Skip to content

readme.md

Latest commit

 

History

History
214 lines (150 loc) · 10.2 KB

readme.md

File metadata and controls

214 lines (150 loc) · 10.2 KB

XtraMCP Orchestration Layer

XtraMCP is a custom MCP-based orchestration server that powers PaperDebugger’s higher-level workflows:

  • 🧑‍🔬 Researcher – find relevant papers and position your work within the literature
  • 🧑‍⚖️ Reviewer – critique drafts like a top-tier ML conference reviewer
  • ✍️ Enhancer – perform fine-grained, context-aware rewrites
  • 🧾 Conference Formatter (WIP) – adapt drafts to conference templates (NeurIPS, ICLR, AAAI, etc.)

This document describes the core tools exposed by XtraMCP and how they combine into these workflows.

Note: XtraMCP (see 4ndrelim/academic-paper-mcp-server) is currently closed-source while the API and development stabilize.
PaperDebugger runs fully without it; connecting XtraMCP unlocks the advanced research/review pipelines described here.

Tool Overview

Tool Name Role Purpose Primary Data Source Presently Deployed
search_relevant_papers Researcher Fast semantic search over recent CS papers in a local vector DB, enhanced with semantic re-ranker module Local DB Yes
deep_research Researcher Multi-step literature synthesis & positioning of your draft Local DB + retrieved papers analysis Yes
online_search_papers Researcher Online search over external academic corpora OpenReview + arXiv Yes
review_paper Reviewer Conference-style structured review of a draft Your draft + section-level review (static & semantic) Yes (slightly scaled down to balance operational cost)
verify_citations Reviewer Ensure citations are grounded, valid, and traceable Your draft's bibliography Yes
generate_citations Reviewer Generates BibTeX-style citations by simplying providing arxiv ID / DOI / URL / title Your draft Yes
enhance_academic_writing Enhancer Context-aware rewriting and polishing of selected text Your draft + XtraGPT Temporarily Disabled
get_user_papers Misc / Researcher Fetch all published papers with description, by a specific user OpenReview Yes
search_user Misc Fetch user's profile, including info such as publications, co-authors OpenReview Yes

1. search_relevant_papers

Purpose:
Search for similar or relevant papers by keywords or extracted concepts against a local corpus of ~800,000 academic papers.
This tool uses semantic search with vector embeddings to find the most relevant results, enhanced with a re-ranker module to better capture nuance. It is fast and the default and recommended tool for paper searches.

How it works:

  • Recent CS papers (last few years) are vectorized into a local index.
  • Queries (from your topic or draft) are embedded and matched via similarity search.
  • Results are reranked by an LLM-based reranker for better semantic alignment.

Typical usage:

  • “Find the 10 most relevant papers to this draft.”
  • “Search for relevant works on diffusion models for imbalanced medical imaging.”

2. deep_research

Purpose:
Given a research topic or draft paper, perform multi-step literature exploration and synthesis. Summarize their findings, and provide insights on similarities and differences to assist in the research process. This helps you to position your work.

How it works:

  1. Uses search_relevant_papers (and optionally online_search_papers) to retrieve candidate works.
  2. Summarizes key ideas, methods, and results from retrieved papers.
  3. Performs chain-of-thought style analysis to:
    • highlight similarities/differences vs your draft,
    • surface missing baselines or evaluation settings,
    • suggest how to position your contribution.

Typical usage:

  • “deep_research to compare my draft to recent work on retrieval-augmented generation.”
  • “For this topic, deep_research 5-10 relevant papers and explain where the open gaps are and what I can adopt.”

3. online_search_papers

Purpose:
Expand beyond the local DB to search online academic corpora (OpenReview + arXiv). This tool is ideal for discovering recent or broader papers beyond those available in the local database.

How it works:

  • Called when local search is too sparse (new topic) or you explicitly want the latest work.
  • Queries both OpenReview and arXiv for up-to-date results.
  • Results can then be fed into deep_research for synthesis.

Typical usage:

  • “My topic is very new. Look online for the latest preprints from OpenReview/arXiv.”

4. review_paper

Purpose:
Analyze and review a draft against the standards of top-tier ML conferences (ICLR, ICML, NeurIPS). Identifies improvements and issues in structure, completeness, clarity, and argumentation, then provides prioritized, actionable suggestions.

