Research
Deep Fusion Memory proposes chunk-level parametric memory for LLMs
DFM predicts memory from context and fuses it into hidden states via gated cross-attention with a frozen backbone, needing 32x fewer retrieval queries for training data. Reported QA gains reach 12.71% on Mistral-7B-v0.3 and 11.19% on Qwen3-8B-Base.
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
Deep Fusion Memory: A Chunk-Level Parametric Memory for Large Language Models