Lewati ke isi

[GS05] MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution

Kunci Stabil: GS05 | Sitasi BibTeX: 1

Ringkasan Metadata

  • Penulis & Tahun: Wei Tao and others, 2024
  • Venue & Tier: NeurIPS 2024 (Vol 37)
  • DOI / URL: 10.52202/079017-1647
  • Skor Kualitas (QA Score): - (QA korpus baru OPEN, belum dinilai)
  • Peran Studi: QGS (penguji recall P1+P2)
  • Tingkat Ekstraksi: metadata-calibration
  • Cakupan (Scope): Repository-level GitHub issue resolution (SWE-bench 500 instances).

1. Masalah Utama (Core Problem)

Resolusi GitHub issue repo-level secara otonom.

2. Arsitektur yang Diajukan (Architecture)

Hierarki 4 agen: Manager, Custodian, Developer, QA (supervisor-worker).

3. Mekanisme Koordinasi & Kontrol (Coordination)

Manager membagi tugas dan mengarahkan agen lain; Custodian membatasi search space berkas; 100% otonom tanpa HITL gate.

4. Foundation Model & Infrastruktur (FM)

GPT-4 era.

5. Tolok Ukur & Dataset (Benchmark)

SWE-bench (500 issue repo-level).

6. Artefak & Kode Sumber (Artifact)

weitao92/MAGIS.

7. Metrik Primer Eksak (Primary exact)

Resolved rate 13.94% = 8x GPT-4 1.74% (PDF hal 1+7/Table 2 hal 7).

8. Metrik Sekunder (Secondary)

-

9. Pola Kegagalan (Failure Modes)

86% isu gagal terselesaikan tanpa batas eskalasi terukur.

10. Ancaman Validitas (Threats)

UNDOCUMENTED (QA korpus baru OPEN).

11. Jejak Bukti & Provenansi (Provenance)

Bukti Verifikasi: EVIDENCED-FACT (GS_quasi_gold.md GS05; Scopus-769 HIT judul eksak; arXiv 2403.17927v2 HIT).

  1. Wei Tao and 2024 others. MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution. 2024. Venue: NeurIPS 2024 (Vol 37), DOI: 10.52202/079017-1647, Role: QGS (penguji recall P1+P2). ↩