LEAP: Learning Efficient Action Proposals for LLM Agents

Preprint

The idea

A tiny draft. A faster agent.

LLM agents spend much of their time waiting for the next decision. LEAP teaches a small model to anticipate those decisions, so a larger target model can verify several proposed action contexts concurrently and advance through a task faster.

The draft learns directly from the target's own actions, without reproducing its reasoning. The target stays frozen and verifies every committed action. A latency framework explains when the time saved outweighs the cost of drafting, verification, and tools.

LEAP overview. Target traces train a small action drafter. The drafter proposes actions, the target verifies their contexts concurrently, and verified actions are committed. Committed actions can also provide online training data.
LEAP learns a small action drafter from the target agent, then accelerates execution through concurrent verification.

How it works

Learn. Propose. Verify.

  1. Learn the target's actions

    Train a lightweight LoRA adapter on context–action pairs collected from the target agent.

  2. Propose the next steps

    The small draft predicts a short sequence of actions without generating reasoning text.

  3. Verify and move forward

    The target verifies contexts concurrently. Commit the matching prefix and the target's next action, unless the task has ended.

Citation

BibTeX 
@article{xu2026leap,
  title   = {{LEAP}: Learning Efficient Action Proposals for {LLM} Agents},
  author  = {Xu, Zhen and Zhang, Qizheng and Wan, Gerry and
             Zhu, Shang and Zhang, Ce},
  journal = {Preprint},
  year    = {2026}
}