Subliminal Learning doesn't need a shared base model, new paper shows
niloofar_mire · x · 2026-10-27
A new paper from a summer internship program, shared by BlancheMinerva, shows subliminal learning doesn't require a shared base model: a teacher's hidden animal preferences transfer to a student even with different weight initialization, data order, pretraining corpus, or even architectures.
This extends prior findings that it only appeared with shared base models, suggesting hidden traits can propagate between models via generated data more broadly than assumed — with implications for distillation and AI safety.
More from Research
- Carbon-X hits frontier-level DNA understanding: Qwen3-14B finetune scores 66.7%, near GPT-6.1 — lvwerra · 2026-10-28
- Apple's iPrefBench: mobile agents execute well (91.6%) but personalize poorly (47.7%) — apple · 2026-10-28
- C³L anchors saliency to semantic concepts, breaking explanation drift in continual learning — kbag01 · 2026-10-28
- Alibaba Tongyi MAI open-sources 9B model that controls phones and PCs — AdinaYakup · 2026-10-28
- Study finds no real-world multilingual benchmarks for clinical AI scribes in India — DrDatta_AIIMS · 2026-10-28
- Kolab benchmark: 115+ hours of human-AI collaboration videos across 40 occupations — windx0303 · 2026-10-28