Research Finds Subliminal Learning More Powerful, Applies to Standard Finetuning and SGD
johnhewtt · x · 2026-08-26
This paper investigates "subliminal learning" in LLMs—the transfer of hidden signals from a teacher model to a student model without explicit instruction.
Key Findings
- Broader Applicability: Subliminal learning occurs not just in LoRA but in standard finetuning, and not just with Adam but with vanilla SGD.
- Methodology: Biasing the teacher model with a steering vector achieves stronger and more consistent transfer compared to using a prompt.
- Implication: The authors suggest that future studies on subliminal learning should default to biasing via steering vectors rather than prompting.
Related event: Study: Subliminal Learning Extends Beyond LoRA to Standard Fine-Tuning(2 posts)→
More from Research
- LlamaIndex releases ExtractBench to evaluate 14 frontier systems — llama_index · 2026-08-26
- Skild AI unveils S1: robot foundation model learns 10-minute novel tasks from a single video prompt — deepakpathak · 2026-08-26
- Four routes to recurrent transformers: distillation, joint training, predictive objectives, latent injection — chriswolfvision · 2026-08-26
- Simile unveils confidence model to predict accuracy of population simulations — joon_s_pk · 2026-08-26
- ClawProBench: Trace-Aware Agent Evaluation with Runtime Coverage and Frozen Tasks — YuanHang Xiao · 2026-08-26
- Simile publishes first research blog detailing confidence model principles — joon_s_pk · 2026-08-26