Research map / Architectures and efficiency / RNNs and normalization
End-to-end Algorithm Synthesis with Recurrent Networks: Extrapolation without Overthinking
Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, Tom Goldstein
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 |
|---|---|---|---|
| add recall (input re-concatenated to features) to DT (ablation) | Deep Thinking net · Prefix sums 512-bit | peak accuracy: 0.00 \pm 0.00 -> 96.19 \pm 3.73 | Improved |
| add recall to DT (ablation) | Deep Thinking net · Mazes 59x59 | peak accuracy: 0.00 \pm 0.00 -> 82.72 \pm 15.14 | Improved |
| add progressive loss (alpha=0.01) to DT-Recall (ablation) | DT-Recall · Mazes 59x59 | peak accuracy: 82.72 \pm 15.14 -> 97.30 \pm 0.68 | Improved |
| add progressive loss (alpha=1.0) to DT-Recall (ablation) | DT-Recall · Prefix sums 512-bit | peak accuracy: 96.19 \pm 3.73 -> 97.12 \pm 1.88 | Improved |
| remove progressive loss from DT-Recall (ablation) | DT-Recall · Chess puzzles | average accuracy: 83 -> 97 | Improved |
| add progressive loss (alpha=0.5) to DT-Recall (ablation) | DT-Recall · Chess puzzles 600K-700K | peak accuracy: 82.12 \pm 0.69 -> 82.69 \pm 0.27 | Improved |
| randomly sample n instead of fixing n=0 in progressive loss (ablation) | DT-Recall · Prefix sums 512-bit | accuracy: 90.27 -> 97.12 | Improved |
| add recall to DT (ablation) | Deep Thinking net · Chess puzzles 600K-700K | peak accuracy: 78.81 \pm 1.04 -> 82.12 \pm 0.69 | Improved |
| add progressive loss (alpha=1.0) to DT without recall (ablation) | DT · Prefix sums 512-bit | peak accuracy: 0.00 \pm 0.00 -> 11.26 \pm 6.90 | Improved |
| add progressive loss (alpha=0.01) to DT without recall (ablation) | DT · Mazes 13x13 | peak accuracy: 85.59 \pm 2.81 -> 86.08 \pm 3.96 | No clear effect |
| add progressive loss to DT-Recall (ablation) | DT-Recall · Mazes 13x13 | peak accuracy: 99.94 \pm 0.03 -> 99.88 \pm 0.05 | No clear effect |
| replace progressive loss with run-time weighted loss (ablation) | DT-Recall · Prefix sums 512-bit | accuracy: 97.12 \pm 1.88 -> 72.45 \pm 8.91 | Hurt |
Closest papers by meaning
- On the Markov Property of Neural Algorithmic Reasoning: Analyses and Methods (ICLR 2024 · similarity 0.87)
- Boosting Inference with Guided Reasoning: Stochastic Exploration for Recursive Models (NeurIPS 2026 · similarity 0.86)
- NeuralSolver: Learning Algorithms For Consistent and Efficient Extrapolation Across General Tasks (NeurIPS 2024 · similarity 0.86)
- Reasoning Cache: Continual Improvement Over Long Horizons via Short-Horizon RL (ICML 2026 · similarity 0.86)
- Let LRMs Break Free from Overthinking via Self-Braking Tuning (NeurIPS 2025 · similarity 0.86)
- The Overthinker's DIET: Cutting Token Calories with DIfficulty-AwarE Training (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.
