Investigating 1-bit quantization in transformer-based top tagging

The increasing scale of deep learning models in high-energy physics research has posed significant challenges for their deployment on low-power, latency-sensitive platforms, such as FPGAs and ASICs used in trigger systems, as well as CPUs and GPUs employed in offline data reconstruction and processing pipelines. In this context, top-quark tagging, a key task in precision measurements and searches for new physics at the Large Hadron Collider (LHC), provides a representative benchmark due to its stringent latency and accuracy requirements under extremely large data volumes. In this work, we introduce BitParT, a 1-bit Transformer-based architecture designed specifically for top-quark tagging. Building upon recent advances in ultra-low-bit large language models, we extend these ideas to the collider-physics domain by developing a binary-weight variant of the Particle Transformer (ParT) model. Our results indicate that this approach enables a substantial reduction in model size and computational complexity while maintaining high tagging performance. We benchmark BitParT on the public Top Quark Tagging Reference Dataset and show that it achieves performance competitive with its full-precision counterpart. This work demonstrates that extreme quantization can provide a practical route toward real-time inference in collider experiments with minimal and optimized resource usage.