Keyword frequency and topic trend analysis of 5,356 accepted papers at ICLR 2026, based on author-supplied keywords and OpenReview primary areas.
📊 View Charts | 📋 Data | 🔗 GitHub
Keyword frequency word cloud
Thematic grouping of related keywords into 16 major research tracks. Area size represents total keyword hits within each group.
| Rank | Topic | Keyword Hits | Key Signals |
|---|---|---|---|
| 1 | 🔥 LLM (General) | 998 | Still dominant but shifting toward depth over architecture |
| 2 | 🚀 Reinforcement Learning & RLHF | 552 | RLHF / preference learning is the hottest sub-area |
| 3 | 🔥 Diffusion & Flow Models | 409 | Flow matching is replacing DDPM as the default |
| 4 | 📈 Multimodal / VLM | 331 | Moving from modality stitching to native unified intelligence |
| 5 | 📈 Alignment / Safety | 243 | From academic curiosity to engineering necessity |
| 6 | 🔥 Reasoning / Test-Time Scaling | 218 | The hottest incremental direction — near-zero two years ago |
| 7 | ⚡ Efficient Training & Inference | 195 | Quantization, speculative decoding, MoE |
| 8 | 🧠 Interpretability | 169 | Mechanistic interpretability becomes a standalone track |
| 9 | 🚀 Agents / Tool Use | 152 | From novel concept to stable track |
| 10 | 🌍 World Models / Video | 104 | Still emerging (~2% coverage) — not yet red ocean |
| Keyword | Enrichment |
|---|---|
| mamba | 6.84x |
| pre-training | 5.70x |
| optimal control | 5.70x |
| expressivity | 5.70x |
| scaling laws | 4.80x |
| agents | 4.34x |
| generative modeling | 4.34x |
| optimal transport | 4.15x |
→ Methodological depth beats application breadth with reviewers.
Keywords plotted by total appearances (y-axis) vs. oral enrichment ratio (x-axis). Bubble size represents frequency. Higher enrichment indicates reviewer preference for oral presentation.
Key insight: Keywords in the upper-right quadrant (high frequency + high enrichment) represent both popular and prestigious research directions — scaling laws, agents, and generative modeling stand out.
The most frequent keywords across all 5,356 papers. LLM-related terms dominate, followed by reinforcement learning and diffusion/flow models.
iclr2026-research-trends/
├── README.md ← You are here
├── LICENSE ← MIT
├── data/
│ ├── iclr2026_keywords.csv ← Keyword frequency table (top 200)
│ ├── iclr2026_areas.csv ← Primary area distribution
│ ├── iclr2026_oral_enrichment.csv ← Oral vs Poster keyword enrichment
│ └── iclr2026_trend_groups.json ← Thematic group definitions + counts
├── scripts/
│ ├── analyze.py ← Main analysis script (keyword extraction + grouping)
│ └── visualize.py ← Generate charts (matplotlib + AntV)
└── charts/
├── top25_keywords.png ← Top 25 keyword bar chart
├── trend_groups.png ← Thematic group treemap
├── area_distribution.png ← Primary area donut chart
├── oral_enrichment.png ← Oral enrichment bubble chart
└── keyword_wordcloud.png ← Keyword frequency word cloud
- Raw dataset: DmytroLopushanskyy/iclr2026-affiliations — 5,356 ICLR 2026 accepted papers with PDF-derived affiliations, author-supplied keywords, and OpenReview primary areas.
- Conference: ICLR 2026, 5,356 accepted papers out of ~15,700 submissions (34.1% acceptance rate).
# 1. Clone this repo + the data source
git clone https://github.com/ZTCY/iclr2026-research-trends.git
cd iclr2026-research-trends
git clone --depth 1 https://github.com/DmytroLopushanskyy/iclr2026-affiliations.git _source
# 2. Install dependencies
pip install pandas matplotlib seaborn squarify adjustText wordcloud
# 3. Run analysis
python scripts/analyze.py
# 4. Generate charts
python scripts/visualize.py- Keyword extraction: Author-supplied keywords from the ICLR 2026 submission form (semicolon-separated), lowercased and stripped.
- Thematic grouping: Related keywords manually grouped into 16 thematic clusters (e.g., "LLM" merges "large language models", "llm", "llms", etc.).
- Oral enrichment: For each keyword, computed the ratio of its appearance rate in Oral papers vs. Poster papers. Higher enrichment → reviewers value this direction more.
- Visualization: Charts generated using matplotlib (bubble chart) and AntV API (treemap, donut, bar chart, word cloud) for publication-quality output.
If you use this analysis, please cite both this repository and the original data source:
@misc{iclr2026-research-trends,
title={ICLR 2026 Research Trends: Keyword Frequency and Topic Analysis},
author={ZTCY},
year={2026},
url={https://github.com/ZTCY/iclr2026-research-trends}
}
@misc{iclr2026-affiliations,
title={ICLR 2026 Institutional Affiliations Dataset},
author={Lopushanskyy, Dmytro},
year={2026},
url={https://github.com/DmytroLopushanskyy/iclr2026-affiliations}
}MIT — see LICENSE.