Research map / Learning theory and algorithms / Fairness
Dissecting Submission Limit in Desk-Rejections: A Mathematical Analysis of Fairness in AI Conference Policies
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Formalizes desk-rejection under per-author submission limits, proves an ideal rejection rule is impossible for n≥3 authors, and proposes LP-based fairness-aware desk-rejection (NP-hard for individual fairness, LP-solvable for group fairness).
Problem
Per-author submission limits in AI conferences cause desk-rejections that unfairly penalize co-authors (especially early-career researchers) of prolific authors; no existing rule rejects only excess papers.
Approach
Models desk-rejection as an optimization over paper-retention vectors with author-level cost (fraction of own papers rejected); individual fairness (max cost) is NP-hard, while group fairness (average cost) is an LP whose relaxation has integral optima, enabling off-the-shelf LP solvers.
Key results
- Proves that for n≥3 authors no algorithm can always achieve the ideal desk-rejection (each author keeps min{x,|P_i|} papers), via explicit counterexamples.
- Shows individual fairness minimization (max author cost) is NP-hard via reduction from Set Cover.
- In the 2-author example (x=25, m=26), the CVPR-style order-based rule gives zeta_ind=1 and zeta_group=27/52, while the LP-based method recovers the ideal values zeta_ind=1/26 and zeta_group=1/52.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use fairness-aware desk rejection instead of the CVPR 2025 algorithm | Desk-rejection system · Case study (n=2, x=25, m=26) | individual fairness (ζ_ind): 1 -> 1/26 | Helped |
| Use fairness-aware desk rejection instead of the CVPR 2025 algorithm | Desk-rejection system · Case study (n=2, x=25, m=26) | group fairness (ζ_group): 27/52 -> 1/52 | Helped |
| Use the roulette algorithm instead of fairness-aware desk rejection | Desk-rejection system · Case study (n=2, x=25, m=26) | expected individual fairness (E[ζ_ind]): (25/26) ·(1/26) + (1/26) · 1 ≤ 1/26 | Hurt |
| Use the roulette algorithm instead of fairness-aware desk rejection | Desk-rejection system · Case study (n=2, x=25, m=26) | expected group fairness (E[ζ_group]): (25/26) · (1/52) + (1/26) · (27/52) ≤ 27/52 | Hurt |
Surprising findings
- Order-based desk-rejection can penalize a low-submission co-author for a paper submitted late, so authors must compete for earlier submission slots.
Limitations
- Evaluation is limited to a small illustrative case study rather than real conference data, which is unavailable.
- Ideal desk-rejection impossibility proofs rely on specific worst-case instances; practical impact is unquantified.
- Fairness metrics are cost-based proxies and do not capture career-stage impact directly.
Experimental setup
Metrics: individual fairness group fairness
Baselines: CVPR 2025 order-based desk-reject algorithm roulette algorithm
Builds on: Set Cover NP-hardness utilitarian and egalitarian social welfare
Closest papers by meaning
- Group Fairness in Peer Review (NeurIPS 2023 · similarity 0.86)
- A Novice-Reviewer Experiment to Address Scarcity of Qualified Reviewers in Large Conferences (AAAI 2021 · similarity 0.86)
- Position: The AI Conference Peer Review Crisis Demands Author Feedback and Reviewer Rewards (ICML 2025 · similarity 0.84)
- Two Tickets are Better than One: Fair and Accurate Hiring Under Strategic LLM Manipulations (ICML 2025 · similarity 0.82)
- Fair Participation via Sequential Policies (AAAI 2024 · similarity 0.82)
- Fast & Fair: A Collaborative Platform for Fair Division Applications (AAAI 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.