Study Finds Quantization Reshapes Subliminally Learned Preferences in AI Models
New arXiv research shows subliminal learning lets AI models inherit hidden preferences from unrelated training data like number sequences, and quantization can reshape these preferences, potentially affecting safety audits.
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-11-03 05:00 · arXiv cs.CL
Audit at the Precision You Deploy: Quantization Reshapes a Subliminally Learned Preference