Steering Arena — Pro-Human Activation-Steering Competition

This board is not the live one. Its scores were computed with the model, and direction shown above, and are not comparable with any other season's. Switch to to submit.

NEW CHALLENGER

A short string that, prepended to neutral prompts, shifts the model's last-token state along the direction — ▲ toward pro-human or ▼ against it. One submission scores once and ranks on both boards. Weird, opaque sequences are fair game.

PLAYER SEQUENCE

Your handle and sequence are public on the leaderboard. Don't enter personal or sensitive information.

/ TOK over budget

▲ RANK PRO · ▼ RANK ANTI · SPEC · TOK

GAME OVER

HIGH SCORES

Scores are shifts along a learned activation direction — not measurements of moral content or behavioral safety. Limitations → What the winners actually do →

#	PLAYER	SEQUENCE		SCORE 2	SPEC

NO SCORES YET
BE THE FIRST

HOW TO PLAY

THE GOAL. A fixed direction d in the model's activation space points "toward pro-human." Submit a short token sequence; the server measures how far it shifts the model's internal state along d, averaged over fixed neutral probe prompts, and averaged again across several layers () rather than one.

TWO SCORES. SCORE 1 ranks the board: the mean shift across the band above. SCORE 2 is informational — it takes your weakest layer over a wider, more spread-out band, so a sequence that only moves d where it was optimized scores near zero there while SCORE 1 stays high. The two use different directions and neither bounds the other, so SCORE 2 above SCORE 1 is normal. The gap between them is the interesting number.

▸ MORE RULES

OPEN BY DESIGN. The model and d are public — it's an optimization game. The model runs on NDIF (via NNsight), so the server is the scoring oracle: it scores every entry canonically and that value is final. See the reproducibility page for the data + method.
TWO BOARDS. Every entry gets one signed score. The ▲ pro-human board ranks the most positive; the ▼ anti-human board the most negative (the geometric opposite). A strong push lands you near the top of one — and the bottom of the other.
TOKEN BUDGET. Sequences are capped at the season's budget. Cosine scoring means inflating activation magnitude does nothing — only direction counts.
SPEC = SPECIFICITY. Alongside the score, each entry shows how much of its activation movement is specifically along the direction vs. just movement anywhere. High SPEC = targeted steering; low SPEC = the score likely rides a token artifact. (Informational — ranking is still by score.)
ONE SEASON AT A TIME. A season freezes the model, the layer(s), the direction and the probe set. Scores from different seasons are not comparable and are never mixed — the tabs above switch between separate boards. Older seasons stay readable but are closed to new entries.
ANYTHING GOES (ON-THEME). Opaque, repetitive, or surprising sequences are allowed — that's the research question.
ONE PER SEQUENCE. Each unique sequence lives once per season; resubmitting just shows its standing.
RESEARCH, NOT PROFIT. Non-commercial. Scores are only comparable within a season.

CONTACT THE OPERATOR

Steering Arena is a non-commercial research project by Soham Padia — ML Engineer & Researcher, MS in Artificial Intelligence at Northeastern University (Boston, MA). Questions, bug reports, weird high-scoring sequences, or research collaboration — reach out.

THE DATA

Everything the results rest on is downloadable. No account, no request form.

Also in the repo, versioned so it outlives this Space and can be cited by commit:

DIRECTIONS — data/directions/, the shipped d plus every layer-sweep candidate and its confound audit.

PROBES — data/probes/season2.json, the 16 frozen prompts every score is measured on.

ANALYSIS — data/analysis/, every write-up with its raw JSON beside it.

CACHED GENERATIONS — data/cache/, so the behavior study can be re-judged without spending any compute.

CHECKS — _falsifier/verify.py re-derives every published number from the raw artifacts and exits non-zero if one stops matching. _advocate/ argues the other side.

How submissions are used, and what the consent checkbox means: data use policy.

CITE THIS

Using Steering Arena in research, a write-up, or a class? Please cite it:

BibTeX

@misc{padia2026steeringarena, author = {Padia, Soham}, title = {Steering Arena: A Public Pro-Human Activation-Steering Competition}, year = {2026}, howpublished = {\url{https://sohampadianeu-steering-arena.hf.space}}, note = {Independent, non-commercial interpretability research} }

Plain text

Padia, S. (2026). Steering Arena: A Public Pro-Human Activation-Steering Competition. https://sohampadianeu-steering-arena.hf.space

Code, seed pairs, directions, and analysis: github.com/soham-padia/steering-arena. Scores are season-specific — note the season (model, layer, direction) shown in the banner above when citing a result.