Subliminal Learning is Non-Semantic Distillation
This paper studies subliminal learning in the context of mechanistic interpretability. It was accepted as a spotlight paper for the ICML 2026 Mechanistic Interpretability Workshop.
From: Ethan Hadley, Eren Gultepe
The work appears to examine how information is transferred or compressed inside model training or distillation settings.
Its framing places it within interpretability and internal representation analysis rather than consciousness research.
It does not show that models have subjective experience, and it does not establish reliable introspection by the model itself.
Source: arXiv ↗. This is a summary; the source is the authority.
This is relevant to how internal representations move through models, but it adds little direct evidence about AI 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…
- Interpretability Transfer from Language to Vision via Sparse Autoencoder Transfer AlignmentThis work transfers labeled sparse-autoencoder features from language to vision inside a multimodal model. It uses those features to make vi…