#ICML2025 excellent place paper: Interview with Jaeho Kim on addressing the issues with convention reviewing

At this yr’s Worldwide Convention on Machine Studying (ICML2025), Jaeho Kim, Yunseok Lee and Seulki Lee received an excellent place paper award for his or her work Place: The AI Convention Peer Evaluation Disaster Calls for Creator Suggestions and Reviewer Rewards. We hear from Jaeho in regards to the issues they have been making an attempt to handle, and their proposed writer suggestions mechanism and reviewer reward system.

May you say one thing about the issue that you simply deal with in your place paper?

Our place paper addresses the issues plaguing present AI convention peer assessment programs, whereas additionally elevating questions in regards to the future path of peer assessment.

The upcoming drawback with the present peer assessment system in AI conferences is the exponential progress in paper submissions pushed by growing curiosity in AI. To place this with numbers, NeurIPS acquired over 30,000 submissions this yr, whereas ICLR noticed a 59.8% improve in submissions in only one yr. This large improve in submissions has created a basic mismatch: whereas paper submissions develop exponentially, the pool of certified reviewers has not saved tempo.

Submissions to a number of the main AI conferences over the previous few years.

This imbalance has extreme penalties. The vast majority of papers are now not receiving ample assessment high quality, undermining peer assessment’s important perform as a gatekeeper of scientific information. When the assessment course of fails, inappropriate papers and flawed analysis can slip via, probably polluting the scientific document.

Contemplating AI’s profound societal influence, this breakdown in high quality management poses dangers that reach far past academia. Poor analysis that enters the scientific discourse can mislead future work, affect coverage selections, and finally hinder real information development. Our place paper focuses on this crucial query and proposes strategies on how we are able to improve the standard of assessment, thus main to higher dissemination of data.

What do you argue for within the place paper?

Our place paper proposes two main modifications to deal with the present peer assessment disaster: an writer suggestions mechanism and a reviewer reward system.

First, the writer suggestions system allows authors to formally consider the standard of opinions they obtain. This method permits authors to evaluate reviewers’ comprehension of their work, determine potential indicators of LLM-generated content material, and set up primary safeguards in opposition to unfair, biased, or superficial opinions. Importantly, this isn’t about penalizing reviewers, however moderately creating minimal accountability to guard authors from the small minority of reviewers who could not meet skilled requirements.

Second, our reviewer incentive system offers each instant and long-term skilled worth for high quality reviewing. For brief-term motivation, writer analysis scores decide eligibility for digital badges (similar to “Prime 10% Reviewer” recognition) that may be displayed on tutorial profiles like OpenReview and Google Scholar. For long-term profession influence, we suggest novel metrics like a “reviewer influence rating” – basically an h-index calculated from the next citations of papers a reviewer has evaluated. This treats reviewers as contributors to the papers they assist enhance and validates their function in advancing scientific information.

May you inform us extra about your proposal for this new two-way peer assessment methodology?

Our proposed two-way peer assessment system makes one key change to the present course of: we cut up assessment launch into two phases.

The authors’ proposed modification to the peer-review system.

At the moment, authors submit papers, reviewers write full opinions, and all opinions are launched without delay. In our system, authors first obtain solely the impartial sections – the abstract, strengths, and questions on their paper. Authors then present suggestions on whether or not reviewers correctly understood their work. Solely after this suggestions can we launch the second half containing weaknesses and scores.

This method gives three foremost advantages. First, it’s sensible – we don’t want to alter present timelines or assessment templates. The second section might be launched instantly after the authors give suggestions. Second, it protects authors from irresponsible opinions since reviewers know their work will likely be evaluated. Third, since reviewers sometimes assessment a number of papers, we are able to observe their suggestions scores to assist space chairs determine (ir)accountable reviewers.

The important thing perception is that authors know their very own work greatest and may rapidly spot when a reviewer hasn’t correctly engaged with their paper.

May you speak in regards to the concrete reward system that you simply recommend within the paper?

We suggest each short-term and long-term rewards to handle reviewer motivation, which naturally declines over time regardless of beginning enthusiastically.

Quick-term: Digital badges displayed on reviewers’ tutorial profiles, awarded primarily based on writer suggestions scores. The objective is making reviewer contributions extra seen. Whereas some conferences listing prime reviewers on their web sites, these lists are arduous to seek out. Our badges can be prominently displayed on profiles and will even be printed on convention identify tags.
Instance of a badge that might seem on profiles.

Lengthy-term: Numerical metrics to quantify reviewer influence at AI conferences. We advise monitoring measures like an h-index for reviewed papers. These metrics may very well be included in tutorial portfolios, just like how we presently observe publication influence.

The core thought is creating tangible profession advantages for reviewers whereas establishing peer assessment as an expert tutorial service that rewards each authors and reviewers.

What do you suppose may very well be a number of the execs and cons of implementing this technique?

The advantages of our system are threefold. First, it’s a very sensible answer. Our method doesn’t change present assessment schedules or assessment burdens, making it simple to include into present programs. Second, it encourages reviewers to behave extra responsibly, understanding their work will likely be evaluated. We emphasize that almost all reviewers already act professionally – nevertheless, even a small variety of irresponsible reviewers can significantly injury the peer assessment system. Third, with ample scale, writer suggestions scores will make conferences extra sustainable. Space chairs may have higher details about reviewer high quality, enabling them to make extra knowledgeable selections about paper acceptance.

Nevertheless, there’s sturdy potential for gaming by reviewers. Reviewers would possibly optimize for rewards by giving overly constructive opinions. Measures to counteract these issues are positively wanted. We’re presently exploring options to handle this challenge.

Are there any concluding ideas you’d like so as to add in regards to the potential future
of conferences and peer-review?

One rising development we’ve noticed is the growing dialogue of LLMs in peer assessment. Whereas we consider present LLMs have a number of weaknesses (e.g., immediate injection, shallow opinions), we additionally suppose they are going to ultimately surpass people. When that occurs, we’ll face a basic dilemma: if LLMs present higher opinions, why ought to people be reviewing? Simply because the fast rise of LLMs caught us unprepared and created chaos, we can’t afford a repeat. We must always begin getting ready for this query as quickly as doable.

About Jaeho

Jaeho Kim is a Postdoctoral Researcher at Korea College with Professor Changhee Lee. He acquired his Ph.D. from UNIST below the supervision of Professor Seulki Lee. His foremost analysis focuses on time sequence studying, notably creating basis fashions that generate artificial and human-guided time sequence knowledge to scale back computational and knowledge prices. He additionally contributes to bettering the peer assessment course of at main AI conferences, along with his work acknowledged by the ICML 2025 Excellent Place Paper Award.

Learn the work in full

Place: The AI Convention Peer Evaluation Disaster Calls for Creator Suggestions and Reviewer Rewards, Jaeho Kim, Yunseok Lee, Seulki Lee.