Post-hoc KV cache editing

Method introduced in mini experiment 2 to steer persona activations in stored KV entries at specific layers and positions

Neighborhood — ranked by edge-count

paper

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

KV cachingmethod0.804
Caching of key-value pairs to avoid recomputation; also provides a mechanism for introspection of earlier computations.
KV cacheconcept0.802
Stored key and value vectors from previous tokens used by attention heads to carry information across token-time
KV caching overcomes statelessness and provides a mechanism for introspection of computations at earlier token positions.claim0.717
Janus's claim about KV caching as an introspection mechanism.
KV State Persistenceconcept0.693
The key-value cache from steered tokens is retained during no-steering continuation, allowing causal effect of steering to propagate
KV state persistence across steered tokensconcept0.658
The causal steering experiment persists KV state over steered tokens so downstream effects can be observed without continued steering