Speaker
Description
Standard Level-1 trigger algorithms treat collision events as statistically independent, a design choice that simplifies implementation but prevents models from leveraging short term variations in detector performance. These fluctuations can transiently distort reconstructed features and weaken the stability of fast classification algorithms. To address this limitation, we introduce Context Enrichment, a lightweight neural mechanism that augments instantaneous jet-tagger inputs with a compact summary of recent behaviour. This context vector provides a learned local reference frame reflecting short-term detector conditions, enabling the tagger to anchor each decision in a small window of history.
The method is fully modular and integrates as a drop-in addition to existing jet tagging architectures without altering their internal structure. Under controlled degradations designed to mimic realistic short term detector effects, the baseline model exhibits an accuracy reduction of ~8%, whereas the context-enriched variant limits the loss to ~2%. Hardware deployment on a Xilinx VU9P FPGA shows that the context module incurs a small latency overhead while maintaining sub-microsecond inference time, ensuring compatibility with Level-1 trigger timing constraints. These results demonstrate that incorporating an amount of recent context can substantially stabilise real time jet tagging performance without compromising hardware efficiency.