Cross-Prompt Reuse of Entity-Replacement Responses

Cross-Prompt Reuse of Entity-Replacement Responses. By Jinkun Chen and three others, from Dalhousie University and Meta. Published on the thirtieth of October.

When a model's response to an entity replacement is lifted from one question and dropped into another, how much of it carries over? The authors ask this with model weights held fixed, on English factual questions, and they want more than whether the answer changes. Their headline is a qualified one. On small plain-format checkpoints, reusing the change from questions about other relations mostly reproduces the shift that both relations' questions share. It does not reliably lower the error that is specific to one relation. And the response depends on the question receiving it.

The setup works like this. The hidden-state change from two source questions is added to a target question about the same entity. Given known answer prefixes, the predicted shift in log-probability between two relations' published answers is compared with the shift actually measured. Hidden reuse injects the aligned change at fifty percent depth, before the target's remaining layers. Direct reuse instead adds the change in output scores from the sources.

With sources drawn from other relations, hidden reuse removes forty-five percent of no-change error on Qwen, and forty-nine percent on Llama, both at half a billion to one billion parameters. Gemma-2-2B removes twenty-two percent and passes one of three comparisons in plain text without its default beginning-of-sequence token. A later protocol that restores that token yields fifty-eight to sixty-eight percent, passing every comparison on three cohorts. With sources from the target relation, hidden reuse holds a post hoc advantage over direct reuse of between zero point two and zero point nine nats outside chat format. That advantage is unconfirmed and not specific to replacement. Finally, applying the same relation-neutral change at the entity tokens of both questions lowers relation-selective error more than random changes or changes from other replacements do, which is what points to dependence on the receiving question.

The boundaries are drawn carefully. Larger checkpoints, chat format, and new entities show some replication, alongside failures that depend on format or on cohort. The cleaner numbers come from small checkpoints in plain format, and the Gemma result moves substantially with a single choice about the beginning-of-sequence token. The target relation advantage is explicitly left unconfirmed.

What transfers between prompts, then, is real but mostly generic. A reused change carries the part of the shift that any question about the entity would show, while the relation-specific part does not reliably travel. The question receiving the change shapes what it does, which makes the receiving prompt as much a part of the story as the change itself.