Paper Evaluation

Guidelines for EvaluatePapers and FilterPapers. Use before these tools to ensure consistent scoring and selection.

EvaluatePapers Scoring (1–5)

Score	Meaning	Action
1	Irrelevant	should_retry: true if not yet retried
2	Mostly irrelevant	should_retry: true
3	Partially relevant	Proceed; consider retry if idea is still fuzzy
4	Relevant	Proceed
5	Highly relevant	Proceed

should_retry Logic

• score < 3 and not yet retried: Call ExtractKeywords again with refined idea, then SearchArxiv again
• score >= 3: Proceed to FilterPapers
• suggestion: Use the suggestion field to guide keyword refinement (e.g. "Try adding domain terms like 'medical imaging'")

FilterPapers Principles

• Count: Select 5–8 most relevant papers
• Order: By relevance to idea (not by arXiv date)
• Indices: Use 1-based indices from papers_summary (e.g. [1, 3, 5, 7, 9] for top 5)
• Diversity: Prefer papers that cover different aspects of the idea when possible
• Avoid: Duplicate or near-duplicate papers (same authors, same method)

Relevance Criteria

When evaluating, consider:

• Direct: Paper directly addresses the idea's core question
• Method: Paper introduces methods relevant to the idea
• Domain: Paper is in the same domain (e.g. NLP, ML)
• Recency: Prefer recent papers when idea is time-sensitive
• Citation: Highly cited papers may indicate foundational work

Output Format

EvaluatePapers returns: {"score": 1–5, "should_retry": bool, "suggestion": "string"}

FilterPapers receives indices and returns: {"count": N, "indices": [1, 2, ...]}