brycewang-stanford ECAI Writing Style
ECAI reviews a 7-page body across the full breadth of AI (symbolic reasoning, KR, planning, search, multi-agent systems, ML, and applications). Two things follow, and they drive most of the advice here: the contribution must be legible to a general AI audience, and the writing must be dense — at 7 pages there is no room for a paragraph that does not carry the argument.
The ECAI first-page arc
Land the whole contribution before the fold:
- A well-posed AI problem — stated as a problem, not as "X has become popular."
- Why current methods are inadequate — the specific gap, in one or two sentences.
- The contribution — the mechanism, and where the claim is provable, the guarantee.
- Evidence proportional to the claim — a theorem + construction, and/or a fair empirical comparison; named on the first page, delivered in the body.
- What it means for AI — why a general AI audience should care.
The worked example (../../resources/worked-examples/01-introduction.md)
shows this arc rebuilt from a benchmark-first draft.
Lead with the AI contribution, not the application of a model
The most common re-route signal is a paper that leads with "we apply to ." ECAI
rewards a contribution that generalizes — a mechanism, a guarantee, a characterization, an
understanding — over a single benchmark delta. Apply the model-swap test: if you replaced the
underlying model/solver with another, would a lasting AI lesson remain? If not, the paper may belong
at a pure-ML venue (ecai-topic-selection).
Match evidence to the claim shape
ECAI's breadth means the right evidence differs by contribution:
| Claim shape | Evidence ECAI expects |
|---|---|
| "This always holds / is complete / is optimal" | A proof, with all assumptions explicit |
| "This is more efficient / expands fewer nodes" | A controlled comparison vs a fair baseline, with spread |
| "This learns better / calibrates better" | A fair empirical comparison, seeds, and a reason why (not just a number) |
| "This works in deployment" | A credible real-world demonstration (route to PAIS) |
A provable claim asserted only empirically is a weakness a reviewer will name; a purely empirical claim dressed as a theorem is worse.
Density discipline (the 7-page reality)
- Paragraph one carries the contribution. Do not spend the opening on the importance of AI.
- Cut the roadmap. At 7 pages the paper cannot afford a "Section 2 does X, Section 3 does Y" preview; a single orienting sentence is enough.
- Define once, precisely. Symbolic-AI reviewers check every later lemma against your definitions; sloppy notation costs you the proof's credibility.
- One figure doing three jobs beats three figures. Merge panels; caption them to be self-contained.
- Push detail, keep the idea. Full proofs and extra tables go to the supplement; the idea and
the decision-critical result stay in the body (
ecai-supplementary).
Threats / limitations as argument, not boilerplate
State the honest boundary of the claim where it lives — the assumption the theorem needs, the
regime where the method stops helping, the confound the experiment cannot rule out. In a
single-round, no-revision process (ecai-review-process), a limitation you name yourself is far
cheaper than one a reviewer discovers.
Language and audience
- Write for a broad AI reader: define subfield jargon, motivate why a planning/KR/ML reader should care even if it is not their area.
- ECAI is an international European venue; keep the English clear and the claims measured — EurAI's reviewer pool spans many first languages and subfields.
- Avoid overclaiming ("revolutionizes," "solves"); ECAI rewards a precise, bounded contribution.
Anti-patterns
- Benchmark-first abstract that never states an AI problem.
- Model-as-contribution with no lesson surviving a model swap.
- Proof by assertion — a completeness/optimality claim with no proof.
- Roadmap padding eating the 7-page budget.
- Decision-critical content in the supplement because the body ran long.
Output format
[First-page arc] problem / inadequacy / contribution+guarantee / proportional evidence / meaning — all present?
[Model-swap test] does an AI lesson survive swapping the model/solver? yes/no
[Evidence match] claim shape -> proof and/or fair comparison present? gaps: <list>
[Density] roadmap trimmed? paragraph one carries the contribution? figures merged?
[Limitations] stated as argument where the claim lives? yes/no
[Budget] decision-critical content inside 7 pages? yes/no