K/V-cache changes affect language model persona without matching word use
K/V-Cache Interventions Dissociate Representation Alignment from Persona Expression in Decoder-Only Language Models
Computation and Language
Summary
This study looks at how changing certain internal memory parts in large language models affects their ability to adopt different writing personalities or personas. The researchers found that some memory swaps align internal representations closely but don’t always change the model’s actual word choice like you’d expect. Also, messing with the order of memory entries usually stops the model from showing a target persona reliably. The work shows that just comparing internal model signals doesn’t fully predict how well a model expresses a new persona.
What this means in practice
- •For ai developers: Guide persona control techniques by selecting mid-layer K/V-cache replacements to better preserve natural language diversity in AI generation.
- •For chatbot designers: Improve chatbot persona customization by understanding how internal memory changes impact target persona behavior beyond simple representation matching.
Authors
Yu Sun, Mengyin Lu, Cong Feng, Guangming Lu, Huimin Han
Abstract
We study K/V-cache interventions -- transplanting a target-conditioned K/V trajectory into a source-persona generation -- as a structured surface for persona control in decoder-only language models. Across 13 intervention configurations applied to Llama-3.1-8B for a fixed source-to-target persona pair, we report two consistent dissociations between representation-level alignment and behavioral expression, plus a common failure under position perturbations. First, all layer-band K/V replacements (early, mid, late) achieve strong local V-space alignment (V-gap 0.91, 0.89, 0.84), but only mid-layer replacement (layers 9-20) combines substantial target-marker expression with preserved lexical diversity. Second, full and mid-layer replacement induce comparable alignment (V-gap 0.94 vs. 0.89) yet produce different lexical-diversity profiles (TTR 0.65 vs. 0.77). Third, position perturbations (lag and shuffle) apply distinct operations yet uniformly suppress target-persona expression -- a common behavioral failure rather than a strict dissociation. Representation-level similarity metrics alone are thus not sufficient predictors of downstream persona expression in the regimes we study; the K/V cache emerges as a controllable but structurally constrained intervention surface. Because the transplanted trajectory carries the target's own generated token history, we characterize the intervention as trajectory-level transplantation rather than isolated persona-representation injection; a same-token-sequence control, decoding an identical token sequence under source vs. target conditioning, reproduces the sign and layer localization of the L28 representational shift, indicating the shift is not explained solely by imported token history. These findings characterize representation-behavior dissociation in a high-signal setting rather than establishing universality across models or persona pairs.