Interpretability tools transfer between AI models with just unlabeled prompts

A linear map fit on unlabeled data carries refusal, correctness, and sentiment directions across model families and sizes, matching native performance when the map's training data covers the tool's region.

Academic

Marcus Armstrong · Navid Ayoobi · Pradham Mummaleti · Alexander Chulzhanov · Arjun Mukherjee

University of Houston

Research Digest·11 November 2026·2 min read

In one line

Linear maps trained on prompts from a tool's region transfer interpretability directions across models at native performance.

Before you read this

Two ideas this paper builds on

activation steeringActivation steering modifies a model's internal vector representations during inference to shift behavior, such as editing outputs, removing capabilities, or controlling trust.72 papers →mechanistic interpretabilityThe study of how neural networks compute through internal circuits, features, and causal interventions, rather than just input-output behavior.129 papers →

Read it as

AI models have internal signals that control behavior, like a direction that makes the model refuse to answer. Such tools are usually built separately for each model. This study asks if they can be transferred using a simple mapping trained on unlabeled text.

The researchers tested this across many model pairs from several families, ranging from small to very large models. They transferred a refusal direction and a correctness probe, comparing maps built from generic questions versus task-specific prompts.

With generic questions, the transferred tool recovered 54% of the native performance. With task-specific prompts, recovery reached 102%, matching or exceeding the original. This shows interpretability findings can be reused across models, saving time and resources.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

activation steering mechanistic interpretability representation similarity model stitching coverage score