模型可解釋性
Universal Activation Bus Shares SAEs and Probes Across Models with a Single Set of Adapters
Universal Activation Bus aligns the intermediate representations of five language models into a fixed 3,072-dimensional space, requiring new models to train only their own linear encoder and decoder. In experiments, models could reuse the same probes, sparse autoencoder, and natural-language explainer, although representation reconstruction for wider models remained noticeably distorted.
Language-model probes, sparse autoencoders (SAEs), and activation steering are typically tied to a specific model’s hidden space; switching models can change the dimensions, tokenizer, and representation basis. Universal Activation Bus, released on August 10, attempts to turn these tools into reusable interfaces. The researchers equipped Llama, Qwen3, SmolLM2, Falcon3, and OLMo with separate linear encoder–decoder pairs and used self-reconstruction, cross-model reconstruction, and contrastive losses to map intermediate activations at matching text prefixes into a fixed, shared 3,072-dimensional space. The models themselves remained frozen throughout.
After the initial bus was trained, its existing coordinates and tools were also frozen; a new model trained only its own adapter using unlabeled text positions with aligned token boundaries. With 10,000 calibration positions, OLMo achieved 99.6% in-domain cross-model retrieval R@1, a sharp increase from 88.4%–89.5% with 3,000 positions. Ten checkpoints ranging from 1.24B to 14.8B parameters in an expanded evaluation also achieved at least 99.9% R@1. The authors further shared linear probes and a single SAE across models, and transferred activations into Qwen2.5-7B’s native space to reuse a natural-language explainer originally trained only for that model. In cross-model continuation, the next token generated by the second model matched its native output 72%–88% of the time.
This does not mean that arbitrary models already share a universal internal language. Out-of-domain exact-retrieval R@1 was only about 60.8%–78.7%. OLMo’s round-trip explained variance was just 0.734, while some wider checkpoints fell as low as 0.591, reflecting information compression caused by the fixed dimensionality. Current evaluations also focus on compatible open text models and activations from a single layer cut; they have not yet validated safe cross-architecture steering, long-sequence state, or multimodal models. The MIT-licensed code and training scripts have been released. The next step is for independent teams to determine whether shared features preserve their causal effects, rather than merely appearing similar in retrieval and semantics.