MoME: Mixture-of-Memory Embeddings for Context-Aware Sparse Lookup
MoME introduces context-aware memory embeddings with gated multi-slot lookup, improving LLM efficiency across Llama-3 and Qwen3 backbones.
Mixture of Memory Embeddings (MoME) replaces each token's single memory row with a mixture of M slots selected by a learned gate over the hidden state, resolving sense-collision in deterministic lookups. In controlled pretraining on nanochat, Llama-3/MobileLLM, and Qwen3 backbones, MoME beats Value Embedding, Bigram, and STEM baselines in iso-parameter and iso-FLOP settings. Routing analyses show semantic interpretability, dispatching polysemous tokens to distinct slots by context.
- Replaces deterministic memory lookup with gated mixture of M slots
- Evaluated on nanochat, Llama-3/MobileLLM, and Qwen3 backbones
- Beats Value Embedding, Bigram, and STEM baselines at matched compute
- Learned routing shows interpretable sense-based token dispatch
Full article175 words · extracted from huggingface.co · click to collapse
Scaling large language models efficiently has motivated sparse capacity mechanisms such as Mixture-of-Experts and, more recently, conditional memory: token-indexed embedding tables that augment the backbone with cheap parametric lookups. Existing memory-embedding methods retrieve via a deterministic function of the surface form, which collapses different contextual senses of the same token (e.g., python the language vs. the animal) into a single fixed entry. We introduce Mixture of Memory Embeddings (MoME), a context-aware memory mechanism that replaces each token's single memory row with a mixture of M slots and uses a learned gate over the hidden state to choose which slots to read at each position. In controlled pretraining experiments across nanochat, Llama-3/MobileLLM, and Qwen3 backbones, MoME improves over Value Embedding, Bigram, and STEM baselines in iso-parameter and iso-training-FLOP settings, shows a more promising memory-size scaling trend at sub-billion scale, and remains efficient in training and inference. Qualitative routing analyses on polysemous tokens further suggest that the learned mixture exhibits a degree of semantic interpretability, dispatching the same surface token to distinct memory slots under different senses.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.15126