XDomainBench Diagnoses Reasoning Collapse in LLMs for Science

A new benchmark, XDomainBench, reveals that large language models (LLMs) suffer from systematic reasoning collapse when composing knowledge across scientific disciplines. The benchmark, introduced in a paper on arXiv (2605.14754), simulates interactive interdisciplinary scientific workflows with 8,598 sessions across 20 domains and 4 task categories. It formalizes composition order and mixture structure to stress-test models from single-discipline to interdisciplinary reasoning. Evaluations show that as composition order increases, LLMs exhibit a collapse in reasoning, attributed to two root causes. The study aims to address the gap in existing benchmarks that focus on single-turn scenarios, failing to capture the complexities of real-world AI4S (AI for Science) applications.