ArXiv Paper Tracker

按领域和关键词追踪 arXiv 新论文，两阶段漏斗筛选高价值论文。

Tool

python3 .claude/skills/arxiv-tracker/scripts/arxiv_ctl.py <command> [args]

Commands

# Run daily paper tracking (default: yesterday's papers)
python3 .claude/skills/arxiv-tracker/scripts/arxiv_ctl.py run
python3 .claude/skills/arxiv-tracker/scripts/arxiv_ctl.py run --date 2026-03-10

# List current topic configuration
python3 .claude/skills/arxiv-tracker/scripts/arxiv_ctl.py topics

# View tracking history
python3 .claude/skills/arxiv-tracker/scripts/arxiv_ctl.py history
python3 .claude/skills/arxiv-tracker/scripts/arxiv_ctl.py history --days 14

# Trigger keyword evolution analysis
python3 .claude/skills/arxiv-tracker/scripts/arxiv_ctl.py evolve

Architecture

Two-stage funnel:

1. Keyword pre-filter: arXiv API → keyword match on title + abstract (~200 → ~30 papers)
2. LLM deep evaluation: Gemini 3.1 Pro → 4-dimension score (novelty × 0.25 + rigor × 0.15 + relevance × 0.35 + collaboration × 0.25) → threshold ≥ 3.5

Output → Two Feishu documents:

• 业务日报: 平铺 + 多标签（无分类重复），含实操建议 + 趋势雷达
• 你可能感兴趣的: 个人兴趣论文（仅在有 personal 标记时生成）

3. 趋势雷达（双模型交叉分析）:
   • Round 1: Gemini Pro + Opus 并行独立分析
   • Round 2: 交叉校准（各看对方 R1 结论后修正）
   • Opus 综合渲染最终报告（共识/争议/知乎价值/趋势变化）
   • 持久化 trend_state.json：信号强化/衰减/淡出机制

完成后自动通过 Notifier 发送文档链接通知。

Configuration

Topic and keyword config: config/topics.yaml

• Topics are data-driven — change topics by editing YAML, no code changes needed
• Keyword evolution: automatic suggestions based on hit rate analysis

Schedule

Cron: 0 9 * * 2-6 (Tue-Sat 9:00 AM, covering Mon-Fri arXiv publications) Handler: arxiv:daily