Camila Blank

Camila Blank, Agam Bhatia, Senthooran Rajamanoharan et al.

Subliminal learning refers to a student language model acquiring a teacher's traits (e.g. a system-prompted preference for owls) when fine-tuned on the teacher's outputs, despite the outputs being semantically unrelated to those traits. It remains poorly understood how data without semantic meaning can transfer specific semantic traits. In this work, we show that subliminal learning is mediated by a single steering vector, i.e. a vector added to the model's activations. Across two open-source models, we find that the teacher's system prompt is well approximated by a steering vector, and that the student's behavior is driven by learning an aligned vector over fine-tuning. System prompts that are not well approximated by steering vectors are not subliminally learned. This is a special case of steering vector distillation, in which a student trained on the outputs of a steered teacher learns to imitate that steering. We demonstrate steering vector distillation on a range of semantic and random vectors. Adding a semantic vector to a model's activations can have both model-independent and model-specific (i.e. non-semantic) effects on its behavior, so generated data that is non-semantic can transmit a vector with semantic effects, enabling subliminal learning. This also explains why subliminal learning does not transfer between models. We find that adaptive optimizers are necessary for subliminal learning in language models: activation gradients on steered data carry a small but consistent component along the steering direction, and non-adaptive optimizers impede this by allowing outlier gradients to dominate.

Gustavo Chau Loo Kung, Mohammad Abbasi, Camila Blank et al.

Diffusion Magnetic Resonance Imaging (dMRI) plays a critical role in studying microstructural changes in the brain. It is, therefore, widely used in clinical practice; yet progress in learning general-purpose representations from dMRI has been limited. A key challenge is that existing deep learning approaches are not well-suited to capture the unique properties of diffusion signals. Brain dMRI is normally composed of several brain volumes, each with different attenuation characteristics dependent on the direction and strength of the diffusion-sensitized gradients. Thus, there is a need to jointly model spatial, diffusion-weighting, and directional dependencies in dMRI. Furthermore, varying acquisition protocols (e.g., differing numbers of directions) further limit traditional models. To address these gaps, we introduce a diffusion space rotatory positional embedding (D-RoPE) plugged into our dMRI transformer to capture both the spatial structure and directional characteristics of diffusion data, enabling robust and transferable representations across diverse acquisition settings and an arbitrary number of diffusion directions. After self-supervised masked autoencoding pretraining, tests on several downstream tasks show that the learned representations and the pretrained model can provide competitive or superior performance compared to several baselines in these downstream tasks (even compared to a fully trained baseline); the finetuned features from our pretrained encoder resulted in a 6% higher accuracy in classifying mild cognitive impairment and a 0.05 increase in the correlation coefficient when predicting cognitive scores. Code is available at: github.com/gustavochau/D-RoPE.