Activation Reconstruction Fidelity

Activation reconstruction fidelity is the degree to which a generated explanation preserves enough information to recreate the original model activation.

Key points

• In Anthropic’s NLA setup, researchers cannot directly know whether a natural-language explanation is “true”, because the ground-truth thought encoded by an activation is not directly observable ^[src-066].
• The workaround is a round trip: original activation -> text explanation -> reconstructed activation ^[src-066].
• A better explanation is one that enables a more accurate reconstructed activation, so reconstruction similarity becomes a training signal and evaluation metric ^[src-066].
• This makes interpretability partly testable: the explanation is still text for humans, but the method has a quantitative pressure to retain information from the activation ^[src-066].
• High reconstruction fidelity does not guarantee semantic truth. Anthropic still warns that NLA explanations can hallucinate and need independent corroboration ^[src-066].

Related entities

• Natural Language Autoencoders
• Anthropic

Related concepts

• Model Interpretability
• Evaluation Awareness
• Model Auditing Games

Source references

• ^[src-066] Anthropic – “Natural Language Autoencoders: Turning Claude’s thoughts into text” (2026-05-07)