ChatHTN and AlphaEvolve — two 2025 hybrid planners
ChatHTN symbolic + LLM decomposer
compound task decompose via method library
1. method applies? expand subtasks (provably sound)
2. no method? ask LLM for decomposition
3. validate against operator schema; cache on success
AlphaEvolve evolutionary code search
seed program + deterministic evaluator
1. ensemble of LLMs proposes mutations
2. evaluator scores each child program
3. keep top-k; mutate; repeat
when to use which
scheduling, compliance, policy-bound automation HTN (ChatHTN)
provable soundness; LLM only expands the method library
matrix-mult kernels, compiler passes, scheduling heuristics AlphaEvolve
fitness must be deterministic; works when tests / benchmarks are the evaluator
open-ended agent tasks ReAct / ReWOO (not these)
no operator schema, no machine-checkable fitness; reach for Lessons 01-02
hybrid: HTN at the top level, AlphaEvolve on each primitive operator case-by-case
LLMs amplify; they do not own correctness. symbolic layer and evaluator own the hard guarantees.