Task planning for LLM agents must produce workflows that honor user intent while respecting complex sub‑task dependencies. Existing planners excel on sequential or DAG‑like structures but falter on patterns such as verification‑correction loops, convergent branch merges, and reusable intermediate states that naturally arise in real‑world tool orchestration.
TopoPlanner introduces a topology‑consistent framework that lifts tool dependency graphs into cellular workflow complexes and feeds them as topology‑aware context to an LLM planner. The pipeline consists of three steps: (1) cosheaf‑consistent cellular retrieval extracts a request‑relevant closed subcomplex from the global complex; (2) multidimensional structural reasoning operates on the subcomplex to capture loops, merges, and loop‑merge topologies; (3) the resulting cellular representation is converted into a prompt for the planning LLM, which then generates a tool sequence that respects the identified topology.
Experiments on four benchmarks featuring topology‑guided loop, merge, and loop‑merge workflows show consistent gains over prompt‑only and graph‑enhanced baselines across various local LLM backbones.
Review