A dataset capturing the structured intellectual synthesis behind high-quality AI research.
Sci-Reasoning is the first dataset designed to capture reasoning patterns behind AI breakthroughs in a structured format.
- 3,819 papers (999 Oral, 2,820 Spotlight) from NeurIPS, ICML, and ICLR (2023-2025)
- Structured Lineage Graphs capturing intellectual predecessors and relationships
- 15 Distinct Thinking Patterns identified through systematic analysis
sci-reasoning/
├── ml_paper_acquisition/ # Stage 1: Paper Collection
│ └── code/
│ ├── extract_papers.py
│ └── generate_report.py
│
├── prior_work_extraction/ # Stage 2: Lineage Tracing
│ └── code/
│ ├── prior_work_pipeline.py
│ └── batch_api_processor.py
│
├── thinking_patterns_llm_analysis/ # Stage 3: Pattern Analysis
│ └── code/
│ ├── pattern_analyzer.py
│ ├── classify_all.py
│ └── visualize.py
│
├── research_idea_evaluation/ # Stage 4: LLM Evaluation
│ └── code/
│ ├── evaluate_idea_generation_v4_exa_improved.py
│ ├── evaluate_claude_sonnet.py
│ └── evaluate_gemini_pro.py
│
└── paper.pdf
{
"target_paper": { "title": "...", "venue": "...", "year": 2024 },
"intellectual_predecessors": [
{
"title": "Predecessor title",
"role": "BASELINE|INSPIRATION|GAP_IDENTIFICATION|FOUNDATION",
"relationship_type": "EXTENDS|COMBINES_WITH|ADDRESSES_LIMITATION_OF"
}
],
"synthesis_narrative": "...",
"primary_pattern": "Gap-Driven Reframing",
"secondary_pattern": "Representation Shift"
}| Pattern | % | Description |
|---|---|---|
| Gap-Driven Reframing | 24.2 | Reframing problems to map onto better-suited methods |
| Cross-Domain Synthesis | 18.0 | Importing ideas from other fields |
| Representation Shift | 10.5 | Replacing primitives to simplify the problem |
@article{liu2025scireasoning,
title={Sci-Reasoning: A Dataset Decoding AI Innovation Patterns},
author={Liu, Jiachen and Harmon, Maestro and Zhang, Zechen},
year={2025}
}MIT License - see LICENSE for details.
