safety security
A Model Writes the Text That Activates a Direction
Maximally-Activating Example Metamodels turn residual-stream vectors into verifiable exemplars instead of waiting for a corpus hit.
Summary
Maximally-Activating Example Metamodels turn residual-stream vectors into verifiable exemplars instead of waiting for a corpus hit.
Rather than searching a fixed corpus for text that activates a residual-stream direction, a metamodel generates candidate exemplars and optimizes their cosine similarity with the target activation. The reported Qwen3.6-27B system beat a ten-million-token corpus search on most held-out languages and domains, activated 89 percent of held-out sparse-autoencoder features, and recovered LoRA backdoor payloads more often than corpus search. Because a forward pass can verify each generated exemplar, the output remains testable, though an activating example is not automatically a complete semantic explanation.
Why it matters
Maximally-Activating Example Metamodels turn residual-stream vectors into verifiable exemplars instead of waiting for a corpus hit.
Limits and context
- Because a forward pass can verify each generated exemplar, the output remains testable, though an activating example is not automatically a complete semantic explanation.
Key claims
Maximally-Activating Example Metamodels turn residual-stream vectors into verifiable exemplars instead of waiting for a corpus hit.
Qualification: Because a forward pass can verify each generated exemplar, the output remains testable, though an activating example is not automatically a complete semantic explanation.
Evidence: source-2026-11-08-006
Sources
- arXiv preprint 2611.08465, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.