Cross-Prompt Reuse of Entity-Replacement Responses
| Source: arXiv AI
Tags: mechanistic interpretability, Qwen, Llama, Gemma-2-2B, hidden-state intervention
A new arXiv paper measures how much of an entity-replacement hidden-state change transfers between factual questions in Qwen, Llama and Gemma models, finding gains that are partial and format-dependent.
Details
The authors add hidden-state changes from two source questions to a target question about the same entity, injecting at 50% depth, and compare predicted versus measured log-probability shifts. With other-relation sources, hidden reuse removes 45% and 49% of no-change error on 0.5-1B Qwen and Llama, and 58-68% on Gemma-2-2B once its BOS token is restored. Larger checkpoints, chat format and new entities show format- or cohort-dependent failures. The gains mostly reproduce shared shifts rather than relation-specific ones, making this a narrow interpretability result.