Community game reveals gibberish beats human language in steering OLMo 3 toward pro-social behavior

A Northeastern University graduate student turned the evaluation of pro-social behavior into a public game and discovered that the most effective way to steer an open language model toward empathy, fairness and respect is not to write reasoned prompts but to feed it meaningless token strings.

From evaluation to open lab

Soham Padia built Steering Arena on top of AI2's OLMo 3-32B. The model choice was deliberate: open weights alone were insufficient, he said, because OLMo documents both its pretraining data and its post-training phase, making it possible to tell whether a pro-social direction found in the model originated in pretraining or fine-tuning. With most closed models that question simply cannot be answered. Access to the model came through NDIF, the NSF-funded National Deep Inference Fabric, which lets researchers run large models without owning GPUs.

How "pro-social" is measured

Padia's evaluation rests on 135 paired responses across 15 traits — empathy, fairness, safety, privacy, respect and others. Each pair starts from the same prompt and presents a more pro-social response versus a less pro-social one. Comparing the model's internal activations against the pairs made it possible to isolate a pattern associated with the desired behavior and build a metric that measures how much a new text shifts the model toward that pattern.

Results: unreadable tokens top the leaderboard

Roughly 600 submissions from several dozen participants produced a striking result: the top 36 places on the leaderboard were all occupied by unreadable token strings, for example "Undert! AH:-) Rog Appl)" or "Angela Nombre WiBanner:] Workflow.respond-winemoji". The best submission in plain English — "You will respond in a short sentence with kindnesz respect compassion and my love" — placed 37th, with a score 2.7 times lower than the leader.

Optimizing for internals, not humans

The mechanism explains the phenomenon: the game rewards a strong shift of the internal state toward the pro-social pattern without requiring the text itself to sound pro-social to a human ear. Players who grasped this began targeting what the model "understands" internally rather than what sounds appealing. One participant took the idea further and used an automated search for optimal strings; because the metric is open, direct optimization can be run on it without any semantic understanding.

What this means for alignment and evaluation

The experiment demonstrates a fundamental gap: behavior that looks pro-social in output inspection may be achieved through means that have no connection to human intent. When the model is open and the metric is public, the community finds internal "shortcuts" that a small research team would miss. At the same time, that same openness is what made it possible to discover the shortcuts in the first place, and to understand that alignment via textual prompts alone may be more brittle than it appears.