acceptodds
Desk rejected at ICLR 2027

Scorer, Search, or Interface? Decision Boundaries for Symbolic Planners

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Maintained symbolic planners evolve around interfaces that can remain fixed while learned scoring and deployment-time search change. This setting calls for a comparison object that separates the interface contract, the shared planning wrapper, the scorer, and the controller budget. We formulate such an object and introduce RavelWM, an action-conditioned dynamic-graph scorer built from interleaved graph propagation and selective per-node state-space updates. The design carries node state across time, handles graph birth and disappearance explicitly, predicts next-state attributes and candidate edges, and exposes per-constraint risk to a shared plan-ranking rule. We also specify a protocol-local structured-versus-raw fallback threshold and a condition registry that makes an eventual scorer, controller, or interface comparison auditable. The resulting framework provides a precise basis for studying maintenance choices in long-horizon symbolic planning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
adaptive reasoning and inference, Scorer, Search, or Interface? Decision Boundaries for Symbolic Planners

Related papers

Discussion (0)

Sign in to comment.