norgitov/ trends
Technology · people · ideas
Back to discovery/LessWrongyesterday

J++ Lens: Jacobian Filtering Enhances Faithful Workspace Lens Readouts

The J++ Lens improves upon the J-Lens by filtering noisy gradients to enhance the accuracy of reading intermediate variables from language model activations. It achieves a 55% success rate in latent variable extraction tasks, outperforming the J-Lens (36%) and R-Lens (38%) with better performance at earlier layers.

Open original
SIGNAL FROM THE SOURCE
50
source points
Tracking sinceOctober 8, 202650 source points

This source has not updated recently. Its observation time is shown above.

Momentum+6.77/hover 1.03 h
PublishedOctober 8, 2026Kola Ayonrinde
BEHIND THE NUMBERS

How interest changes

History starts here

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

50 source points

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

WHY IT MAY MATTER

The J++ Lens provides more accurate and reliable extraction of intermediate variables from language model activations, which can be useful for monitoring and detecting unverbalized reasoning patterns.

A useful discovery?
KEEP EXPLORING

Connected ideas

Explore topic