kkkirito-123/zlagent/paper-search

paper-search

Trigger

Use this skill when the user asks for a small ad-hoc list of academic papers, typically:

• "给我目标检测领域的 5 篇顶刊论文，要最新的"
• "AI agent 方向最近半年的 SOTA 论文有哪些"
• "推荐 3 篇关于 RAG 评估的综述"
• "Find me 5 recent papers on diffusion video generation"

Do NOT use this skill for:

• Scheduled daily / weekly digests → daily-paper-digest.
• "Today's cs.AI listing on arxiv" (category dump, not topic-targeted)

→ arxiv.

• Reading / summarising a single paper the user already linked →

web-summary + read_url.

Inputs

Resolve from the user message before fetching:

• topic — the research area in plain words ("目标检测 / object

detection", "RAG evaluation"). Required; ask one short follow-up if truly ambiguous.

• count — how many papers, default 5, hard cap 8.
• recency — "最新 / latest" → last 12 months; "近年 / recent" → last

3 years; explicit window like "2024 以来" → respect verbatim.

• kind — empirical / SOTA / survey / benchmark / methods. Default:

whichever the user phrased; if silent, prefer SOTA + 1 survey when both are obviously useful.

Steps

1. Source priority (cheapest + highest signal first):

• arxiv API — https://export.arxiv.org/api/query?search_query=ti:<topic>+OR+abs:<topic>&start=0&max_results=20&sortBy=submittedDate&sortOrder=descending

via read_url. Returns Atom XML; parse for title / authors / summary / arxiv id. Best for ML / CS / EE.

• Semantic Scholar Graph API — https://api.semanticscholar.org/graph/v1/paper/search?query=<topic>&limit=20&fields=title,authors,year,venue,citationCount,abstract,url

via read_url. Better venue + citation metadata than arxiv. If it returns 429, wait and retry once with a 60s gap; then move on without it.

• web_search — only as a last fallback when arxiv + S2 both

yielded fewer than count candidates. Query template: "<topic>" survey 2024 site:arxiv.org then site:openreview.net.

1. Filter + rank:

• Drop preprints whose abstract is empty / "Withdrawn".
• Prefer top-tier venues when the user asked for 顶刊 / 顶会 (CV:

CVPR / ICCV / ECCV / TPAMI; ML: NeurIPS / ICML / ICLR; NLP: ACL / EMNLP / NAACL; speech: INTERSPEECH; systems: SOSP / OSDI; security: S&P / USENIX). Surface arxiv-only entries only when no venued match exists.

• Sort: recency desc → citation desc → venue tier.
• Dedupe by arxiv id / DOI / canonical URL.

1. Render (one bullet per paper, newest first):

• Title (in original language).
• Primary authors (max 3 then et al.).
• Year + venue if known.
• One-line plain-language takeaway (≤ 80 chars).
• Source URL (arxiv abs preferred over PDF).

1. Tool budget: aim for ≤ 3 tool calls total. List-task

adaptive budget already grants up to 10 — reserve the headroom for retries, not for browsing every candidate.

Verification

• Every paper has a real source URL.
• Title / authors / year are never invented; if unknown, omit the

field and say so plainly.

• If the result set is thin (< count solid hits), say "找到 N 篇

够稳的，剩下的来源都不太可信，先给你这 N 条" — do not pad.

• Do NOT ask the user to "确认是否保存到知识库" unless they signalled

long-term storage intent. Default is in-chat output only.

Pitfalls

• Semantic Scholar 429 is common; honour the Retry-After header

and degrade silently to arxiv-only.

• DDG / Bing captcha will trigger the v0.45 provider chain in

web_search; if all 3 web providers fail, fall back to whatever arxiv + S2 produced even if < count.

• "顶刊" sometimes means journal (TPAMI / TMLR) and "顶会" means

conference (CVPR / NeurIPS); when the user uses both, include at least one of each.

• Don't translate paper titles — keep the originals so the user can

search them as-is.