Research map / Large language models / Language, safety and interpretability
Systematic Generalization: What Is Required and Can It Be Learned?
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 a tree layout rather than a chain layout (ablation) | NMN · SQOOP, especially #rhs/lhs=1 | systematic generalization | Helped |
| Add shortcut connections from image features to all chain modules (ablation) | NMN-Chain · SQOOP | systematic generalization | Helped |
| Use soft-attention learned module parametrization instead of hard-coded word assignments (ablation) | NMN-Tree (Find) · SQOOP #rhs/lhs=1 | test error rate (%): 27.19 ± 16.02 | Hurt |
| Initialize induced layout with p0(tree)=0.1 rather than a high tree probability (ablation) | Stochastic N2NMN (Residual) · SQOOP #rhs/lhs=1 | test error rate (%): 0.16 ± 0.11 -> 31.89 ± 0.75 | Hurt |
| Initialize induced layout with p0(tree)=0.1 rather than a high tree probability (ablation) | Stochastic N2NMN (Find) · SQOOP #rhs/lhs=1 | test error rate (%): 0.41 ± 0.07 -> 47.54 ± 0.95 | Hurt |
| Increase MAC units from 12 to 24 (ablation) | MAC · SQOOP #rhs/lhs=1 | test error rate (%): 13.67 ± 9.97 -> 9.11 ± 7.67 | Helped |
| Reduce MAC units from 12 to 1 (ablation) | MAC · SQOOP #rhs/lhs=1 | test error rate (%): 13.67 ± 9.97 -> 28.67 ± 1.91 | Hurt |
| Set MAC dimensionality to 256 instead of 128 (ablation) | MAC · SQOOP #rhs/lhs=1 | test error rate (%): 13.67 ± 9.97 -> 4.62 ± 5.07 | Helped |
| Set MAC dimensionality to 64 instead of 128 (ablation) | MAC · SQOOP #rhs/lhs=1 | test error rate (%): 13.67 ± 9.97 -> 23.61 ± 6.27 | Hurt |
| Set MAC dimensionality to 512 instead of 128 (ablation) | MAC · SQOOP #rhs/lhs=1 | test error rate (%): 13.67 ± 9.97 -> 8.37 ± 7.45 | Helped |
| Add weight decay 0.00001 (ablation) | MAC · SQOOP #rhs/lhs=1 | test error rate (%): 13.67 ± 9.97 -> 19.21 ± 9.27 | Hurt |
| Add weight decay 0.0001 (ablation) | MAC · SQOOP #rhs/lhs=1 | test error rate (%): 13.67 ± 9.97 -> 31.19 ± 0.87 | Hurt |
Closest papers by meaning
- Delving into the Reversal Curse: How Far Can Large Language Models Generalize? (NeurIPS 2024 · similarity 0.88)
- Nexus: Same Pretraining Loss, Better Downstream Generalization via Common Minima (NeurIPS 2026 · similarity 0.87)
- Quantifying Generalization Complexity for Large Language Models (ICLR 2025 · similarity 0.87)
- Generalization v.s. Memorization: Tracing Language Models’ Capabilities Back to Pretraining Data (ICLR 2025 · similarity 0.87)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.87)
- Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning (NeurIPS 2025 · similarity 0.86)
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.