Stage 4 (Closest Related Work) and every 2024–2026 paper has been checked via live search — publication status, venue, and DOI are current as of this September 2026 snapshot (not a November 2026 re-verification). Older foundational papers (Stages 1, 5, 6) are well-established, high-confidence citations. Re-verify all Stage 4 papers before final thesis submission — this subfield moves fast (three status changes happened during compilation).
Stage 1 — Background Knowledge
Foundational papers establishing vocabulary and context. Read before Phase 0.
Stage 2 — Multi-Agent Systems
Core multi-agent debate papers — baselines and closest architectural competitors.
Stage 3 — Scientific Reasoning & Evaluation Datasets
Dataset-adjacent papers and scientific claim verification methodology.
Stage 4 — Closest Related Work (Contribution / Limitation / Gap Analysis)
The most load-bearing papers — closest competitors whose differentiation is critical for thesis framing. Re-read before writing.
Stage 5 — Supporting Technologies & Infrastructure
Engineering references for implementation. Read on demand when building specific components.
Stage 6 — Evaluation Methodology
Experimental design, statistical validation, and metric foundations.
Reading Timeline (per Phase)
Before Implementation
Papers 1-3, 7-9, 14-15, 21-22, 24, 35. ~20h across team, parallelizable.
During Implementation
Papers 4, 10, 15 (re-read), 26, 30-31. Read before building specific components.
Before Experiments
Papers 6, 18-20, 28-29, 32-34, 37-38. Read before finalizing eval pipeline.
Before Writing
Papers 21-25, 27, 36, 39. Re-read for sharp related-work differentiation.
Living Literature Tracker
Alerts
arXiv (cs.CL, cs.AI, cs.MA) + Google Scholar alerts for "multi-agent debate", "sycophancy", "trust calibration", "evidence-grounded"
Venues to Monitor
ACL, EMNLP, NAACL, NeurIPS, ICLR, ICML, AAAI — where closest competing work has historically landed
Tooling
Zotero (free, BibTeX export) + shared Notion/spreadsheet matrix with columns: venue (verified), arXiv ID, contribution, limitation, relevance, last-verified-date
⚠ Best practice: Re-check every Stage 4 paper's status if its "last verified" date is more than ~2 months old by the time it's actually cited in the thesis. This single practice would have caught both surprises (iMAD venue update, Yao et al. withdrawal) automatically.