1 item with this tag.

  • papers

    ReALLM

    Paper note on compressing language-model matrices into residual low-rank structure plus a shared neural decoder over vector-quantized latent representations.