Large language models can use conditional memory to expand capacity through selective lookup, but the full embedding table must remain available for every possible input. That requirement creates a significant memory overhead. A new arXiv paper proposes a compressed representation called Conditional Program (CP) to address this.

The paper's title indicates a focus on two-bit quantization of lookup tables, suggesting that CP may store these tables at very low precision. The abstract, however, is truncated before explaining how CP works or what trade-offs it introduces. As a result, the specific mechanism and expected savings are not yet clear from the announcement.