Example Reviews
- An Efficient Massively Parallel Constant-Factor Approximation Algorithm for the k-Means Problem in SODA 2026
by Vincent Cohen-Addad, Fabian Kuhn, Zahra Parsaeian
Vincent: "The set of comments provided by the human reviews and the set of comments provided by the generated reviews are disjoint. The were completely correct regarding typos and math equations. The comments were particularly helpful in making all statements very formal and buggy statements. I believe this is useful as a tool to iterate and have an external view on the paper."
review pdf
- Dynamic Algorithms for Matroid Submodular Maximization in SODA 2024
by K. Banihashem, L. Biabani, S. Goudarzi, M. Hajiaghayi, P. Jabbarzade, M. Monemizadeh
Mohammad: "My students and I were genuinely surprised by the quality of these AI-generated reviews. The identified issues are almost uniformly correct, and several are non-trivial and address points that would be exceptionally challenging for a human referee to observe. I am now firmly convinced that these reviews will be very useful to the TCS community."
review pdf
- Unbiased Insights: Optimal Streaming Algorithms for Sampling, the Forget Model, and Beyond on arXiv
by Honghao Lin, Hoai-An Nguyen, William Swartworth, David P. Woodruff
William: "We proposed a lower bound against F_p estimation in the forget model for turnstile streams. The lower bound was a reduction from a particular three person communication game. Initially we claimed that our game could be solved by distintuishing a frequency vector with support {1,1} from a frequency vector with support {3} which could be accomplished via any F_p moment estimator. The model pointed out a mistake in our calculation. As it turned out we needed to distinguish a support of {1,1} from a support of {2} which means that the analysis fails for p=1. Despite not singling out p=1 in our analysis, the model correctly identified that this failure was the key issue. We were then able to correct the analysis to handle all values of p, and updated the ArXiv version of the paper accordingly. The model also correctly pointed out several typos. It incorrectly claimed two additional errors. One was due to parsing subscripts and superscripts incorrectly, which lead the model to believe that a calculation was incorrect. The other issue was just semantic -- we claimed that our bounds were "nearly optimal", and it claimed that this was incorrect due to a polylog gap between the upper and lower bounds."
review pdf