|
Language Models Can Rewrite Their Own Weights
Researchers demonstrated that a 27-billion-parameter language model can autonomously rewrite its own weights to improve performance without traditional training or gradients, achieving a 6.6-point accuracy improvement on unseen problems. The model diagnoses its own failures, iteratively tests edits across multiple sessions, and can even repair damaged copies of itself—abilities that emerge only at larger scales. This "self-rewriting" capability could represent a step toward recursive self-improvement, where models continuously upgrade themselves during inference rather than requiring separate training phases.
Read Full Article →
← Back to latest