An accepted ICML 2026 position paper argues unenforceable guidelines have failed, and that only a fungible credit currency, points earned by reviewing and spent on perks, can rescue ML peer review.
ICML 2026's Position Track has accepted a paper that proposes to remake machine learning peer review on something closer to a market than a guideline: a community credit system in which reviewers earn points for service and spend them on perks such as conference fee waivers, registration discounts, priority submission slots, recognition badges, and contribution credits.
The paper, "Want Better ML Reviews? Stop Asking Nicely and Start Incentivizing with a Credit System", was accepted to the ICML 2026 Position Track and frames two failure modes the field has not solved: how to limit the number of submissions per cycle, and how to discourage sloppy reviewing when the cost of being sloppy is zero. Reciprocal-review policies, OpenReview's transparency features, and desk-rejection letters have moved the floor on honesty, but they cannot price good review above the time-burden of writing a low-effort critique. The result is a publication system that cannot distinguish the worst reviews from the best, and the best increasingly decline to serve.
The proposed mechanism is a fungible credit currency. A reviewer earns roughly one point for completing a review and three points for an editor-flagged outstanding review. The currency is community-issued, not paid out of a conference's general fund. It costs nothing to mint, but the perks it buys are real.
The argument lands because the gap is not novelty but enforcement. Reviewer Guidelines have no tail: nothing happens to a reviewer who skips, files late, or posts a thin critique. A credit system has one. At the next deadline, the reviewer's balance is visible to them and the perks they could have redeemed are gone. That is the mechanism the paper says has been missing for a decade.
This is not the first iteration. In 2019, NeurIPS rewarded its top reviewers with free registrations, a one-off version of the same logic. The CMU Machine Learning Department's postmortem of the ICML 2020 reciprocal-review experiment documented how a forced-recusal, mutual-review design affected both quality and reviewer satisfaction. And an external platform, ReviewerCredits, already issues a portable reviewer-credit currency that universities and journals recognize across venues. The ICML proposal lands on a roughly six-year-old substrate of experiments and tools, not a blank page.
A companion preprint, "The AI Conference Peer Review Crisis Demands Author Feedback and Reviewer Rewards", reaches a similar conclusion from a different angle: author-side feedback plus reviewer rewards, in lieu of more guidelines. Both papers converge on the same diagnosis: the review pipeline's failure mode is now a coordination problem, not a process problem, and coordination requires prices, not pleadings.
The honest gap is governance. A credit system imposes market logic on a community that has run on volunteer norms for decades. ReviewerCredits exists, but it is opt-in and small. ICML adopting an internal currency would be the largest test yet of whether scientific publishing can absorb a price signal without breaking the implicit contract that says reviewers serve the field, not redeem against it. The Position Track functions as a community referendum at the conference itself, so the proposal moves from argument to opt-in trial if its audience at ICML 2026 backs it. Every other major ML conference will then have a forced choice: adopt, fork, or decline.