GaugeQuant

A new method called GaugeQuant, introduced in a paper on arXiv, reduces perplexity in quantized large language models by learning quantization-optimal bases from internal symmetries during training. U…