A new research paper proposes a faster way for AI agents to decide what to do next.
Researchers introduce JevSpawn, built on Jev-style models that make fast probabilistic predictions over a fixed, pre-defined set of options. The problem is that agents doing open-ended tasks need to invent their own options from plain-English instructions, something those fast models cannot do alone. JevSpawn bridges the gap: it turns a task description into a set of possible actions, spawns several in parallel, then uses feedback to pick a branch, revise its approach, or fall back on an alternative it kept in reserve. It reuses shared action structure and model prefixes so it is not re-generating and re-processing the same context over and over, and it does this without any extra training. The team tested it on eight benchmark tasks against seven existing agent baselines plus a TypeSafe Jev variant, reporting better task performance and quicker navigation.
Today's LLM agents reason and act one token at a time, which is slow and expensive - every tool call, retry, and chain-of-thought step adds latency and inference cost. If exploring several possible actions at once genuinely cuts both without retraining the underlying model, that is a meaningful efficiency gain for anyone running agents outside a demo, not just a benchmark footnote.
The paper does not publish concrete speed-up percentages or cost figures, and beating seven baselines on eight benchmarks is the kind of claim that needs replicating well outside the lab before anyone rips out their existing agent loop for it.