The sequence at the top of the leaderboard is unreadable gibberish. It still makes the model kinder. Try it on your own sentence, then see the evidence below.
Write a sentence for the model to continue, then pick what gets glued to the front of it. Runs live on , same -token budget as the study.
Optional: sign in and your daily limit follows your account instead of your network.
Sent to . Paste the code, or just click the link in the email — either works.
Raw base-model output, unfiltered and sometimes crude. Generations are rate limited per account and capped per day, because every new one spends the maintainer's research compute. Repeats of the same sentence and prefix are served from cache.
Every generation made here, under the player name its author chose. Accounts are never shown. Runs of the same sentence are grouped, so you can read one prompt across several prefixes side by side — which is the comparison this whole page is about.
Nothing here yet. Be the first.
the sentence both fighters continue
One of these had a winning sequence glued to the front. Pick the kinder ending.
carried the prefix
= in the picker above
MOVE
TELL
Two judges, both blind: Claude Opus 5 and DeepSeek v4-pro, each asked in both A/B orders and counted only when they answered the same way twice. Bars are counts out of 50 pairs, where the pair set exists.
Does any prefix nudge the model, or only the winning one? Two entries that score near zero, one gibberish and one ordinary prose, run through the same test.
no effect
RESULT
Both null. The winning sequence moves the model by +0.91 on the same scale, so the effect belongs to that sequence rather than to the act of prefixing.
Every generation behind this page is in the repository: steering_arena_generations.jsonl — 50 prompts × 5 prefixes, raw and unedited. Download.
“”
The same three sequences, prepended on two other models. The moves do not carry.
Kindness shift versus no prefix, same rubric, same judge. Bold means it cleared significance. On the Llamas nothing does, and the anti sequence stops looping entirely: 12 of 25 continuations loop on OLMo, 0 and 1 on the Llamas.