How it works:

  • Pass A – Deterministic checks (fast, high-precision)

    • Required sections present (e.g., Abstract, Method, Experiments, Limitations/Broader Impact).
    • Abstract contains problem, approach, core results, significance.
    • Acronyms defined at first use; “TODO”, “FIXME”, “Figure ??” flags.
    • Figures/tables referenced; equation references consistent; citation style uniform.
    • Reproducibility signals: code/data availability, hyperparameters, seeds, compute, eval protocol.

  • Pass B – Section-aware LLM critiques

    • Run per section with venue-aware rubrics (NeurIPS/ICML/ICLR style).
    • Suggest minimal, targeted edits (what to add/remove/clarify).
    • Focus on clarity, completeness, and logical flow.

  • Pass C – Cross-checks (claims vs evidence)

    • Are “state-of-the-art” claims backed by numbers + baselines?
    • Are method components properly ablated?
    • Are there red flags for data leakage, HPO on test sets, or missing uncertainty reporting?

  • Prioritization

    • Each issue is scored by severity (blocker/major/minor), impact, and confidence.
    • Duplicates are merged and top-N issues are surfaced as “quick fixes” vs “substantial rewrites”.

Typical usage:

  • “review_paper this draft like a NeurIPS reviewer and give me the top 10 issues to fix.”
  • “review_paper on method clarity and experimental rigor.”

5. verify_citations

Purpose: Verify that citations in your draft are valid, grounded, and traceable, helping reduce the risk of desk rejection due to incorrect or unverifiable references. There have been embarrassing cases where accepted papers were found with hallucinated citations. This tool helps to avoid such occurrences.

How it works:

  • Parses your bibliography and in-text citations.
  • Verifies that cited papers, preprints, and URLs:
    • exist and are reachable,
    • correspond to real publications or preprints,
    • match their claimed titles/authors/venues where possible.
  • Flags any issues or un-verifiable sources.

Typical usage:

  • “verify_citations”
  • “Check whether any citations in this draft are invalid or unverifiable.”

6. generate_citations

Purpose: Generate BibTeX-style citations easily by simply providing the paper's arxiv ID or DOI or URL or just its title.

How it works:

  • Searches online for the paper's official source
  • Formats into BibTeX-style citation ready for copy-pasting
  • Inform user if proposed information is invalid / unmatched

Typical usage:

7. enhance_academic_writing

Purpose:
Suggest context-aware academic writing enhancements for selected text.

How it works:

  • Powered by XtraGPT models tuned for academic style and LaTeX-heavy text.
  • Uses surrounding context (section, paper intent, venue) to:
    • improve clarity and flow,
    • reduce redundancy and filler,
    • keep technical content intact,
    • align tone with ML/AI papers.

Typical usage:

  • "enhance_academic_writing this paragraph to be clearer and more concise, preserving all technical details.”
  • "enhance_academic_writing the abstract to be suitable for NeurIPS.”

8. get_user_papers

Purpose:
Retrieve all papers authored by a given user (OpenReview), identified by email.
Useful for quickly assembling a researcher’s publication list or grounding context for comparison/positioning.

How it works:

  • Queries the paper database for matching author email(s).
  • Returns structured metadata: title, authors, venue, year, abstract, and identifiers.
  • Often used as a preprocessing step before deep_research.

Typical usage:

  • “get_user_papers for in summary mode.”
  • “Retrieve all publications by this researcher and then compare my draft using deep_research.”

9. Conference Formatter (WIP)

Upcoming workflows will:

  • map your draft onto specific conference templates (NeurIPS, ICLR, AAAI, etc.),
  • adjust sectioning, citation style, and boilerplate requirements,
  • highlight formatting and policy mismatches (e.g., ethics, broader impact sections).

Putting It Together: Example Orchestrated Flows

  • Researcher Flow

    1. Use search_relevant_papers on your draft or topic.
    2. If results are thin or stale, fall back to online_search_papers.
    3. Call deep_research to synthesize and position your work.

  • Reviewer Flow

    1. Run review_paper on the full draft.
    2. For high-impact issues, call enhance_academic_writing on the relevant spans.
    3. Verify your citations with verify_citations or if you haven't yet formatted them, include with generate_citations.

  • Enhancer Flow

    1. Select a paragraph or section in Overleaf.
    2. Call enhance_academic_writing with your preferences (e.g., “more formal”, “shorter”).
    3. Use edit-diff tool to effect changes.