moyix: looped models may resist distillation attacks by hiding reasoning in latent space
moyix · x · 2026-10-09
Security researcher moyix wonders whether looped models are less susceptible to distillation attacks, since more of the "reasoning" occurs in latent space rather than as executable chain-of-thought — reframing CoT as "extractable" rather than "executable." A notable technical take on model distillation defenses.
Related event: Looped Models May Resist Distillation Attacks(2 posts)→
More from Research
- Kuaishou's LIFT unifies retrieval and ranking, beats baselines by 4.9% on ML-20M — _reachsumit · 2026-10-09
- IBM's RIT-RAG induces document sub-trees from retrieved chunks, lifting RAG accuracy by up to 11.4 points — _reachsumit · 2026-10-09
- EVIE: multimodal-judge-trained visual document retrievers with 6 nested embedding sizes per checkpoint — _reachsumit · 2026-10-09
- Autoregressive Retriever (ARR) Refines Queries with Retrieved Item Feedback via SFT and RL — _reachsumit · 2026-10-09
- Project Greenhouse: Jimmy Lin Team Trains Fully Open Reranker from Scratch on a Handful of GPUs — _reachsumit · 2026-10-09
- Sony's Syn-Omni: Shared + Expert LoRA Paths Beat Omnimodal Embedding Baselines Across 81 Tasks — _reachsumit · 2026-10-09