arxiv-reader

name	arxiv-reader
description	Read and analyze arXiv papers via HTML. Use when user mentions arxiv, paper, research paper, academic paper, arxiv ID, or asks to read/summarize/analyze a scientific paper.

arXiv Paper Reader

Read, summarize, and analyze academic papers from arXiv using Kiro's built-in web tools.

Scope

This skill handles: reading arXiv papers, extracting abstracts, summarizing sections, analyzing methodology, searching arXiv. Does NOT handle: PDF-only papers outside arXiv, citation graph analysis, bulk paper downloads.

When to Use

• User shares an arXiv URL or paper ID (e.g., 2301.07041)
• User asks to read, summarize, or analyze a research paper
• User wants to search arXiv for papers on a topic
• User asks about a specific paper's methodology, results, or contributions

Key Insight: ar5iv for Full Paper Reading

arXiv papers are LaTeX → PDF. Agent cannot read PDF directly via webFetch. Solution: Use ar5iv.labs.arxiv.org which renders LaTeX → HTML. Agent reads HTML perfectly.

URL Pattern Quick Reference

Need	URL Pattern	Tool
Abstract + metadata	https://arxiv.org/abs/{id}	webFetch (truncated)
Full paper as HTML	https://ar5iv.labs.arxiv.org/html/{id}	webFetch (selective/full)
Search papers	https://arxiv.org/search/?query={q}	remote_web_search preferred
API metadata (XML)	https://export.arxiv.org/api/query?id_list={id}	webFetch

See references/arxiv-url-patterns.md for full details.

Core Workflows

Workflow 1: Read a Specific Paper

Given an arXiv ID or URL:

1. Extract the paper ID from input (e.g., 2301.07041 from https://arxiv.org/abs/2301.07041 or 2301.07041v2)
2. Fetch abstract page: webFetch url=https://arxiv.org/abs/{id} mode=truncated Validate: Response contains title, authors, abstract text.
3. If user needs full paper content, fetch HTML version: webFetch url=https://ar5iv.labs.arxiv.org/html/{id} mode=selective, searchPhrase="{section keyword}" Validate: Response contains readable paper sections (not just navigation).
4. If selective mode returns insufficient content, retry with mode=full for complete paper.
5. Summarize findings based on user's request (overview, methodology, results, etc.)

Validate: Summary references specific sections, figures, or equations from the paper.

Workflow 2: Search for Papers on a Topic

1. Use remote_web_search with query: arxiv {topic} site:arxiv.org
2. Extract paper IDs from search results
3. For each relevant result, fetch abstract via webFetch url=https://arxiv.org/abs/{id} mode=truncated
4. Present papers with: title, authors, date, abstract summary, relevance note

Validate: Each paper listed has concrete metadata, not just URL.

Workflow 3: Deep Analysis of a Paper Section

1. Complete Workflow 1 steps 1-2 first (get overview)
2. Identify target section from user request (e.g., "methodology", "experiments", "related work")
3. Fetch with selective mode: webFetch url=https://ar5iv.labs.arxiv.org/html/{id} mode=selective searchPhrase="{section name or key term}"
4. If ar5iv returns empty/minimal content (some papers not yet converted), fallback:
   • Try webFetch url=https://arxiv.org/html/{id} mode=selective (arxiv's own HTML, newer papers)
   • Last resort: inform user that paper is PDF-only, suggest they attach the PDF directly

Validate: Extracted section contains substantive content, not just headers.

ar5iv Fallback Strategy

ar5iv returns empty/error?
├─ Paper too new (< few days old)?
│   └─ Try arxiv.org/html/{id} (native HTML, available for newer papers)
├─ Paper has complex LaTeX?
│   └─ ar5iv may fail on some papers. Try arxiv.org/html/{id}
└─ Neither works?
    └─ Inform user: "This paper is only available as PDF. 
       You can attach the PDF file directly in chat for me to analyze."

Anti-Patterns

Agent thinks	Reality
"I'll fetch the PDF URL"	webFetch cannot parse PDF binary. Always use HTML versions.
"I'll read the entire paper at once"	Full papers are huge. Use selective mode with searchPhrase first.
"ar5iv doesn't work, give up"	Try arxiv.org/html/ as fallback. Only then suggest PDF attachment.
"I'll summarize without reading"	NEVER summarize from title/abstract alone when user asks for deep analysis.

Output Format

When presenting paper analysis:

📄 **{Title}**
Authors: {authors} | Published: {date} | arXiv: {id}

**Abstract:** {1-2 sentence summary}

**Key Contributions:**
1. {contribution 1}
2. {contribution 2}

**Methodology:** {brief description}

**Results:** {key findings}

**Relevance:** {why this matters for user's context}

Related Skills

Situation	Activate Skill	Why
Paper describes a training method you want to reproduce	hf-transformers-trainer	Fine-tuning workflows with Trainer/TRL/PEFT
Paper references a HuggingFace model or dataset	hf-hub-datasets	Download models/datasets from HF Hub
Need to brainstorm which approach from the paper to adopt	ml-brainstorm	ML decision-making and tradeoff analysis

References

• arXiv URL Patterns — Load when: need to construct URLs or use arXiv API