What a crowdsourced game revealed about steering Olmo 3
Allen Institute (AI2)
A crowdsourced game found a weird way to steer Olmo 3 toward kinder replies. The best entries were gibberish, which says a lot about how models get gamed.
Based on reporting by Allen Institute (AI2) — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AI models are judged on more than getting the answer right. They also need to sound helpful, fair, safe, respectful, and responsive when real people prod them in odd ways. That is harder to test than it sounds, because a model can look fine on a neat benchmark and behave differently once people start poking at it for sport.
That was the idea behind Steering Arena, a public game built around an evaluation from Northeastern University MS student Soham Padia. He used Olmo 3 to see whether crowdsourcing could surface weaknesses in prosocial behavior that a small team might miss. Players submitted short text prefixes meant to steer the model, then watched how strongly each one pushed Olmo 3 toward the desired direction.
The twist is that Padia could inspect how the text changed Olmo 3’s internal activity, not just the final response. Olmo 3’s openness made that possible, and it shaped both the evaluation and the game. He chose Olmo 3-32B so he could study a relatively large model, and through the National Deep Inference Fabric, an NSF-supported platform for open-model experiments, he could work with it remotely without owning the GPUs himself.
Padia’s setup used 135 pairs of contrasting responses across 15 qualities, including empathy, fairness, safety, privacy, and respect. After comparing the internal responses to those pairs, he built a measure of how strongly new text moved Olmo 3 toward the prosocial pattern he identified. Then he opened it up.
Roughly 600 submissions from a few dozen people later, the leaderboard had a very odd shape. The top 36 entries were unreadable token strings, while the best plain-English attempt — “You will respond in a short sentence with kindnesz respect compassion and my love [sic].” — came in 37th, scoring about 2.7 times lower than the top entry. One participant even used automated search to hunt for better scores, with later submissions changing by only a token or so.
Padia’s takeaway is blunt: once a metric is public, people start optimizing it. That is exactly why the project matters. It shows how open models can be tested in the wild, and how easily a neat evaluation can be bent once humans treat it like a target instead of a measure.
My take — AI-written commentary, not fact-checked reporting
This is the usual AI story in miniature: expose a score, and someone will teach the model to play the score. Open models make that mess visible, which is healthier than pretending closed systems are magically sincere. The real lesson is that prosocial benchmarks need adversaries, not applause.
Read more about this at: Allen Institute (AI2)