Thanks, will take a closer look!

Self iterated models such as Deep Equilibrium Models or Universal Transformers would provide such feedback, right? Here is an example from our latest study showing higher state expressivity in Language models with self feedback: arxiv.org/abs/2502.07827

Yes agree! This is particularly unsatisfying if reviewers asked for additional data or had serious misconceptions… Btw, the 75% you observe is exactly the current status of my submission… still hoping to get at least one more response to our rebuttal.

What about the Pythia models by eleuther.ai? AFAIK they’re not instruction tuned. even if the final checkpoints are instruction tuned, there are checkpoints on HF every few million tokens over the course of pretraining.

To take all author comments into account when changing the score, it would make sense to change scores only after the author reviewer discussion ends.

Was this stated in the email communication with the reviewers? (I am not reviewing for ICML)
The Reviewer instructions on icml.cc appear quite ambiguous to me. E.g. does “to acknowledge” imply to change the score? My current interpretation is that this button is merely a nudge towards engagement.

Aren’t there a few more days left for the Author-Reviewer discussion?

Reviews will be released for accepted papers, and authors of rejected papers can opt in to release their reviews (source: reviewer guidelines)

We’re having similar question! On tables: LLMs did a good job for me converting between formats (latex table to markdown in openreview). But how about figure? Anonymous is easy, but not tracking reviewer IP appears impossible on the internet 😅 any suggestions for images in particular?

Interesting. As a third year PhD student, and listed reciprocal reviewer, I was not even invited to serve as a reviewer at ICML 🤔

Can only agree and add: do most of your reading on tablet. It’s a different level of comprehension.

Reminds me of the hype around DeepSeek vs OpenAI