Tiered goals improve System One model control

System One models restrict output to decisions over user-provided multiple-choice options, trading LLM flexibility for consistent latency. Any LLM with access to logits and prompt prefilling can be adapted into this pattern by batching prompts that generate a single token with structured output, turning it into a general-purpose classifier without changing the model. In a Doom demo, Qwen3-8B reacted faster under this setup than with regular tool calls, making six or seven batched decisions every 190ms instead of one decision every 600ms or so.

Directly mapping game inputs to choices produced aimless behavior, so periodic short-term goals such as collecting armor or killing enemies were added to the fast control loop. The same choice-only interface struggled with Wikiracing because the “baseball” page had over a thousand internal links and Jev supports 255 choices for a single question. Scoring links independently performed poorly with Qwen3-8B, while tournament sampling, in which batches of links were compared and winners advanced, found the ideal three-link route through “scientific american” and “amateur astronomy” to “sun”.