The Ladder mechanism¶
Definition¶
The Ladder (Blum & Hardt, 2015) is a leaderboard algorithm that stays accurate under adaptive resubmission: because participants who repeatedly score against a holdout begin to overfit it, the Ladder controls what the leaderboard reveals, achieving strong guarantees in a fully adaptive estimation model while remaining practical — including a parameter-free variant deployable with no tuning.
Explanation¶
The problem is structural, not adversarial in intent: every published score is information about the holdout, and a sequence of submissions guided by those scores is an optimization loop running against the evaluation data itself — the leaderboard drifts from measuring quality to measuring holdout-fitting. Prior practice resisted with folk heuristics (limited resubmission rates, truncated score precision) that the paper characterizes as poorly understood. The Ladder's contribution is a principled mechanism with a competition-tailored accuracy notion, guarantees that hold under full adaptivity, demonstrated resistance to practical attacks, and validated utility on real Kaggle submission data — while sidestepping a hardness result that rules out seemingly similar accuracy notions. The design lesson for any score-gated eval: the reporting channel is part of the mechanism, and what you reveal per run determines how fast your holdout depletes. The staged source is the abstract; procedure detail is in the paper.
Key Properties¶
- Adaptive resubmission overfits the holdout through the scores themselves
- Guarantees in a fully adaptive model; resistant to practical attacks; validated on real Kaggle data
- Parameter-free deployable variant
- Reporting policy is part of the eval mechanism, not an afterthought
Relationships¶
- Adaptive data analysis — is the deployed instance of that theory for the leaderboard special case
- Golden-dataset retrieval evals — supplies the mechanism-level justification for that harness's ratchet-style gating and restrained score reporting
Applications¶
Designing what a CI eval publishes per run (ratchets, noise bands, thresholds) so the sealed subset's validity survives repeated consultation.
Sources¶
- https://arxiv.org/abs/1502.04585