About

PapersWithELO is an open, incremental attention ranking protocol for machine-learning conferences. It exists because reader attention — not paper quality assessment — is the binding constraint.

What we are, what we are not

We are

  • A negative filter: what you can safely skip
  • An auditable ranking with uncertainty
  • A protocol: incremental, model-agnostic, reproducible

We are not

  • An AI reviewer or accept/reject oracle
  • A “best paper” picker — the top decile is 50/50, as it should be
  • A replacement for human judgment

The headline numbers

On the ICLR 2026 sample (n = 730, 10% of submissions, seed 42): the base reject rate is 68.9%. In our bottom 25%, 90.2% were rejected. In our bottom 50%, 85.5% were rejected. Skipping the bottom half of the ranking costs you very few accepted papers and returns half your reading time.

Study years: 2024. See analytics for the full validity, reliability, and convergence evidence — including our published failure cases.

Principles

  • Static-first: no database, no accounts, $0 hosting. All JSON precomputed offline.
  • Auditable: every θ, observation, and judge identity traceable to raw outputs.
  • Honest about noise: uncertainty is a first-class output, and our misses are linked from the analytics page.
  • No shaming: author statistics are aggregate-only, by design.