The Trillion-Parameter RL Paper Is Really About Letting the Model Find the Workflow

Short summary

An arXiv paper trains a 1T-parameter MoE reasoning model using zero RL (no human chain-of-thought examples) with verifiable rewards, reaching 84.2% pass@1 on AIME 2026. The key insight is that reasoning behaviors like self-verification and structured formatting emerge from training dynamics rather than prompt engineering, suggesting scaffolds are becoming training artifacts. The paper's tiered inference modes (4k/16k/64k token budgets) offer a practical knob for production cost management.

• •1T-parameter MoE model trained with zero RL reaches 84.2% on AIME 2026 without human reasoning traces
• •Reasoning behaviors like self-verification and parallel reasoning emerge from RL training dynamics, not prompt scaffolding
• •Tiered inference modes let the model adapt reasoning depth to problem difficulty, controlling production costs

Generated with AI, which can make mistakes.