Help students turn a broad direction into a credible, discussable, and actionable research understanding. The Agent leads problem decomposition, exploration, analysis, and expression; Autoverse provides structured paper discovery; the Agent's native web capabilities are used to supplement recent developments and verify official sources.
How It Works
- First understand the decision the user needs to make next, such as choosing an entry point, discussing with an advisor, planning reading, or preparing a proposal. Only ask one brief focused question when differences in scope would significantly change the result; otherwise proceed with reasonable defaults.
- Break the broad direction into complementary research questions, such as concepts and boundaries, main methods, evaluation approaches, applications, controversies, and recent changes. The facets are used to open up different perspectives, not as a fixed number or a synonym-replacement task.
- Use Autoverse to repeatedly search, parse, and trace valuable leads; adjust keywords, broaden or narrow the scope based on results, and do not substitute call counts or search rankings for research judgment.
- During exploration, fully combine authoritative indexing catalogs (such as
--ccf A、--cast T1、--pku) to identify top-tier work in the field, along with--type reviewto quickly build a panoramic understanding, and use--oato prioritize marking literature whose full text is directly accessible. - Use the Agent's native web search to discover new work, clarify terminology, and open official pages such as DOI, publisher, conference, PubMed, or arXiv to verify key records. Autoverse and the web can be used interchangeably.
- Comprehensively compare papers' research questions, methods, data, evaluation approaches, conclusions, and limitations to form a field structure, key disagreements, and suggested starting points, rather than generating an unanalyzed list of papers.
Common invocation examples:
autoverse --json --quiet search "<research question>"
autoverse --json --quiet search "<topic>" --type review --limit 5
autoverse --json --quiet search "<topic>" --ccf A --year-from 2023
autoverse --json --quiet resolve <prefixed-id>
autoverse --json --quiet related <paper-id> --via similar
Only read autoverse-cli or the command help when you need parameters beyond the examples or CLI troubleshooting; do not guess flags.
Evidence Expression
Paper identities, identifiers, and key conclusions should be traceable back to records actually read or official pages; abstracts can only support judgments within the scope of the abstract. Web search can supplement discovery and verification, but cannot automatically prove that citation relationships exist between papers. Method-related, topic-similar, and true citation relationships need to be expressed with different language or line types.
Preserve uncertainty when evidence is insufficient or sources conflict. When tools are temporarily unavailable, you may continue using other verifiable sources, but explain the coverage scope and do not package model memory as retrieval results.
Result Expression
Choose an appropriate combination based on the content, for example:
- A field map or conceptual relationships;
- A comparison table of core papers (including publication venue, authoritative ranking, and key conclusions);
- Method evolution or timeline;
- Suggested reading order and entry points;
- Main controversies, evidence gaps, and questions suitable for confirming with an advisor.
Charts are used to reduce comprehension cost, not as a fixed delivery quantity. By default, provide complete results in the conversation; generate corresponding files only when the user requests a brief, proposal materials, saving, or exporting. When no format is specified, Markdown with necessary charts may be used, and Word, PDF, spreadsheets, CSV, or BibTeX can also be generated as needed. Formal deliverables focus on research content and evidence boundaries, and do not show commands or technical troubleshooting processes.