LESSWRONG
LW

All of ICLR reviewer V9gX's Comments + Replies

Reducing LLM deception at scale with self-other overlap fine-tuning
ICLR reviewer V9gX10

Thanks! Apparently this is not visible to me as a reviewer. 

Then my current guess is that the area chair forgot to initiate the AC-reviewer discussion, or that they set the visibility wrong.

Reply
Reducing LLM deception at scale with self-other overlap fine-tuning
ICLR reviewer V9gX10

Interesting. This is all surprising to me, but I honestly don't know what to conclude from the surprise:

  • Even when logging into OpenReview, I do not see a single comment from the area chair.
  • So if what you are saying is true, then the area chair communicated their unhappiness to you, but not to us reviewers. My usual expectation is that the area chair initiates the reviewer-AC discussion, which, as far as I can tell, simply did not happen at all. (This can sometimes happen if the AC thinks no discussion is necessary, but according to what you say the AC woul
... (read more)
Reply
4Marc Carauleanu
Reducing LLM deception at scale with self-other overlap fine-tuning
ICLR reviewer V9gX50

I reviewed this work anonymously for ICLR under the reviewer name V9gX:

https://openreview.net/forum?id=q9g13IoWmk&noteId=Pd6o3qhGlw

This is the most detailed review of the technical details of the work that I'm aware of, and I still stand by those points. Though note that the authors conclude by disagreeing with me and I did not answer since the reviewing period was over. The answer did not convince me, but it might convince you.

I may not find the time to engage extensively here, I simply link to the review since I think it makes important points that are not made in this forum by anyone else. 

Reply
1Marc Carauleanu
Thanks for sharing the discussion publicly and engaging with our work! We seem to disagree on what the claims being made in the paper are more than anything. For what it's worth the area chair has claimed that they were not happy with the way the discussion unfolded and that the reviewers that were on the fence did not respond after addressing their points which could have changed the final decision.