Research map / Large language models / Instruction and fine-tuning
Decomposing Extrapolative Problem Solving: Spatial Transfer and Length Scaling with Map Worlds
Yao Tong, Jiayuan Ye, Anastasia Borovykh, Reza Shokri
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 |
|---|---|---|---|
| allocate budget to distinct questions (1 solution each) rather than 32 solutions per question (ablation) | SFT on shortest paths · Spatial transfer, low budget | success rate: 82% -> 94% | Improved |
| More Questions, High Coverage vs More Solutions (ablation) | Qwen2.5-7B-Instruct SFT · MathQA gain | accuracy: 0.72 -> 0.82 | Improved |
| More Questions, High Coverage vs baseline (ablation) | Qwen2.5-7B-Instruct SFT · MathQA physics | accuracy: 0.68 -> 0.77 | Improved |
| High Coverage instead of High Diversity (ablation) | Qwen2.5-7B-Instruct SFT · MathQA gain | accuracy: 0.74 -> 0.82 | Improved |
| High Coverage instead of High Diversity (ablation) | Qwen2.5-7B-Instruct SFT · MathQA physics | accuracy: 0.74 -> 0.77 | Improved |
| High Coverage instead of High Diversity (ablation) | Qwen2.5-7B-Instruct SFT · MathQA probability | accuracy: 0.792 vs 0.792 | No clear effect |
| increase coverage at 4% coverage via diversity increase (ablation) | SFT on shortest paths · Spatial transfer, coverage 4% | SR: 0.08 -> 0.29 | No clear effect |
| increase diversity at 32% coverage (ablation) | SFT on shortest paths · Spatial transfer, coverage 32% | SR: 0.47 -> 0.91 | Improved |
| increase coverage at diversity d = 1 (ablation) | SFT on shortest paths · Spatial transfer, d=1 | SR: 0.08 -> 0.78 | Improved |
| add ~1% training paths of slightly longer length (l=32,34) (ablation) | SFT on shortest paths · Length scaling to target length 30 | success rate: to nearly 90% | Improved |
| add shorter paths (l=22,24, 12% of data) (ablation) | SFT on shortest paths · Length scaling to target length 30 | success rate | No clear effect |
| add very long paths (l=80) (ablation) | SFT on shortest paths · Length scaling to target length 30 | success rate | Hurt |
Closest papers by meaning
- Attention as a Hypernetwork (ICLR 2025 · similarity 0.85)
- Horizon Generalization in Reinforcement Learning (ICLR 2025 · similarity 0.85)
- Compositional Reasoning in Language Models under Reinforcement Learning Post-Training (NeurIPS 2026 · similarity 0.85)
- Extrapolation by Association: Length Generalization Transfer In Transformers (NeurIPS 2025 · similarity 0.85)
- Memorize Theorems, Not Instances: Probing SFT Generalization through Mathematical Reasoning (NeurIPS 2026 · similarity 0.84)
- SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training (ICML 2025 · 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.
