Iterative top-venue paper tracking and research insight generation system. Input a research topic, and Paper Radar automatically collects, analyzes, clusters, and generates trend reports with actionable research ideas.
- Automated paper collection from arXiv + Semantic Scholar with top-venue and high-velocity filtering
- Manual paper addition via arXiv IDs or local PDF upload
- PDF parsing with intelligent section extraction (Abstract, Method, Conclusion, etc.)
- Claude-powered structured analysis of each paper (Haiku for extraction, Sonnet for synthesis)
- UMAP + HDBSCAN clustering to identify research sub-directions
- Trend analysis with emerging/hot/declining/stable classification
- Blind spot discovery from aggregated limitations and future work
- Research idea generation with novelty and feasibility assessment
- Idea refinement with targeted literature search and deepened analysis
- Interactive Web UI with D3.js cluster visualization, real-time SSE progress
- Markdown report generation with auto-push to GitHub
# 1. Clone the repo
git clone <your-repo-url> && cd paper-radar
# 2. Install all dependencies (Python + Node)
./start.sh install
# 3. Set API keys
export ANTHROPIC_API_KEY="your-key"
export OPENAI_API_KEY="your-key" # optional, for embeddings; 不设则自动用本地模型
# 4. Run the full pipeline
./start.sh run --topic "offline reinforcement learning"./start.sh install # 安装 Python + Node 依赖（首次运行）
./start.sh run # 使用 config.yaml 默认配置运行 pipeline
./start.sh run --topic "LLM reasoning" # 指定研究主题
./start.sh run --months 3 # 指定时间范围
./start.sh run --papers 2401.12345,2402.00001 # 手动添加论文
./start.sh web # 启动 Web UI（前后端同时启动）
./start.sh api # 只启动后端 API
./start.sh refine 1 # 深化指定 Idea
./start.sh help # 查看帮助./start.sh web启动后访问：
- 前端 UI：
http://localhost:3000 - 后端 API：
http://localhost:8000 - 按
Ctrl+C同时关闭前后端
pip install -r requirements.txt
python run.py --topic "reinforcement learning" # 运行 pipeline
python run.py --refine 1 # 深化 Idea
python web/app.py # 启动后端All settings are in config.yaml:
| Field | Type | Default | Description |
|---|---|---|---|
topic |
string | "reinforcement learning" |
Research topic to analyze |
time_range_months |
int | 6 |
How far back to search |
anthropic_api_key |
string | "" |
Anthropic API key (env var preferred) |
openai_api_key |
string | "" |
OpenAI API key for embeddings |
models.extraction |
string | "claude-haiku-4-5" |
Model for paper parsing |
models.analysis |
string | "claude-sonnet-4-5" |
Model for trend/idea generation |
models.refinement |
string | "claude-sonnet-4-5" |
Model for idea deepening |
embedding.provider |
string | "openai" |
openai or local |
embedding.model |
string | "text-embedding-3-small" |
OpenAI embedding model |
embedding.local_model |
string | "all-MiniLM-L6-v2" |
Local sentence-transformers model |
collection.max_papers_per_keyword |
int | 50 |
Max papers per search keyword |
collection.citation_velocity_threshold |
float | 0.5 |
Min citations/day for velocity filter |
collection.max_papers_to_analyze |
int | 50 |
Max papers entering full analysis |
refinement.max_papers |
int | 20 |
Max papers in targeted refinement search |
git.reports_repo_path |
string | "" |
Path to report repo (empty = skip push) |
| Input | Notes |
|---|---|
reinforcement learning |
Broad sub-field, auto-expands keywords |
offline reinforcement learning |
Specific direction, more precise results |
LLM reasoning and planning |
Multi-word combo, covers cross-cutting work |
diffusion models for robotics |
Cross-domain combo, great for finding blind spots |
| Input | Problem |
|---|---|
AI |
Too broad, paper count explodes |
MAPPO with communication constraints |
Too narrow, insufficient papers |
Choose a "research sub-direction" level, e.g., multi-agent reinforcement learning is better than reinforcement learning, but MAPPO with communication constraints is too specific.
Paper Radar collects papers meeting any of these criteria:
1. Top Venue Papers
Verified via Semantic Scholar's venue field against built-in whitelist (NeurIPS, ICML, ICLR, ACL, CVPR, etc.).
2. arXiv High-Velocity Papers For papers published within 180 days, normalized citation velocity is computed:
citation_velocity = citationCount / max(days_since_published, 30)
Papers with velocity >= threshold (default 0.5, i.e., ~1 citation every 2 days) are included even without top-venue publication.
3. Manually Specified Papers Added via arXiv ID or PDF upload, bypassing all filters.
Adjust citation_velocity_threshold and venue lists in config.yaml.
All Claude prompts are defined in prompts.yaml — no prompts are hardcoded. You can modify system/user prompts to customize behavior:
| Prompt | Purpose | Model |
|---|---|---|
keyword_expansion |
Topic to search keywords + venue category | Haiku |
paper_extraction |
Paper sections to structured JSON | Haiku |
cluster_naming |
Paper cluster to name + summary | Haiku |
trend_analysis |
Time data to trend report | Sonnet |
blind_spot_discovery |
Limitations to blind spots | Sonnet |
idea_generation |
Blind spots + methods to ideas | Sonnet |
report_generation |
All data to Markdown report | Sonnet |
idea_keyword_extraction |
Idea to targeted search keywords | Haiku |
idea_deepening |
Idea + new papers to refined plan | Sonnet |
Default whitelist by category (auto-selected based on topic):
| Category | Venues |
|---|---|
| Machine Learning | NeurIPS, ICML, ICLR, JMLR, TMLR |
| NLP | ACL, EMNLP, NAACL, TACL, EACL |
| Computer Vision | CVPR, ICCV, ECCV, TPAMI, IJCV |
| Robotics | CoRL, ICRA, RSS, IROS |
| AI General | AAAI, IJCAI, UAI |
Customize by editing the venues section in config.yaml.
Reports are saved to output/reports/{topic}_{YYYY-MM}.md with this structure:
# {Topic} Research Landscape Report
## Key Takeaways
## Overview
## Sub-fields & Clusters
## Trend Analysis
## Research Blind Spots
## Research Ideas
## Notable Papers
## Methodology Note
Refinement reports are appended as ## Idea Refinement: {title} sections.
Per run (assuming 30 papers analyzed):
| Component | API Calls | Estimated Cost |
|---|---|---|
| Keyword expansion | 1x Haiku | ~$0.01 |
| Paper extraction | 30x Haiku | ~$0.30 |
| Cluster naming | 3-5x Haiku | ~$0.05 |
| Trend analysis | 1x Sonnet | ~$0.10 |
| Blind spot discovery | 1x Sonnet | ~$0.10 |
| Idea generation | 1x Sonnet | ~$0.10 |
| Report generation | 1x Sonnet | ~$0.15 |
| Embeddings (OpenAI) | 30 texts | ~$0.01 |
| Total | ~$0.80 |
Idea refinement adds ~$0.30-0.50 per idea.
- Multi-topic comparison reports
- Citation graph analysis
- Automated weekly/monthly scheduling
- Paper recommendation based on user history
- Integration with Zotero/Mendeley
- Email/Slack notifications for new high-impact papers
- Fine-grained PDF parsing with layout analysis
- Collaborative annotations and comments