Falcon: Functional Assembly and Language for Compositional Reasoning in X-ray

A new multimodal framework called Falcon aims to improve threat detection in X-ray baggage screening by reasoning about how separate components could combine to form a hazard, rather than identifying objects in isolation [1]. The system, detailed in a preprint submitted to arXiv on 24 June 2026, addresses what its authors term "compositional threat reasoning" [1][2]. In security screening, risk often arises from the functional compatibility of spatially dispersed items — such as batteries, detonators, and explosive charges — not from any single object on its own [2]. Conventional vision-language models are largely object-centric and struggle with this relational assessment [1][2]. Falcon abstracts segmentation-aware region features into a structured safety state that captures component presence, pairwise functional compatibility, and scene-level risk [2]. This structured representation is injected into the language model as an explicit intermediate interface to encourage relationally consistent reasoning [2]. The submission, authored by Yonathan Michael and colleagues, was posted as a 4,995 KB file [1]. Alongside the model, the researchers introduce Falcon-X, a benchmark that unifies dense grounding with structured supervision over component completeness and risk inference in cluttered X-ray imagery [1][2]. Experiments indicate that while existing multimodal models can adapt to appearance, they fail to produce coherent threat assessments when compositional logic is required [2]. Falcon improved functional grounding and delivered more consistent safety evaluations, establishing compositional safety reasoning as a distinct evaluation paradigm for multimodal systems [2]. The work appears on arXiv, an open-access repository for electronic preprints that has hosted scientific papers since 1991 and now receives about 24,000 submissions per month [5].