Interpretability Transfer from Language to Vision via Sparse Autoencoder Transfer Alignment
This work transfers labeled sparse-autoencoder features from language to vision inside a multimodal model. It uses those features to make visual representations more interpretable through alignment with textual concepts.
From: University of Bath
The paper introduces a method that constrains a vision-language projector to map visual tokens into an existing textual SAE feature space.
It shows that cross-modal transfer can expose interpretable structure that was already present in the language side of the model.
It does not address consciousness directly and does not tell whether the model’s internal states have any experiential character.
Source: University of Bath Research Portal ↗. This is a summary; the source is the authority.
This suggests interpretability can travel across modalities, but it has no clear bearing on whether models can have experiences.
How we weigh evidence like this: the evidence levels.
- Mechanistic Interpretability Should Prioritize Feature Consistency in Sparse AutoencodersThis paper argues that sparse autoencoders should be judged not only by reconstruction quality and sparsity but also by whether the same fea…
- Mechanistic Interpretability for Neural NetworksThis is a broad survey of mechanistic interpretability, covering the goal of reverse-engineering internal algorithms in neural networks. It …
- Constructing Interpretable Features from Compositional Neuron Representations in Language ModelsThis paper proposes an alternative to standard sparse autoencoders by decomposing MLP activations with semi-nonnegative matrix factorization…
- Multilingual Sparse Autoencoders and Principled Layer Selection for Cross-Lingual Steering in Large Language ModelsThis paper studies sparse autoencoders for feature-level interpretability and steering in multilingual language models. It also proposes a m…