Research map / Large language models / Agents, code and math
How to transfer algorithmic reasoning knowledge to learn new algorithms?
Louis-Pascal A. C. Xhonneux, Andreea Deac, Petar Veličković, Jian Tang
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 |
|---|---|---|---|
| use NE++ (more expressive encoder) instead of NE (ablation) | NeuralExecutor · Dijkstra teacher forcing, 100 nodes | Predecessor error: 0.064 -> 0.522 | Hurt |
| use NE++ instead of NE (ablation) | NeuralExecutor · Most reliable teacher forcing | Next node error: 0.238 -> 0.174 | Improved |
| use NE++ instead of NE (ablation) | NeuralExecutor · Most reliable teacher forcing, 100 nodes | Predecessor error: 0.167 -> 0.171 | No clear effect |
| remove intermediate steps (no algorithm) vs teacher forcing (ablation) | NE · Dijkstra, 100 nodes | Key MSE: 4.79 -> 3.34e16 | Hurt |
| fine-tune pretrained Prim model | NE · Dijkstra no algorithm, 100 nodes | Predecessor error: 0.194 -> 0.388 | Hurt |
| freeze pretrained Prim processor | NE · Dijkstra no algorithm, 100 nodes | Predecessor error: 0.194 -> 0.607 | Hurt |
| 2-Processor transfer from Prim | NE · Dijkstra no algorithm, 100 nodes | Predecessor error: 0.194 -> 0.305 | Hurt |
| 2-Processor transfer from Prim | NE · Dijkstra no algorithm, 100 nodes | Key MSE: 3.34e16 -> 1.04e4 | Improved |
| multi-task training with Prim | NE · Dijkstra, 100 nodes | Key MSE: 3.34e16 -> 126 | Improved |
| multi-task training with Prim | NE++ · Dijkstra, 100 nodes | Predecessor error: 0.388 -> 0.282 | Improved |
| multi-task training with Widest | NE++ · Most reliable, 50 nodes | Predecessor error: 0.566 -> 0.185 | Improved |
| multi-task training with Widest | NE · Most reliable, 20 nodes | Predecessor error: 0.187 -> 0.452 | Hurt |
Closest papers by meaning
- Deep Equilibrium Algorithmic Reasoning (NeurIPS 2024 · similarity 0.87)
- Discrete Neural Algorithmic Reasoning (ICML 2025 · similarity 0.85)
- AlgoTrace: Algorithmic Primitives and Compositional Geometry of Reasoning in Language Models (ICML 2026 · similarity 0.85)
- Understanding Transformer Reasoning Capabilities via Graph Algorithms (NeurIPS 2024 · similarity 0.84)
- Primal-Dual Neural Algorithmic Reasoning (ICML 2025 · similarity 0.84)
- Open-Book Neural Algorithmic Reasoning (NeurIPS 2024 · 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.
