Crowdsourced Game Finds Garbled Strings Like 'Undert! AH :-)' Steer Olmo 3 Prosocial
allen_ai · x · 2026-09-18
Ai2 details Steering Arena, a crowdsourced game built on the fully open Olmo 3-32B via NSF-backed NDIF: players submit text prefixes to elicit kinder model responses, revealing that garbled strings like 'Undert! AH :-) Rog Appl)' are surprisingly effective—and exposing eval gaps small teams would miss.
Related event: Ai2's Steering Arena: Gibberish Prompts Make Olmo 3 Kinder(4 posts)→
More from Research
- Researchers: RL training has made models' theory of mind 'terrible' — voooooogel · 2026-09-18
- The 272-page Multimodal Deep Learning book is free to read online with an AI tutor — burkov · 2026-09-18
- Therna launches Chronos RNA foundation model trained across ~50 human cell lines — LightningAI · 2026-09-18
- Arvind Narayanan's ICML Keynote: RSI Is Not the Singularity, and No Lab Milestone Kills All Jobs — random_walker · 2026-09-18
- Does RL build new representations or just make models flail? LLM researchers debate — doomslide · 2026-09-18
- Statistician argues alignment is impossible: LLM outputs are distributional, not fixed traits — gerardsans · 2026-09-18