AI Paper Atlas

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

AAAI 2026 · Large language models › Agents, code and math · approach: Mathematical and structural reasoning benchmarks · 1 citations (OpenAlex)

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 changedWhereMeasuredEffect
add implicit template constraint (ablation)DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarksaccuracy: 41.25 -> 42.59Improved
add self-distillation (ablation)DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarksaccuracy: 41.65 -> 42.59Improved
add SFT loss (ablation)DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarksaccuracy: 40.14 -> 42.59Improved
add GRPO training (step III) (ablation)DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarksaccuracy: 40.53 -> 42.59Improved
run a second DURIT iteration on GSM8KDURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarksaccuracy: 42.59 -> 42.95Improved
run a second DURIT iteration on GSM8KDURIT, Llama3.2-1B-Instruct · average over 6 benchmarksaccuracy: 39.34 -> 40.03Improved
increase iterations from 1 to 3DURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarksaccuracy: 42.59 -> 43.14Improved
double training steps in a single iteration instead of two iterationsDURIT, Qwen2.5-0.5B-Instruct · average over 6 benchmarksaccuracy: 42.63 -> 42.95Hurt
map inputs with DURIT instead of PRewrite or originalQwen2.5-0.5B-Instruct hidden states · GSM8K-Platinumaverage 5NN distance: 75.16 -> 68.59Improved
use DURIT instead of base modelLlama3.2-1B-Instruct · GSM-Symbolicrelative accuracy drop (%): -23.7 -> -7.2Improved
use DURIT instead of base modelQwen2.5-0.5B-Instruct · GSM-Symbolicrelative accuracy drop (%): -9.6 -> -11.3Hurt
distill from DeepSeek-R1 CoT (CoT-Dis) on LogiQA-trained dataQwen2.5-1.5B-Instruct · average of benchmarksaccuracy: 52.60 -> 47.48Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy