Delta Compression: An Efficient Lossless Compression Method Using Hypernetwork-Generated Parameter Deltas

This paper proposes ”Delta Compression,” a novel method for achieving efficient lossless compression by encoding information as a hypernetwork-generated parameter delta rather than static data. When applied to a shared, deterministic base model trained for reconstructive decoding, this delta enables perfect recovery of the original input. In our proof-of-concept (PoC), we encoded 160 texts (978 KB) into a single LoRA adapter (fp16, 849 KB) on a LLaMA-3-8B model and restored them with complete fidelity. Furthermore, we demonstrate that the adapter's weight distribution is highly concentrated near zero, enabling quantization to 11-bit integers, which reduces the size to 583 KB without compromising reconstruction accuracy. While our method introduces a novel paradigm, it is also applicable to conventional scenarios requiring lossless storage of language data. As AI systems continue to generate large volumes of semantically rich data—particularly in contexts where reproducibility and traceability are essential—Delta Compression is especially promising for long-term archiving applications. At the same time, it enhances existing approaches to lossless language data compression by offering a scalable and efficient alternative.