FEM-ASM: separating storage, execution and neural coordination in LLMs
A. Bochkov · hf · 2026-10-09
A new Hugging Face paper proposes FEM-ASM, an architecture that separates contextual computation, persistent storage, and exact execution in language models instead of routing everything through shared parameters:
- Inspired by the finite element method, independently built document states and deterministic executable skills contribute typed proposals to a shared LM state, reconciled by an explicit residual operator at common interface nodes.
- An attention-free Multi-Mesh prototype learns causal language modeling but isn't competitively capable.
- A versioned store holds 52,809 reconstructive memory elements (1.7B floating values) with 75% token reconstruction accuracy, made addressable via support-aware lexical indices under provenance-controlled queries.
- For executable arithmetic, positional result observations substantially improve neural rendering over a repeated global result vector; output substitutions shift the model's preferred answer.
The results support separating storage, execution, and neural coordination, while flagging open limits in question-only retrieval, unconstrained generation, and end-to-end efficiency.
More from Research
- Inherit-MAS cuts multi-agent token use by up to 34.6% with evolution-inspired inheritance — Songtao Wei · 2026-10-09
- Zero human labels: auto-generated soccer tracking dataset hits 62.5 HOTA, beating fine-tuned FairMOT — RexDouglass · 2026-10-09
- Why AI doesn't actually read words: from BPE subwords to byte-level models like BLT and H-Net — jbhuang0604 · 2026-10-09
- NVIDIA details HSTU recommender inference stack with up to 5.93x lower latency — PyTorch · 2026-10-09
- Preprint: LLMs store numbers as curves and helices, but compute comparisons differently — tweetsatpreet · 2026-10-09
- Best model was cheapest: open-weights model ran 669 clinical decisions for 1.7 cents — antoine_chaffin · 2026-10-09