Research map / Large language models / Agents, code and math
Decoupling Understanding from Reasoning via Problem Space Mapping for Small-Scale Model Reasoning
Li Wang, Changhao Zhang, Zengqi Xiu, Kai Lu, Xin Yu, Kui Zhang, Wenjun Wu
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 implicit template constraint (ablation) | DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarks | accuracy: 41.25 -> 42.59 | Improved |
| add self-distillation (ablation) | DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarks | accuracy: 41.65 -> 42.59 | Improved |
| add SFT loss (ablation) | DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarks | accuracy: 40.14 -> 42.59 | Improved |
| add GRPO training (step III) (ablation) | DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarks | accuracy: 40.53 -> 42.59 | Improved |
| run a second DURIT iteration on GSM8K | DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarks | accuracy: 42.59 -> 42.95 | Improved |
| run a second DURIT iteration on GSM8K | DURIT, Llama3.2-1B-Instruct · average over 6 benchmarks | accuracy: 39.34 -> 40.03 | Improved |
| increase iterations from 1 to 3 | DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarks | accuracy: 42.59 -> 43.14 | Improved |
| double training steps in a single iteration instead of two iterations | DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarks | accuracy: 42.63 -> 42.95 | Hurt |
| map inputs with DURIT instead of PRewrite or original | Qwen2.5-0.5B-Instruct hidden states · GSM8K-Platinum | average 5NN distance: 75.16 -> 68.59 | Improved |
| use DURIT instead of base model | Llama3.2-1B-Instruct · GSM-Symbolic | relative accuracy drop (%): -23.7 -> -7.2 | Improved |
| use DURIT instead of base model | Qwen2.5-0.5B-Instruct · GSM-Symbolic | relative accuracy drop (%): -9.6 -> -11.3 | Hurt |
| distill from DeepSeek-R1 CoT (CoT-Dis) on LogiQA-trained data | Qwen2.5-1.5B-Instruct · average of benchmarks | accuracy: 52.60 -> 47.48 | Hurt |
Closest papers by meaning
- Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.90)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.89)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.89)
- Decompose, Analyze and Rethink: Solving Intricate Problems with Human-like Reasoning Cycle (NeurIPS 2024 · similarity 0.89)
- RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.88)
- Improving Reasoning Performance in Large Language Models via Representation Engineering (ICLR 2025 · similarity 0.88)
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.
