Paper Scout (arXiv pipeline)
Built-in tools implement a three-step workflow inspired by PaperFlow (arXiv query + metadata-style scoring) and PaperBrain (LLM screening of title/abstract + narrative digest).
Tools
| Tool |
Purpose |
arxiv_fetch_range |
Crawl submittedDate in [start_date, end_date], optional keywords (OR in title/abstract), categories (default cs.AI, cs.LG, cs.CL, math.OC). Returns JSON papers. |
arxiv_rank_papers |
Enhanced metadata score every paper; optionally enriches the top pool with Semantic Scholar / OpenAlex signals and optionally runs one batch LLM call for shortlist scoring; returns selected (top N). |
arxiv_paper_digest |
Introduction report (Markdown) for the selected list; uses metadata, bibliometric cues, and LLM rationale when available. |
Recommended workflow
Fetch
Call arxiv_fetch_range with start_date, end_date, keywords, and optionally categories / max_results.
Pass the returned JSON (whole object) into the next step.
Rank
Call arxiv_rank_papers with:
papers_json: the string from step 1 (or json.dumps of the object).interest_keywords: what “high value” means for the user (used for scoring and LLM).top_n: how many papers to keep.use_external_ranking: defaults to true; uses Semantic Scholar first and OpenAlex fallback for stronger selection.use_llm_refinement: true for stronger selection (needs OpenRouter / VLM configured for ClawPhD).
Digest
Call arxiv_paper_digest with:
selected_papers_json: json.dumps of the selected array from step 2, or the full rank output object.interest_keywords: same as step 2.language: zh or en.
Notes
- No
arxiv Python package is required; queries use the public Atom API via httpx.
- External ranking uses Semantic Scholar first and OpenAlex as fallback. It works without a key, but if you have one you can pass
semantic_scholar_api_key or set SEMANTIC_SCHOLAR_API_KEY / S2_API_KEY.
- LLM refinement and rich digest need a configured multimodal/text provider (same stack as diagram tools — typically OpenRouter).
- For local cached papers and interactive CLI workflows, still use the
arxivterminal skill (arxiv fetch, scripts/arxiv_search.py). All papers are stored under paper_library/{subject}/{date}/ in the workspace root, where {subject} is a snake_case topic slug and {date} is YYYY-MM-DD.
Example (conceptual)
arxiv_fetch_range(start_date="2026-03-01", end_date="2026-03-10", keywords=["world model", "VLA"], categories=["cs.RO","cs.LG"])
→ arxiv_rank_papers(papers_json=<output>, interest_keywords=["world model","VLA","embodied AI"], top_n=5, use_external_ranking=true, use_llm_refinement=true)
→ arxiv_paper_digest(selected_papers_json=<selected>, interest_keywords=[...], language="zh")
1---2name: paper-scout3description: Fetch arXiv papers by date range and topics, rank them for research value, and produce introduction digests. Use when the user wants a literature sweep, daily or weekly paper triage, or a written overview of the best papers in a niche — without relying on the arxivterminal local database.4---56# Paper Scout (arXiv pipeline)78Built-in **tools** implement a three-step workflow inspired by **PaperFlow** (arXiv query + metadata-style scoring) and **PaperBrain** (LLM screening of title/abstract + narrative digest).910## Tools1112| Tool | Purpose |13|------|---------|14| `arxiv_fetch_range` | Crawl **submittedDate** in `[start_date, end_date]`, optional **keywords** (OR in title/abstract), **categories** (default cs.AI, cs.LG, cs.CL, math.OC). Returns JSON `papers`. |15| `arxiv_rank_papers` | **Enhanced metadata score** every paper; optionally enriches the top pool with **Semantic Scholar / OpenAlex** signals and optionally runs **one batch LLM call** for shortlist scoring; returns `selected` (top N). |16| `arxiv_paper_digest` | **Introduction report** (Markdown) for the `selected` list; uses metadata, bibliometric cues, and LLM rationale when available. |1718## Recommended workflow19201. **Fetch** 21 Call `arxiv_fetch_range` with `start_date`, `end_date`, `keywords`, and optionally `categories` / `max_results`. 22 Pass the returned JSON (whole object) into the next step.23242. **Rank** 25 Call `arxiv_rank_papers` with:26 - `papers_json`: the **string** from step 1 (or `json.dumps` of the object).27 - `interest_keywords`: what “high value” means for the user (used for scoring and LLM).28 - `top_n`: how many papers to keep.29 - `use_external_ranking`: defaults to `true`; uses **Semantic Scholar first** and **OpenAlex fallback** for stronger selection.30 - `use_llm_refinement`: `true` for stronger selection (needs **OpenRouter / VLM** configured for ClawPhD).31323. **Digest** 33 Call `arxiv_paper_digest` with:34 - `selected_papers_json`: `json.dumps` of the `selected` array from step 2, or the full rank output object.35 - `interest_keywords`: same as step 2.36 - `language`: `zh` or `en`.3738## Notes3940- **No `arxiv` Python package** is required; queries use the public **Atom API** via `httpx`.41- **External ranking** uses **Semantic Scholar** first and **OpenAlex** as fallback. It works without a key, but if you have one you can pass `semantic_scholar_api_key` or set `SEMANTIC_SCHOLAR_API_KEY` / `S2_API_KEY`.42- **LLM refinement** and **rich digest** need a configured multimodal/text provider (same stack as diagram tools — typically OpenRouter).43- For **local cached papers** and interactive CLI workflows, still use the **`arxivterminal`** skill (`arxiv fetch`, `scripts/arxiv_search.py`). All papers are stored under `paper_library/{subject}/{date}/` in the workspace root, where `{subject}` is a snake_case topic slug and `{date}` is `YYYY-MM-DD`.4445## Example (conceptual)4647```text48arxiv_fetch_range(start_date="2026-03-01", end_date="2026-03-10", keywords=["world model", "VLA"], categories=["cs.RO","cs.LG"])49→ arxiv_rank_papers(papers_json=<output>, interest_keywords=["world model","VLA","embodied AI"], top_n=5, use_external_ranking=true, use_llm_refinement=true)50→ arxiv_paper_digest(selected_papers_json=<selected>, interest_keywords=[...], language="zh")51```
zhihaoairobotic