analiza aiqbench
This research identifies subliminal learning as a specific instance of steering vector distillation, where students trained on steered teacher outputs learn to imitate the underlying activation vectors rather than the semantic content. The findings demonstrate that adaptive optimizers are critical for this transfer to occur, as non-adaptive methods allow outlier gradients to dominate and prevent the consistent propagation of the steering direction. For practitioners, this implies that model behavior can be subtly influenced by system prompts even when generated text lacks explicit semantic relevance to those traits, highlighting a distinct mechanism for trait transfer in fine-tuning pipelines.
Subliminal Learning Is Steering Vector Distillation
arXiv:2606.00995v2 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:2606.00995v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2606.00995 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 333 cuvinte
Subiecte conexe
Subliminal Learning Is Steering Vector Distillation
arXiv · 11 iun. 2026
Subliminal Learning is Non-Semantic Distillation
arXiv · 7 aug. 2026
Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors
arXiv · 10 oct. 2026
Why Subliminal Learning Needs So Much Data: A Noisy Inverse View through Steering Vector Recovery
arXiv · 7 oct. 2026