brycewang-stanford ECAI Related Work
ECAI's related work has a peculiar difficulty: the reviewer pool is broad (a planning paper may draw a KR reviewer and an ML reviewer), yet the space is tiny — the body is 7 pages and references overflow into only 1 page (standalone ECAI) or 2 pages (IJCAI-ECAI 2026). You must position against the right lanes for a general audience and stay compact.
Cover the lanes a general-AI reviewer will expect
Identify which AI lanes your contribution touches and place it in each — a broad reviewer will fault a section that only cites its author's home subfield:
- Symbolic / foundational lane — the KR, planning, search, reasoning, or argumentation line your problem sits in.
- Learning lane — the ML/statistical methods your work uses or competes with.
- Multi-agent / decision lane — if agents, game-theoretic, or sequential-decision aspects apply.
- Application lane — the real-world domain, especially for a PAIS-flavored contribution.
- Cross-venue neighbors — the closest work at IJCAI, AAAI, AAMAS, KR, or a specialist venue, so a reviewer from that community sees you know the field's flagship results.
You will not cover all lanes deeply at 7+1(2) pages — cover the ones your claim actually engages, and cite the canonical entry point for the rest.
Delta-first positioning
Do not write an annotated bibliography. For each closest competitor, state the delta in one sentence: what they do, and precisely what you do differently or additionally.
"Prior learned-heuristic planners plug the model in directly and lose completeness [refs]; we preserve completeness and a (1+ε) bound via per-state switching."
Delta-first writing does three jobs at once in minimal space: it credits prior work, states your novelty, and pre-empts the "how is this different from [X]?" reviewer question.
Keep it double-blind
ECAI is double-blind. Self-citations are the classic leak:
- Cite your own prior work in the third person ("Smith et al. [7] showed...") — never "in our previous work [7]."
- Do not let the density or phrasing of self-citations reveal the author group.
- Anonymize a system name that identifies you if it appears in the comparison.
# Catch first-person self-citation and identity leaks in related work
grep -nEi 'our (previous|prior|earlier) (work|paper|system|approach)|we previously' paper.tex | head
Budget discipline
- The reference pages are 1 page (standalone) or 2 pages (IJCAI-ECAI 2026) — finite. Prune to the works that position the contribution, not every paper you read.
- Prefer the canonical citation for a well-known result over three secondary ones.
- Move an extended survey-style discussion to the supplement if it is not decision-critical
(
ecai-supplementary); keep the delta-establishing citations in the body. - Do not pad the reference list to signal thoroughness — a broad reviewer notices citations that do no positioning work.
Common failure modes
- Home-subfield tunnel vision — a planning paper that ignores the learning lane it competes with, or an ML paper blind to the symbolic literature it reinvents.
- Bibliography, not positioning — listing works without stating the delta.
- First-person self-citation breaking double-blind.
- Missing the obvious neighbor at IJCAI/AAAI/AAMAS/KR that a specialist reviewer will know by heart.
- Reference-budget overflow forcing body cuts (
ecai-submission).
Output format
[Lanes covered] symbolic / learning / multi-agent / application / cross-venue — which apply, all present?
[Delta statements] each closest competitor -> one-sentence delta? gaps: <list>
[Double-blind] self-citations third-person; no identifying system names? yes/no
[Budget] reference pages within 1 (standalone) / 2 (2026); extended survey moved to supplement?
[Neighbor check] closest IJCAI/AAAI/AAMAS/KR result cited? yes/no