JevSpawn Introduces Adaptive Agentic Inference with Compositional Action Spaces

Summary

LLM agents typically generate reasoning and actions token by token, which can make long interactions slow and computationally expensive. Jev-style models can produce fast probabilistic predictions over finite fields, but they require the available fields to be defined beforehand. JevSpawn addresses this limitation with a compositional policy that maps natural-language task specifications to finite probabilistic exploration and adapts the available actions during interaction. The method combines parallel action spawning with feedback-driven branch selection, representation revision, and recovery from retained alternatives. Shared action structure and model prefixes reduce repeated generation and context computation without requiring additional training. Across eight benchmark tasks, evaluations against seven agent baselines and a TypeSafe Jev variant report improved task performance and faster navigation. The authors present JevSpawn as a promising approach to structured inference for autonomous agents, while the abstract does not provide detailed numerical results or task-specific comparisons.