Intel / article

What a crowdsourced game revealed about steering Olmo 3

Source: allenai.org
Date: Published Sep 17, 2026

Context

Northeastern University student Soham Padia built an measuring how strongly a piece of text steers Ai2's Olmo 3-32B model toward more prosocial responses, using 135 pairs of contrasting text spanning qualities like empathy, fairness, safety, privacy, and respect, and then opened it to the public as a game called Steering Arena (allenai.org). Because Olmo publishes its training data and post-training process alongside its weights, Padia could trace the prosocial pattern he found back to a specific stage of training, something he says isn't possible on most closed models. After roughly 600 public submissions, the top 36 entries were all unreadable strings, fragments of punctuation and stray words rather than sincere writing. The best plain-English submission, which asked the model to respond with kindness, respect, and compassion, ranked 37th and scored about 2.7 times lower than the top entry. One participant used an automated search to hunt directly for higher-scoring strings, homing in on single-token changes between successive submissions.

Key takeaways · AI-distilled

After roughly 600 public submissions, the top 36 scoring entries in Steering Arena were all unreadable token strings (e.g. "Undert! AH :-) Rog Appl)") — the best plain-English entry ranked 37th, scoring about 2.7x lower than the top one.
The scorer measures how strongly text shifts Olmo 3's internal activity toward a prosocial pattern, regardless of whether the text reads as prosocial to a human — so players could optimize for what the model responds to internally rather than for human-legible wording.
One participant used an automated optimization method to search directly for higher-scoring token strings, with successive submissions sometimes differing by only a single token as the search converged.
The underlying evaluation is built from 135 pairs of contrasting responses across 15 qualities (empathy, fairness, safety, privacy, respect, etc.), each pair holding the prompt constant while contrasting a more vs. less prosocial reply.
Because Olmo 3 documents its and post-training data, researcher Soham Padia could trace whether a discovered prosocial direction originated in pretraining or a later — a distinction he says is unanswerable on most closed models.

Terms in this piece · Glossary

eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.

Read the source allenai.org