Loading ToolBench
ArXiv paper search MCP server for searching academic papers with filtering by date, keywords, field, and direct arxiv ID lookup
Single tool 'papersearch' has a complete input schema with proper JSON Schema structure, but suffers from significant documentation and usability gaps. The tool name lacks a clear action verb, description is present but minimal, and parameter descriptions are inconsistent in detail. Output schema is not documented. The implementation shows functional arxiv integration but lacks production-grade error handling and LLM-friendly design patterns.
Arxiv 论文搜索
Tool name 'papersearch' lacks action verb. Should be 'search_papers' or 'search_arxiv' to clearly signal intent to LLMs.
No output schema documented. LLMs cannot determine what fields to expect in the response (title, authors, url, etc.) for downstream planning.
Parameter 'query_type' description is ambiguous. Conflates two use cases: 'search type (default moe)' vs 'arxiv ID (e.g. 2103.03404)'. Should be split or clearly documented as 'Either a search strategy (moe, moe inference) OR an arxiv ID (format: YYYY.NNNNN)'.
Parameter descriptions use inconsistent detail levels. 'days' and 'max_results' lack bounds documentation (e.g., '1-365 days', '1-1000 results'). LLMs may pass absurd values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 40 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 23 | - | v1 |
Error handling returns generic text ('出现错误: {str(e)}') without actionable guidance. LLMs cannot determine if the error is retryable, user-fixable, or fatal.
Default values for 'days' (180) and 'max_results' (100) are large. Searching 180 days of arxiv and returning 100 results risks token exhaustion. Should default to 30 days and 20 results.
Tool description 'Arxiv 论文搜索' is in Chinese and minimal (4 words). Should be in English and explain WHEN to use it vs other tools, WHAT it returns, and any prerequisites. E.g., 'Search arXiv preprints by keyword, category, or ID. Returns title, authors, abstract, publication date, and PDF link. Useful for literature review, citation tracking, and recent research discovery.'
Response formatting embeds pagination, sorting, and display logic (format_papers). No structured output schema. LLMs receive formatted text, not JSON, harder to extract fields and chain downstream tools.