Qwen3.5 27B model rewrote its own weights and gained 6.6 points on unseen math
UIUC researchers report Qwen3.5-27B edited its own parameters with no gradient training, lifting held-out accuracy from 57.2% to 63.8% in one accepted session.
By Ryan Merket · Published · Updated

UIUC researchers report Qwen3.5-27B edited its own parameters with no gradient training, lifting held-out accuracy from 57.2% to 63.8% in one accepted session.