Announcing Accepted Re-align Challenge Reports

Research in representational alignment converges on central questions but often diverges in its answers. Building on the success of last year’s hackathon, we introduce this year’s Representational Alignment Challenge to promote transparency, reproducibility, and collaboration in this area. The challenge provides access to efficient implementations of representational comparison measures, along with a shared benchmark and leaderboard for open and friendly competition.

We received 7 submissions, with a total of 81 and 90 attempts from the 🟦 Blue (universality) and 🟥 Red (idiosyncrasy) teams, respectively. After a light review process by the program committee, 5 blog posts were accepted. These contributions showcase creative and technically grounded approaches to understanding and challenging representational alignment.

Among this year’s submissions, we would like to highlight:

  • 🟥 Red Team — Maximizing Representational Divergence via Semantic Restriction
    Motivated by the observation that between-class structure dominates overall CKA variance, this work proposes maximizing representational divergence through semantic restriction. The method effectively suppresses the dominant variance component that inflates similarity metrics, leading to strong leaderboard performance and more meaningful comparison of representations.

  • 🟦 Blue Team — Maximizing Alignment via the Densest k-Subgraph
    This submission frames the task as a weighted Densest k-Subgraph problem and solves it with a robust optimization pipeline. The methodology is sound and empirically validated, with top leaderboard performance supporting its effectiveness. A particularly compelling insight is that lightweight CNNs—rather than Vision Transformers (ViTs)—maximize alignment, suggesting an interesting relationship between model capacity and representational convergence.

To encourage engagement, the accepted blog posts will be presented during a poster session at the Re-Align Workshop. In addition, the two winning teams will deliver featured oral presentations showcasing their challenge submissions and insights.

All accepted blog posts are available on the Blog page of the website.