We recommend referring to the ICLR version.
Due to internal company policy restrictions, we are temporarily unable to release the source code. However, we will support the community in reproducing this work, and we also appreciate the excellent reproductions that some community members have already created, such as this repository(https://github.com/alexfdom/attention-influence).
NIPS-2025(https://openreview.net/forum?id=kqfZWhveTP) Rating 3/4/4/5 rejected
ICLR-2026(https://openreview.net/forum?id=xPxoZuosIe) Rating 2/4/4/6 rejected
[URGENT ICLR 2026] Response to leaked reviewer/AC identities
Background
On Nov 27, ICLR was notified of a software bug that leaked the names of authors, reviewers, and area chairs for all ICLR submissions. The bug impacted all conferences hosted on OpenReview, and we are grateful to the OpenReview team for fixing the issue quickly. More information about this bug and its mitigation is available in OpenReview's official statement.
Because this incident enabled deanonymization of all papers and reviews, it significantly increases the possibility of collusion attempts, such as authors intimidating or bribing reviewers to raise their scores. After the public disclosure of this bug, we began to receive reports of collusion attempts and additionally observed an anomalous uptick in reviewers increasing their scores.
Response
We understand that this incident deeply impacts our community, and we take your concerns seriously. We are taking the following actions:
To avoid collusion, we are preventing reviewers from changing their scores or participating in discussion going forward. Due to uncertainty around when the exploit was introduced, we are also reverting all reviews and their scores to their state before the start of the discussion period. We are continuing to allow authors to post responses and comments until the end of the rebuttal period.
Additionally, we will be assigning a new area chair to every submission. The new area chair will be tasked to carefully consider how the author response addresses reviewer concerns when deciding on their accept or reject recommendation. We recognize that it will be frustrating to authors and reviewers to have reviews reverted after the initial rebuttal process, and we additionally recognize that we are asking area chairs to perform additional work, but we believe this is the best course of action to prevent widespread collusion that could undermine the scientific integrity of ICLR.
Reviewer Scores:
I would say that, if the reviewers had engaged more deeply, the scores could have improved from 4-6-2-4 to something like 4-6-4-6, as the authors addressed quite a few of the concerns. Unfortunately, that does not seem to have happened. I believe the authors can incorporate all of the feedback to significantly strengthen their next submission.