Research map / Architectures and efficiency / Attention and transformers
Neural Execution Engines: Learning to Execute Subroutines
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use conditional masking driven by predicted pointers | Transformer sorting · Selection sort, sequences up to length 100 | accuracy | Helped |
| Use a 1D convolutional mask updater instead of positional transformer mask updates | NEE · Long-sequence mask updates | generalization | Helped |
| Use explicit positional encoding with a transformer encoder-decoder for mask updates | NEE mask updater · Unseen sequence positions | generalization | Hurt |
| Train on selection-sort and addition subroutines before composing graph algorithms (S1) | NEE · Mixed graph types, graph size 100 | accuracy: 92.00 -> 100.00 | Helped |
| Train NEE addition on fewer numbers | NEE addition · 8-bit addition with 64 training numbers | accuracy: 100.00 -> 96.53 | Hurt |
| Remove stronger residual connections (C1) (ablation) | all_mod transformer · Selection sort, sequence length 8, mixed test sets | accuracy: 99.00% -> 95.89% | Hurt |
| Remove MLP-based attention (C2) (ablation) | all_mod transformer · Selection sort, sequence length 8, mixed test sets | accuracy: 99.00% -> 97.56% | Hurt |
| Remove symmetrized MLP attention (C3) (ablation) | all_mod transformer · Selection sort, sequence length 8, mixed test sets | accuracy: 99.00% -> 98.33% | Hurt |
| Remove shared query, key, and value projections (C4) (ablation) | all_mod transformer · Selection sort, sequence length 8, mixed test sets | accuracy: 99.00% -> 98.44% | Hurt |
| Remove binary input encoding (C5) (ablation) | all_mod transformer · Selection sort, sequence length 8, mixed test sets | accuracy: 99.00% -> 89.56% | Hurt |
| Use binary input without a linear embedding (C6) (ablation) | all_mod transformer · Selection sort, sequence length 8, mixed test sets | accuracy: 99.00% -> 84.78% | Hurt |
| Add binary input encoding to the vanilla transformer (ablation) | Vanilla transformer · Selection sort, sequence length 8, mixed test sets | accuracy: 93.11% -> 96.67% | Helped |
Closest papers by meaning
- Learning to Add, Multiply, and Execute Algorithmic Instructions Exactly with Neural Networks (NeurIPS 2025 · similarity 0.87)
- Universal Length Generalization with Turing Programs (ICML 2025 · similarity 0.86)
- Slower Generalization, Faster Memorization: A Sweet Spot in Algorithmic Learning (NeurIPS 2026 · similarity 0.85)
- On the Markov Property of Neural Algorithmic Reasoning: Analyses and Methods (ICLR 2024 · similarity 0.85)
- Linear Transformers Implicitly Discover Unified Numerical Algorithms (NeurIPS 2025 · similarity 0.84)
- The Geometric Inductive Bias of Grokking: Bypassing Phase Transitions via Architectural Topology (NeurIPS 2026 · similarity 0.84)
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.