There are two things I noticed. First, it seems that interpretability research got the lowest scores. And that the highest scored paper (444 with highest confidence) is the "Context Language Models" paper which lets models modify their context. This seems to support OpenAI latest findings.

scores by domain: https://www.reddit.com/r/MachineLearning/comments/1xpedgc/iclr_2027_review_scores_n/

scores by paper: https://shiqiangw.github.io/iclr-2027-scores/

OpenAI safety findings: https://alignment.openai.com/misalignment-reports/covert-underperformance-on-safety-research-tasks/