norgitov/ trends
Technology · people · ideas
Back to discovery/LessWrong1 hour ago

CLT Features Support Manifold Steering, but Hold Only Partial Steering Signal

The study extends cyclical manifold methods to Anthropic's CLT feature space within Gemma-2-2B, showing that while manifold steering works in both activation and CLT spaces, CLT features provide a weaker steering signal, indicating they capture only part of the day-of-week information.

Open original
SIGNAL FROM THE SOURCE
7
source points
Tracking sinceSeptember 21, 20267 source points
MomentumMore observations needed
PublishedSeptember 21, 2026Anna Marbut
BEHIND THE NUMBERS

How interest changes

History starts here

The chart will appear after repeat observations. The current metric comes from the source.

7 source points

Real observations only. History before source connection is not reconstructed.

WHY IT MAY MATTER

This research provides insights into how different feature spaces within a language model encode cyclical patterns, which could inform future work on model interpretability and control.

A useful discovery?
KEEP EXPLORING

Connected ideas

Explore topic