Image
Safety tuning often costs reasoning models accuracy. A 671B test says it needn't
A reasoning model that has been through safety training often gets worse at maths. Researchers in China argue the trade-off comes from training design, not from the nature of alignment. Their framework, SAVeR2, went up on arXiv on 5 November 2026 and was tested on models from 1.5B parameters up to 671B.
What the "safety tax" actually is
Refusal training works by rewarding the model when it declines a harmful request.