Definition

Recursive Language Models (RLMs) are an architectural approach where language models recursively invoke themselves or sub-agents to decompose and solve long-horizon tasks, enabling continual learning and self-improvement loops.

Key Points

  • Central thesis of prime-intellect open superintelligence bet
  • Distinct from single-pass inference — emphasizes iterative refinement
  • Pairs with RL training infrastructure for agent improvement over time

Sources