Search papers, labs, and topics across Lattice.
This paper introduces the Universal Activation Bus, a framework that standardizes activation interfaces across various language models, allowing for the reuse of activation-based tools without the need for retraining. By employing a lightweight linear encoder-decoder adapter for each model, the framework enables new models to integrate seamlessly into the shared activation space using only unlabeled matched text. The findings demonstrate that semantically related texts maintain consistent neighborhoods in this shared space, facilitating effective tool reuse and cross-model predictions without retraining.
A single adapter can unlock shared activation tools across different language models, enabling seamless integration and effective reuse without retraining.
Activation-based tools are usually tied to one model's native hidden space, requiring probes, sparse autoencoders, and natural-language interpreters to be rebuilt or rediscovered for each new language model. We present a Universal Activation Bus, a framework that provides a common activation interface across compatible language models. Using a small set of source models, we learn a shared dense space together with one lightweight linear encoder--decoder adapter pair per model. After source training, the interface is frozen; a new model joins by fitting only its adapter pair on unlabeled matched text. The resulting interface allows activation-based tools to be shared across connected models, including common probes and SAE features as well as access to an NLA originally trained for a different model. Across five models, semantically related texts form consistent neighborhoods in the shared space, and an onboarded model reuses these tools effectively without retraining them. We further show that an intermediate activation from one model can be used by another model's frozen upper layers to produce predictions. These results establish a stable, model-wise activation contract for reusable tools across compatible language models.
