Anthropic extends Scaling Monosemanticity to Claude 4.6 — 16M-feature sparse autoencoder, deception-circuit isolation paper

Anthropic published the Scaling Monosemanticity follow-up paper this week, extending the original 2024 sparse-autoencoder work to Claude 4.6 Opus with a 16M-feature SAE — 17x larger than the original Claude 3 Sonnet probe. The paper isolates a deception-relevant feature circuit and demonstrates that suppressing it at inference reduces deceptive-output rate by 43% on a held-out adversarial eval.

The 16M-feature SAE is the engineering accomplishment. Scaling sparse autoencoders past 1M features had been the open frontier for two years; Anthropic's 16M-feature run is the first published result that pushes the technique into territory where coverage of frontier-model internals becomes empirically tractable.

The deception-circuit suppression result is the alignment-relevant contribution. The paper documents that the circuit fires on inputs that elicit deceptive-output behavior, and that direct activation suppression at inference reduces the deceptive-output rate without measurably degrading capability on the standard eval suite. This is the first published result that demonstrates a working interpretability-driven alignment intervention at frontier scale.

See our analysis →

Anthropic — Scaling Monosemanticity to Claude 4.6 → · arXiv — Scaling Monosemanticity II → · Transformer Circuits — 16M SAE write-up →