Research map / Large language models / Agents, code and math
DESIGNER: Design-Logic-Guided Multidisciplinary Data Synthesis for LLM Reasoning
Weize Liu, Yongchi Zhao, Yijia Luo, Mingyu Xu, Jiaheng Liu, Yanan Li, Xiguo Hu, ZhiqiBai, Yuchi Xu, Wenbo Su, Bo Zheng
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 |
|---|---|---|---|
| SFT base model on DLR-Web+Book synthetic data vs official instruct model | Llama-3.1-8B · GPQA-Diamond | accuracy: 23.18 -> 65.45 | Improved |
| SFT Qwen3-8B-Base on DLR-Web+Book vs official thinking-mode model | Qwen3-8B · GPQA-Diamond | accuracy: 59.44 -> 71.01 | Improved |
| SFT Qwen3-4B-Base on DLR-Web only vs official thinking model | Qwen3-4B · GPQA-Diamond | accuracy: 54.70 -> 53.74 | Mixed |
| DLR data vs other synthetic datasets at equal size | Qwen3-8B-Base · GPQA-Diamond | accuracy: 56.67 -> 60.35 (NaturalReasoning vs DLR-Book) | Improved |
| combine Web and Book data | Llama-3.2-3B · GPQA-Diamond | accuracy: 22.42 (Web) / 38.33 (Book) -> 42.27 | Improved |
| use Design Logics instead of exemplar questions (ablation) | DESIGNER · GPQA-Diamond (304,181 questions) | accuracy: 58.89 -> 60.35 | Improved |
| use Design Logics instead of exemplar questions on full 3.04M corpus (ablation) | DESIGNER · GPQA-Diamond | accuracy: 64.10 -> 69.39 | Improved |
| add coarse retrieval ranking (ablation) | DESIGNER · MMLU-Pro | accuracy: 73.76 -> 74.98 | Improved |
| add fine LLM re-ranking (ablation) | DESIGNER · GPQA-Diamond | accuracy: 59.34 -> 60.35 | Improved |
| synthesize from book corpus instead of web corpus (ablation) | DESIGNER · multidisciplinary benchmarks | accuracy | Improved |
| scale synthesized data size (ablation) | DESIGNER · multidisciplinary benchmarks | accuracy | Improved |
Closest papers by meaning
- ORACLE: Optimizing Reasoning Abilities of Large Language Models via Constraint-Led Synthetic Data Elicitation (AAAI 2026 · similarity 0.90)
- Enhancing Reasoning Capabilities of LLMs via Principled Synthetic Logic Corpus (NeurIPS 2024 · similarity 0.89)
- Automated Creation of Reusable and Diverse Toolsets for Enhancing LLM Reasoning (AAAI 2025 · similarity 0.89)
- MindGYM: What Matters in Question Synthesis for Thinking-Centric Fine-Tuning? (NeurIPS 2025 · similarity 0.89)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.89)
- MindLoom: Composing Thought Modes for Frontier-Level Reasoning Data Synthesis (NeurIPS 2026 · similarity 0.89)
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.
