Process-Aware Reward Modeling for Automated Construction Planning
Journal
Proceedings of the International Symposium on Automation and Robotics in Construction
Start Page
2126
End Page
2133
ISSN
2413-5844
ISBN (of the container)
978-064583223-5
ISBN
[9780645832235]
Date Issued
2026-06-22
Author(s)
Singh, Akarsth Kumar
Abstract
Construction planning remains error-prone because existing automated approaches primarily assess final schedules and lack mechanisms to evaluate intermediate planning decisions. As a result, early-stage errors in task decomposition, sequencing, or duration estimation can propagate unnoticed across planning stages. This study proposes a process-aware framework for automated construction planning that embeds planning-time supervision through a construction-specialized Process Reward Model (PRM). The framework integrates sequential multistage plan generation with rubric-based Small Language Model (SLM)-as-a-Judge evaluation to generate step-level quality signals, which are used to train the PRM to predict the quality of partial planning states. Experiments on real-world building construction data show that the PRM achieves 78% agreement with human-verified evaluations while reducing evaluation cost by approximately 50× and inference time by more than 5× compared to SLM-as-a-Judge assessment. The results indicate that process-level reward modeling provides scalable and efficient quality assurance for structured planning elements such as work breakdown consistency and resource feasibility, while serving as a pre-screening mechanism for complex dependency reasoning.
Event(s)
43rd International Symposium on Automation and Robotics in Construction, ISARC 2026,22 June 2026 - 26 June 2026,Singapore
Subjects
Automated Construction Planning
Large Language Model
Process Reward Model
Process-Level Evaluation
SLM-as-a-Judge
Publisher
International Association for Automation and Robotics in Construction (IAARC)
Type
conference paper