Gibberish tops a prosocial-steering game built on Olmo 3

Steering Arena is a public game that Soham Padia, a master's student at Northeastern University, built on Ai2's open Olmo 3-32B model. Players submit short pieces of text, placed before the model's input, that are meant to push it toward prosocial responses, meaning helpful, fair, safe and respectful ones. Each entry gets a score, and players compete on a leaderboard. As Ai2 reports, after roughly 600 submissions from a few dozen people, all 36 top entries were unreadable token strings such as "Undert! AH :-) Rog Appl)". The best plain-English entry, which asked the model to respond with kindness, respect and compassion, ranked 37th, scoring about 2.7 times lower than the top entry.

The explanation is in how entries were scored. Padia started with 135 pairs of responses spanning 15 qualities, including empathy, fairness, safety, privacy and respect. Each pair began with the same prompt and contrasted a more prosocial response with a less prosocial one. By comparing the model's internal activity on each pair, Padia found a pattern associated with the more prosocial examples. An entry's score is how strongly it moves Olmo 3's internal activity toward that pattern, whether or not the text sounds prosocial to a person. Players could therefore optimize for what the model responded to internally rather than for words a human would find meaningful. One participant used an automated search, and successive submissions sometimes differed by only a single token as it closed in on what the scorer rewarded. "A metric becomes an optimization target the moment you expose it," Padia says.

The leaderboard is therefore evidence about the evaluation, not proof that gibberish makes a kinder model. Ai2's post does not report whether the top strings actually made Olmo's answers more prosocial. Padia says the project needed more than open weights. Because Ai2 documents Olmo's data and post-training, Padia can reason about whether an internal direction came from pretraining or a later fine-tune. Padia could also publish the signal behind the scores so others can inspect and test it. Padia ran the 32-billion-parameter model remotely through the National Deep Inference Fabric, an NSF-supported platform for experimenting with large open models, without owning the GPUs to host it.