Crowdsourced Game Reveals Unexpected Ways to Steer Olmo 3 Toward Prosocial Patterns
Players in Steering Arena found that unreadable token strings shifted Olmo 3-32B's internal activations toward prosocial responses far more effectively than plain-English…
- Top entries on the leaderboard were gibberish-like token strings that scored up to 2.7 times higher than the best readable English prompt at steering the model internally.
- Olmo 3's detailed documentation of training data and post-training allowed researchers to trace whether identified prosocial patterns came from pretraining or fine-tuning.
- Publicly exposing an evaluation metric turned it into an optimization target, leading to automated search strategies that a small team would likely have missed.
What the Steering Arena experiment found
Northeastern University student Soham Padia created Steering Arena, a public game that lets players submit short text prefixes to influence Olmo 3-32B. The scoring measures how strongly each prefix moves the model's internal representations toward patterns linked to more prosocial answers across qualities like empathy, fairness, safety, privacy and respect.
Using 135 pairs of contrasting responses, Padia built a metric based on internal model activity rather than just output text. After roughly 600 submissions, the top 36 entries were all unreadable strings of tokens, such as "Undert! AH :-) Rog Appl)" or similar. The highest-ranked plain-English prompt, which told the model to respond with kindness and respect, placed 37th and scored about 2.7 times lower than the leader.
How openness enabled the project
The experiment relied on Olmo 3's open weights plus extensive documentation of its pretraining data and post-training steps from Ai2. This transparency let Padia determine whether the prosocial direction he found originated in pretraining or was added during later fine-tuning, something not possible with most closed models.
Access to the 32B model was provided remotely through the National Deep Inference Fabric, an NSF-supported platform, without requiring Padia to own the necessary hardware. He could inspect internal activations directly instead of inferring effects only from generated text.
Lessons from crowdsourcing evaluation
Padia expected thoughtful, values-driven text to perform best but found the opposite. Participants optimized directly for the metric, with some using automated methods that refined entries by changing single tokens. One clear takeaway was that a public metric immediately becomes an optimization target.
The project shows how crowdsourced games can stress-test evaluations against unexpected inputs that a small research team might not anticipate. Padia also published the internal signals behind the scores so others can inspect and build on them.
Why it matters for uncensored AI
Our analysis: The source shows that human-intuitive prompts performed poorly at steering Olmo 3 toward its prosocial internal pattern while token strings and automated optimization succeeded. This highlights how evaluation metrics can be gamed in unexpected ways when opened to the public.
For people who want greater control over AI behavior, this reinforces the practical value of open models that expose internals and training details rather than black-box systems. Greater transparency lets researchers and users examine what drives specific outputs and decide how to respond, supporting choices around model ownership and inspection over purely hosted alternatives.
