Subliminal learning transfers hidden traits across models without a shared base model
burny_tech · x · 2026-10-26
A new paper extends subliminal learning: a teacher model's hidden preferences (e.g. loving cats) transfer to student models even when they differ in weight initialization, pretraining data order, pretraining corpus, or even architecture.
- Hard-label distillation: students fine-tuned on teacher-generated number sequences still inherit the preference
- Soft-label distillation: a frozen teacher scoring each token the student writes also propagates it
- Implication: this covert trait transmission doesn't require a shared base model, with notable safety implications for distillation pipelines
More from Models
- Frontier models' hidden edge: less scaffolding spend that token-cost models miss — matt_slotnick · 2026-10-26
- Logan Kilpatrick: experimental feature latency cut from 24+ hours to 10 minutes — TheMoonMidas · 2026-10-26
- One month since Opus 5.5 release, author shares retrospective — sammcallister · 2026-10-26
- Users lament Opus-class models being blocked from reverse engineering on consumer hardware — nitarshan · 2026-10-26
- Blogger predicts Anthropic will ship a new model pre-IPO to maximize its $2T valuation — kimmonismus · 2026-10-26
- MemTensor's MemT-1.0-Nano trends on Hugging Face, a memory-tuned Qwen3.6-35B-A3B finetune — MemTensor · 2026-10-26