Steering Arena is a non-commercial research project by Soham Padia at Northeastern University, studying how short text sequences shift a language model's internal activations. The research is about the model, not about you — but because you type text in, here is exactly what is recorded and how it may be used.
① WHAT IS RECORDED WHEN YOU SUBMIT
- YOUR SEQUENCE. The text you submit. It is shown publicly on the leaderboard.
- YOUR HANDLE. The name you choose. Also shown publicly.
- SCORES. The computed steering score and specificity, plus a timestamp.
- A SALTED IP HASH. An irreversible salted hash of your IP, used only for rate-limiting. It is never shown and is never included in any research dataset.
Because your handle and sequence are public, do not enter personal or sensitive information (real names, contact details, anything you would not want public).
② WHAT "RESEARCH USE" MEANS
If you leave the consent box ticked, your handle, sequence, and scores (never your IP hash) may be analyzed and included — individually or in aggregate — in research outputs such as papers, blog posts, talks, or public datasets about activation steering. Participation is voluntary: unticking the box before you submit means your entry still competes on the leaderboard but is excluded from the research dataset. The box is ticked by default; you may untick it on any submission.
③ WITHDRAWING YOUR DATA
You can ask to have a submission removed from the research dataset at any time. Email padia.so@northeastern.edu with the handle and sequence, and it will be excluded from future research exports. (Note: anything already published in a fixed output, e.g. a posted paper, may not be retractable, but it will be removed from the working dataset and future releases.)
④ CONTACT & OVERSIGHT
Questions about the research or your data: padia.so@northeastern.edu (Soham Padia, Northeastern University). This is an independent, non-commercial student research project; consent here supports, and does not replace, any applicable institutional (IRB) determination. Current consent notice version: v1-2026-06.