Pan Zhang

LLMs can be compressed by nearly 50% with minimal accuracy loss using a new tensor decomposition method, but pushing further reveals abrupt, model-specific compressibility limits tied to shifts in internal representations.