Foundation models have the potential to provide reusable representations of complex collider events, but their extension to nuclear collisions remains largely unexplored. In heavy- and light-ion systems, the high particle multiplicity and non-trivial underlying event make full-event representation learning particularly challenging, while also providing an opportunity to learn information beyond isolated reconstructed objects. We present a first study of self-supervised particle-level learning directly on CMS heavy-ion open data, representing collision events as variable-length collections of particle-flow candidates and pretraining transformer- and state-space-based architectures through masked-particle reconstruction. The learned representations are evaluated through several downstream tasks using both experimental CMS data and controlled simulated samples. We find that improved reconstruction performance during pretraining does not systematically translate into better downstream performance, and that no single architecture is consistently favoured across the tasks considered. The present comparison is, however, limited by the available computational resources, which restrict model scale, training statistics, and systematic architecture optimization. A larger-scale study, enabled by our pending EuroHPC allocation, will allow these comparisons to be revisited under substantially improved training conditions. These results provide an initial benchmark for self-supervised learning in high-multiplicity nuclear collisions and motivate the development of scalable representations transferable across heavy- and light-ion systems.<\/p>\n","protected":false},"excerpt":{"rendered":"
Foundation models have the potential to provide reusable representations of complex collider events, but their extension to nuclear collisions remains largely unexplored. In heavy- and light-ion systems, the high particle multiplicity and non-trivial underlying event make full-event representation learning particularly challenging, while also providing an opportunity to learn information beyond isolated reconstructed objects. We present a first study of self-supervised particle-level learning directly on CMS heavy-ion open data, representing collision […]<\/p>\n","protected":false},"author":14,"featured_media":0,"template":"","meta":{"_acf_changed":false,"footnotes":""},"publication-type":[32],"publication-year":[1680],"class_list":["post-43771","publication","type-publication","status-publish","hentry","publication-type-inproceedings","publication-year-1680"],"acf":[],"publishpress_future_workflow_manual_trigger":{"enabledWorkflows":[]},"_links":{"self":[{"href":"https:\/\/lamarr-institute.org\/de\/wp-json\/wp\/v2\/publication\/43771","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lamarr-institute.org\/de\/wp-json\/wp\/v2\/publication"}],"about":[{"href":"https:\/\/lamarr-institute.org\/de\/wp-json\/wp\/v2\/types\/publication"}],"author":[{"embeddable":true,"href":"https:\/\/lamarr-institute.org\/de\/wp-json\/wp\/v2\/users\/14"}],"version-history":[{"count":0,"href":"https:\/\/lamarr-institute.org\/de\/wp-json\/wp\/v2\/publication\/43771\/revisions"}],"wp:attachment":[{"href":"https:\/\/lamarr-institute.org\/de\/wp-json\/wp\/v2\/media?parent=43771"}],"wp:term":[{"taxonomy":"publication-type","embeddable":true,"href":"https:\/\/lamarr-institute.org\/de\/wp-json\/wp\/v2\/publication-type?post=43771"},{"taxonomy":"publication-year","embeddable":true,"href":"https:\/\/lamarr-institute.org\/de\/wp-json\/wp\/v2\/publication-year?post=43771"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}