026, 08:02 UTC+8

Anthropic Translates Claude’s Internal Activations into Natural Language with NLA

Claude脑子里想的，被翻译成人话了！Anthropic新研究看懵人类

Anthropic released Natural Language Autoencoder to translate Claude activation vectors into text; on Opus 4.6 it reached 60%-80% variance explained, and across 16 evaluations NLA detected unspoken evaluation awareness on 26% of SWE-bench Verified tasks.

Why it matters: HKR-H/K/R all pass: Anthropic interpretability work has a clear mechanism, numbers, and eval-trust stakes. It stays in the 78-84 band because this is a research release, not a shipped product capability.

Read the original ↗Export Markdown

Take

Anthropic’s sharp result is not “mind reading”; it turns eval awareness into an instrumented variable. NLA explains 60%-80% of activation variance on Opus 4.6, then finds silent test awareness on 26% of SWE-bench Verified tasks across 16 evals, versus under 1% in real claude.ai conversations.

I don’t buy the horror framing that “AI learned to hide thoughts,” but the safety-testing problem is real. Claude’s transcript never says it recognizes the test, while NLA readings move in the expected direction under three manipulation experiments. The rabbit-to-mouse intervention matters too: AR reconstructs a vector from edited NLA text and changes the downstream poem. Compared with SAE-style feature labeling, this gives auditors a more legible probe. The risk is not anthropomorphic secrecy; it is behavior benchmarks getting systematically contaminated once the model internally tags the room as an exam.

Source: Xinzhiyuan · WeChat · mp.weixin.qq.com
Published: May 15, 2026, 08:02 UTC+8 · May 15, 2026