TodoEvolve: Learning to Architect Agent Planning Systems

2026

Read TodoEvolve: Learning to Architect Agent Planning Systems (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-07839.txt.

Description

Agent planning usually relies on fixed, hand-crafted structures that cannot adapt to the diversity of open-ended tasks. Jiaxi Liu and seven co-authors offer TodoEvolve, a meta-planning approach that synthesizes and revises task-specific planning architectures on the fly.

They first build PlanFactory, a modular design space unifying planning patterns around topology, initialization, adaptation, and navigation. From planning trajectories in that space they train Todo-14B with impedance-guided preference optimization, rewarding plans that are performant, stable, and token-efficient across tasks and agent backbones. Tests show the learned planner adapts where fixed structures falter. Agent builders gain a route from hand-designed planning to learned, evolving planning.

Catalog record

YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-07839

Browse the full library · About the Tower of Babel