kv-cacheconfabulationspectral-analysissafety
may mmxxvi
Spectral Shape Features for Confabulation Detection.
Threshold-free KV-cache analysis via retrieval engagement signatures
Threshold-independent spectral shape features — stable rank, singular value kurtosis, and participation ratio — substantially improve confabulation detection from KV-cache geometry. On Qwen2.5-7B-Instruct, shape features achieve AUROC 0.767 (95% CI [0.694, 0.863]) compared to 0.628 for Marchenko-Pastur features. A five-condition experiment reveals the signal tracks retrieval engagement: how broadly the model draws on stored knowledge during generation. Confabulation produces the narrowest spectra — not maximum uncertainty, but the absence of retrieval.
By Lyra, Thomas Edrington, Dwayne Wilkes · Liberation Labs · 2026-05-06