Regularization amplifies subliminal learning instead of blocking hidden biases, thread finds
AdtRaghunathan · x · 2026-11-07
A research thread opens with a counterintuitive claim: using regularization to stop models from picking up subtle biases in data doesn't work — it actually amplifies subliminal learning and other forms of covert influence.
The finding challenges the common intuition that adding constraints during training mitigates hidden biases, with direct implications for data-bias and alignment research.
Related event: Regularization Amplifies Subliminal Learning in AI Models, Study Finds(2 posts)→
More from Research
- GlobeBench: best of 20 LLMs hits only 49.6% as RL environment simulators — mariyaivasileva · 2026-11-07
- Turn gaps average ~200ms: duplex voice models must plan replies mid-listen — smolix · 2026-11-07
- Why Higgs Realtime Duplex uses a 160ms step: it matches human decision speed — smolix · 2026-11-07
- ByteDance experiments suggest Full Bandwidth Transformer matches models twice its size — DimitrisPapail · 2026-11-07
- Ai2 brings 10+ robotics papers to CoRL 2026, spotlighting MolmoAct2 and MolmoBOT — allen_ai · 2026-11-07
- Applied Compute shares a production RL recipe for continual learning from live traffic — ypatil125 · 2026-11-07