SpectrAI-Initiative/Paper Discussion Full

Paper Discussion Full

Use this skill when the user wants the full structured discussion workflow for one paper rather than a simple Q&A exchange.

Goal

Produce a comprehensive, evidence-grounded discussion report for one paper by emulating the full discussion pipeline:

1. Moderator
2. Librarian
3. Skeptic
4. Reproducer
5. Convergence
6. Final Scribe Report

Workflow

1. Identify the target paper. If it is missing, ask for the title, URL, or PDF link.
2. Use readPaper to ground the analysis in the full paper whenever possible.
3. If the user explicitly wants related-work comparison, use searchArticles to retrieve nearby papers; otherwise stay tightly focused on the selected paper.
4. Reason through the following stages in order and expose the stage outputs clearly:
   • Agenda
   • Evidence Summary
   • Critical Analysis
   • Reproducibility Check
   • Convergence
   • Final Report
5. Keep every stage grounded in available evidence. When evidence is missing, say so explicitly.

Stage Expectations

Agenda

• Define the discussion focus.
• Name the main evaluation axes: novelty, evidence quality, methodology, reproducibility, limitations.

Evidence Summary

• State the core claim, method, setup, and reported results.
• Separate explicit evidence from inference.
• When referencing external work for comparison, cite it using searchArticles results.

Critical Analysis

• Challenge overclaims, weak baselines, missing ablations, and threats to validity.
• Mark issue severity as Critical, Moderate, or Minor.
• When identifying missing baselines or comparing to external methods, cite specific papers found via searchArticles.

Reproducibility Check

• Judge whether the paper is easily, partially, or poorly reproducible.
• List what is specified versus what is missing.
• Propose a minimal reproduction recipe.

Convergence

• Summarize agreement, disagreement, and unresolved questions.

Final Report

Use this exact structure:

Paper Discussion Report

1. Paper Snapshot

2. Key Claims

3. Strengths

4. Weaknesses / Risks

5. Reproducibility Assessment

6. Open Questions

7. Recommended Next Actions

End with:

Overall take: ...

8. References

Quality Rules

• Do not fabricate details not present in the paper context. All external references must come from searchArticles results, never from model memory.
• Keep criticism specific and technically grounded.
• Preserve nuance rather than collapsing everything into a single score.
• If paper text cannot be retrieved, state that the report is based on limited context.

Citation Policy

• Any claim about related work, competing methods, or external benchmarks MUST include a citation retrieved via searchArticles. Do NOT cite from memory.
• When the discussion references methods or results not in the seed paper, call searchArticles first and cite what is returned.
• If no supporting reference is found, mark the claim: [Unverified — no supporting reference found via search].
• The target paper itself must be cited formally at the top of the report.

Inline Citation & Reference Format

Use numbered inline citations in the text body (e.g., [1], [2]) and collect full references in the ## 8. References section. All references MUST be rendered in markdown so they are clickable and well-formatted:

Inline example:

该方法在 ImageNet 上超越了 ViT [1]，但在小数据集上的泛化能力受到质疑 [2]。

References section example:

## 8. References

1. **Dosovitskiy, A. et al.** (2021). *An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale.* ICLR 2021. [arXiv:2010.11929](https://arxiv.org/abs/2010.11929)
2. **Liu, Z. et al.** (2021). *Swin Transformer: Hierarchical Vision Transformer using Shifted Windows.* ICCV 2021. [DOI:10.1109/ICCV48922.2021.00986](https://doi.org/10.1109/ICCV48922.2021.00986)

• Each reference line MUST include: Author(s) (Year). Title. Venue/Journal. Linked DOI or URL when available from searchArticles.
• Use markdown bold for authors, italic for title, and [text](url) for clickable DOI/URL links.
• Number references sequentially as they first appear in the text.