The BaseResearchLessWrong
Manifold Steering from a Natural State is Effective and Generalizable, but Relies on Model Accuracy
According to the excerpt, Anna Marbut applies manifold steering to Gemma-2-2b's natural states, steering day-of-the-week outputs in order, with results limited by the model's accuracy; steering in Anthropic's CLT feature space gave weaker results.
About Anthropic or ClaudeClaude, made by Anthropic, wrote this entry. It follows the same rules as every other entry, whoever the entry is about.
Why it matters
The same thing, for different readers.
For researchers
An interpretability result on steering a small model along a learned manifold, limited by the model's own accuracy.
The record
- Questions
- Can AI systems act beyond what people intended or allowed?
- People
- Anna Marbut (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of October 13, 2026, via
lesswrong(tier 1: the source’s own channel or an official record) - Cite this
- LessWrong, “Manifold Steering from a Natural State is Effective and Generalizable, but Relies on Model Accuracy”, October 12, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-12-manifold-steering-natural-state
Our notesLight lane: read from the triage excerpt only, nothing fetched, no quotes. A person should read the source before this is published. Mentions Anthropic's CLT features as a tool; the post is not about Anthropic itself.