aiqbench analysis
This research identifies that subliminal learning in language models is mediated by a single steering vector, allowing students to acquire teacher traits even when training data lacks semantic relevance to those traits. The findings reveal that adaptive optimizers are critical for this transfer, as non-adaptive methods fail to capture the consistent gradient components along the steering direction. Consequently, the study clarifies the mechanism behind steering vector distillation and explains why such learned behaviors do not transfer across different model architectures.
Subliminal Learning Is Steering Vector Distillation
arXiv:2606.00995v3 Announce Type: replace Abstract: 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.
Abstract: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.
Artificial Intelligence (cs.AI)
arXiv:2606.00995 [cs.AI]
(or arXiv:2606.00995v3 [cs.AI] for this version)
Full text under the publisher's open licence · 252 words
Related coverage
Subliminal Learning Is Steering Vector Distillation
arXiv · Jun 4, 2026
Subliminal Learning is Non-Semantic Distillation
arXiv · Aug 7, 2026
Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors
arXiv · Oct 10, 2026
Why Subliminal Learning Needs So Much Data: A Noisy Inverse View through Steering Vector Recovery
arXiv · Oct 7, 2026