Research map / Architectures and efficiency / RNNs and normalization
A Primer for Neural Arithmetic Logic Modules
Bhumika Mistry, Katayoun Farrahi, Jonathon Hare
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace NALU with NAU for addition | NALU · Single Module Arithmetic Task, addition | success rate | Improved |
| use iNALU instead of NALU for addition | NALU · Single Module Arithmetic Task, addition | success rate | Improved |
| use iNALU instead of NALU for multiplication | NALU · Single Module Arithmetic Task, multiplication | success rate | Improved |
| use G-NALU instead of NALU | NALU · Single Module Arithmetic Task, subtraction and multiplication | success rate | No clear effect |
| add imaginary weight matrix (NPU vs RealNPU) | RealNPU · Single Module Arithmetic Task, multiplication U[-2,2) | success rate: 32% more success | Improved |
| NPU vs RealNPU on remaining multiplication ranges | RealNPU · Single Module Arithmetic Task, multiplication | success rate | No clear effect |
| use RealNPU instead of NPU on negative input ranges | NPU · Single Module Arithmetic Task, division | success rate | Improved |
| add 8 redundant inputs (10 inputs total) | iNALU and NAU · Single Module Arithmetic Task, addition | success rate | Hurt |
| add 98 redundant inputs (100 inputs total) | NAU · Single Module Arithmetic Task, subtraction | success rate | Hurt |
| remove relevance gating from NPU (NaiveNPU) (ablation) | NPU · Single Module Arithmetic Task (prior work) | sparsity | Hurt |
| stack two NMUs instead of NAU-NMU | NMU · Two-layer arithmetic task | success rate: 56% subtraction, 0% addition | Hurt |
| increase NLRL intermediary units per layer | NLRL · Logic tasks | performance: saturates at 8 | Improved |
Closest papers by meaning
- Neural Arithmetic Logic Units (NeurIPS 2018 · similarity 0.85)
- Neural Logic Machines (ICLR 2019 · similarity 0.84)
- Arithmetic Without Algorithms: Language Models Solve Math with a Bag of Heuristics (ICLR 2025 · similarity 0.83)
- KLay: Accelerating Arithmetic Circuits for Neurosymbolic AI (ICLR 2025 · similarity 0.83)
- Mastering Symbolic Operations: Augmenting Language Models with Compiled Neural Networks (ICLR 2024 · similarity 0.82)
- LogicMP: A Neuro-symbolic Approach for Encoding First-order Logic Constraints (ICLR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
