Intelligently choosing how to round weights during quantization based on loss landscape information rather than simple truncation.