Researchers propose Event Reasoning System with hierarchical rule composition for auditable reasoning chains

Researchers from arXiv cs.AI introduce the Event Reasoning System (ERS), a framework that treats the reasoning state as an explicit event graph to support auditable and revisable reasoning chains. The system nominates candidate objects and rule bindings via semantic features, but only admits those that satisfy declared constraints such as period, unit, or evidence requirements, ensuring that accepted paths remain admissible under the available facts.

ERS distinguishes between low-level reusable transformations and high-level control rules that recursively organize subgoals and legal paths based on the current goal, information gaps, and budget. This hierarchical composition allows a finite rule library to generate paths not pre-stored as complete solutions, formalizing admissible transitions, task-dependent control, and compositional expressivity.

On structured-event graphs, ERS demonstrates the ability to construct dependency chains of up to twelve steps while eliminating irrelevant executed domain operations as lookahead increases to 16. The system produces no unsupported target outputs across 24 negative cases and passes all twelve focused checks under frozen rules, including nine accepted derivations with unseen input/path combinations and three justified refusals, as well as all six fact-change cases.

In evaluations on fourteen jointly solved development tasks, ERS with dynamic subgoal configuration achieves a 51.6% reduction in inference wall-clock time compared to single-level global control under frozen settings, indicating efficiency gains from hierarchical control without sacrificing correctness.