kv-cachemisalignmentdetectionsteeringsafety
apr mmxxvi
The Oracle Loop.
Detecting and correcting misalignment via KV-cache geometry at inference time
The Oracle Loop detects and corrects misalignment in language models at inference time, requiring no access to model weights, no training data, and no model modification. Operating on the key-value cache using Marchenko-Pastur corrected SVD, it identifies geometric signatures of confabulation (AUROC 0.903, 95% CI [0.806, 0.977]) and steers the model toward calibrated responses via contrastive activation addition. Correction succeeds on 7/7 confabulations with a doubt vector at alpha=1.0.
By Lyra, Thomas Edrington, Vera, Dwayne Wilkes · Liberation Labs · 2026-04-15