CAMEL: Fixed-Work Residual Memory for CPU-Oriented Language Models

Abstract

CAMEL investigates external residual memory for small language models intended for CPU execution. The main study uses a recurrent byte-level student with two compiled memories: an exact surface 4-gram memory and a learned latent residual memory. A 32-byte table selects between the two using support-count bins.On three held-out TinyStories runs, the selected hybrid improves bits per byte over surface memory, dense latent memory, and a 64-byte sparse latent baseline, without reducing next-byte accuracy relative to surface memory. A native lookup benchmark shows lower lookup-and-apply time than dense latent values. The same method does not outperform dense latent memory on enwik8. An address-inclusive benchmark also fails its exact-equivalence criterion because of one Python/C++ address disagreement.The paper also reports development experiments using Qwen3-0.6B as a teacher for a 2.43-million-parameter token-level GRU. In these experiments, exact token-history memory performs better than the learned product-context address. Residual bigram exceptions improve the token baseline on TinyStories at fixed logical work, but the unchanged method fails one of three enwik8 fits. An argmax-preserving projection keeps most of the loss improvement without changing the token baseline's predictions.The paper reports both successful and failed experiments and limits its claims to the tested models, datasets, and CPU kernels.