Skip to content

Releases: Ar9av/PaperOrchestra

Release list

v0.2.0 — multi-agent skills, PaperBanana & new search backends

v0.2.0 — multi-agent skills, PaperBanana & new search backends Latest
Latest

Choose a tag to compare

View all tags
@Ar9av Ar9av released this 25 Apr 12:22
v0.2.0
c75e125

What's new

Added

  • agent-research-aggregator skill — orchestrates parallel sub-agents to aggregate and synthesise research findings across multiple papers. Auto-detected: activates only when the host environment supports multi-agent spawning.
  • setup.sh — one-shot setup script that wires the multi-agent skills integration into the host environment.
  • build_pdf.py — ReportLab-based fallback PDF builder for environments without LaTeX. Produces two-column IEEE-style PDFs with clickable in-text citations and an auto-injected architecture diagram.
  • PaperBanana backbone for plotting-agent — optional integration for hosted figure rendering; now accepts an OpenRouter key or a Google (Gemini) key.
  • Semantic Scholar API key integration — set SEMANTIC_SCHOLAR_API_KEY to raise the literature-review-agent's rate limit beyond the anonymous tier.
  • Exa search backend for literature-review-agent — semantic search via Exa; activate by setting EXA_API_KEY.

Fixed

  • Pipeline bottlenecks in an end-to-end CitationRL run: redundant re-fetch loops, blocking I/O in the citation-gate helper, and a slow path in the refinement halt logic.

Changed

  • Simplified PaperBanana integration surface; added upstream citation in CITATION.cff.
  • Project renamed to PaperOrchestra in all README headings.
  • README updated with paper preview thumbnail, performance metrics table, skills explanation, and OOS-metrics achievement badge.

Full changelog: v0.1.0...v0.2.0

Assets 2
Loading

v0.1.0 — Initial public release

v0.1.0 — Initial public release

Choose a tag to compare

View all tags
@Ar9av Ar9av released this 09 Apr 15:21
v0.1.0
c2de31b

A host-agent-pluggable skill pack implementing PaperOrchestra (Song et al., 2026, arXiv:2604.05018) for any coding agent.

Song, Y., Song, Y., Pfister, T., Yoon, J. PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing. arXiv:2604.05018, 2026.

What it is

Seven composable skills + a benchmark harness + four autoraters that turn Claude Code, Cursor, Antigravity, Cline, Aider, OpenCode (or any coding agent) into the PaperOrchestra five-agent paper-writing pipeline.

No API keys. No embedded LLM clients. No SDK dependencies. Skills are markdown instruction documents the host agent reads and executes using its own native tools. Deterministic Python helpers do only mechanical work (JSON schema validation, Levenshtein fuzzy citation matching, BibTeX formatting, citation gates, LaTeX sanity, refinement halt logic).

The seven skills

Skill Paper step LLM calls Role
paper-orchestra orchestrator — Top-level driver
outline-agent Step 1 1 Idea + log + template + guidelines → outline JSON
plotting-agent Step 2 ~20–30 Render plots & diagrams + optional VLM critique loop
literature-review-agent Step 3 ~20–30 Web-search + S2 verify (Levenshtein > 70) + ≥90% citation integration
section-writing-agent Step 4 1 One single multimodal call drafting all remaining sections
content-refinement-agent Step 5 ~5–7 Simulated peer review + strict accept/revert halt rules
paper-writing-bench §3 — Reverse-engineer raw materials from existing papers
paper-autoraters App. F.3 — Citation F1, LitReview Quality, SxS Paper, SxS LitReview

Fidelity to the paper

All 13 prompts from Appendices D.4, F.1, F.2, F.3 are reproduced verbatim with per-page citations to the paper. See docs/paper-fidelity.md for the full design-decision → paper-page map.

What's included

  • 76 files, ~6.8K lines
  • 8 SKILL.md instruction documents
  • 35 reference documents (verbatim prompts, JSON schemas, rubrics, halt rules, integration guides)
  • 19 deterministic Python helper scripts (with argparse CLIs)
  • Realistic end-to-end example (examples/minimal/ — a fabricated but coherent Adaptive Top-K Attention paper case)
  • Architecture, paper-fidelity, and per-host integration docs

Quickstart

`bash
git clone https://github.com/Ar9av/paper-orchestra ~/paper-orchestra
cd ~/paper-orchestra
pip install -r requirements.txt

symlink into your host's skill directory (Claude Code shown)

mkdir -p ~/.claude/skills
for s in paper-orchestra outline-agent plotting-agent literature-review-agent
section-writing-agent content-refinement-agent paper-writing-bench
paper-autoraters; do
ln -sf ~/paper-orchestra/skills/$s ~/.claude/skills/$s
done
`

Then ask your coding agent: "Run the paper-orchestra pipeline on ./workspace".

See docs/coding-agent-integration.md for Cursor, Antigravity, Cline, Aider, OpenCode setup.

Smoke-tested

All 10 deterministic helpers pass on real inputs:

  • validate_inputs.py ✓
  • validate_outline.py ✓
  • extract_metrics.py ✓ (4 tables extracted from the example log)
  • levenshtein_match.py ✓ (incl. substring-bypass)
  • check_cutoff.py ✓ (PASS / FAIL paths)
  • dedupe_by_id.py + bibtex_format.py ✓ (real BibTeX entries with deterministic keys)
  • orphan_cite_gate.py ✓
  • latex_sanity.py ✓
  • citation_coverage.py ✓
  • anti_leakage_check.py ✓
  • score_delta.py ✓
  • render_matplotlib.py ✓ (real PNG at requested aspect ratio)
  • render_diagram.py ✓

Citation

If you use this skill pack, please cite the paper:

bibtex @article{song2026paperorchestra, title={{PaperOrchestra}: A Multi-Agent Framework for Automated {AI} Research Paper Writing}, author={Song, Yiwen and Song, Yale and Pfister, Tomas and Yoon, Jinsung}, journal={arXiv preprint arXiv:2604.05018}, year={2026}, url={https://arxiv.org/pdf/2604.05018} }

License

MIT — see LICENSE.

Loading