DrillBench: A Sequential Decision-Making Benchmark for Drilling Operations with Cross-Stage Agent Feedback

Abstract

Abstract Drilling operations require engineers to make sequential, safety-critical decisions under partial observability, physical constraints, and noisy sensor data, yet most existing machine learning approaches treat drilling intelligence as isolated prediction tasks and fail to capture the multi-stage reasoning process used in practice. We introduce DrillBench, a sequential decision-making benchmark that formalizes drilling operation understanding as a structured multi-stage reasoning problem, constructed from 2,155 validated operational episodes derived from a full geothermal drilling campaign and explicitly modeling hypothesis formation, interactive evidence acquisition, physics-aware assessment, control recommendation, and cross-stage dependency. To address the challenges posed by this benchmark, we propose a DrillBench Agent that integrates physics-informed reasoning with an explicit cross-stage feedback mechanism, enabling revision of incorrect early hypotheses when later evidence contradicts prior beliefs. Extensive experiments across strong Large Language Models (LLMs) show that the proposed agent consistently outperforms single-pass and non-feedback baselines, particularly in ambiguous operational regimes where early errors would otherwise propagate, while targeted ablation studies confirm that sequential dependency, interactivity, physics grounding, and feedback are all critical to robust decision-making. Together, DrillBench and the proposed agent framework provide a principled foundation for evaluating and developing AI systems capable of reliable sequential reasoning in real-world drilling operations.