Our original subliminal learning paper showed subliminal learning in MNIST with MLPs. We show this holds across many ablations. In particular, it holds with full fine-tuning (not LoRA) and with SGD (as well as other optimizers).
We also prove a theorem about subliminal learning, which applies to SGD, full-weight updates, and arbitrary neural networks. https://www.nature.com/articles/s41586-026-10319-8/figures/7
Users describe subliminal learning findings as a promising research direction relevant to AI alignment because the effect holds across fine-tuning, SGD, and arbitrary networks.
For clarity, the LLM results in our paper do use LoRA (for open-weight models) or OpenAI's fine-tuning API (which likely uses some kind of parameter-efficient fine-tuning). https://www.nature.com/articles/s41586-026-10319-8
Our original subliminal learning paper showed subliminal learning in MNIST with MLPs. We show this holds across many ablations. In particular, it holds with full fine-tuning (not LoRA) and with SGD (as well as other optimizers).
We also prove a theorem about subliminal learning, which applies to SGD, full-weight updates, and arbitrary neural networks. https://www.nature.com/articles/s41586-026-10319-8/figures/7
@tmkadamcz This is not a new result. But just highlighting this old result in the light of some recent papers.
@OwainEvans_UK curious why ur focusing a lot on SL these days? what do u hope to get from this research direction?
@OwainEvans_UK Where could we find the updated theorem?
@OwainEvans_UK The link is just a link to a particular figure. Confused. Is there a new result here?
@OwainEvans_UK please see
@thkostolansky It's not a major focus right now for me. We published the original paper 11 months ago. But I do think it's a promising research direction and relevant to alignment.
Our original subliminal learning paper showed subliminal learning in MNIST with MLPs. We show this holds across many ablations. In particular, it holds with full fine-tuning (not LoRA) and with SGD (as well as other optimizers).
We also prove a theorem about subliminal learning, which applies to SGD, full-weight updates, and arbitrary neural networks. https://www.nature.com/articles/s41586-026-10319-8/figures/7