Nuclear Physics · Benchmark
Given an unordered ~2 microsecond window of ePIC central-tracker hits containing on average one buried DIS physics event on top of continuous beam/synchrotron background, reconstruct the tracks belonging to that event (the task) and tag which reconstructed objects are signal rather than overlay (the task's other half), scored against the two metrics defined in Section 3 (background hit rejection factor, >~2x or more; signal track retention rate, ~100%) while satisfying fixed system constraints: <=0.6 ms end-to-end latency and an approximate ~1.4 Tb/s (~171 GB/s) sustained input data rate computed from the ~5,700-hit/window, 15-feature-per-hit input over the 2 microsecond window.
The EIC's innermost silicon tracking layers face synchrotron/beam-gas background occupancy roughly three orders of magnitude above the true deep-inelastic-scattering (DIS) signal rate, and ePIC's streaming readout has no hardware trigger to filter this inline. This benchmark asks a model to jointly (a) group an unordered set of tracker hits into tracks belonging to the one true signal event in the window, and (b) tag signal vs. background hits/tracks using relative-timing coincidence (not geometry alone), while satisfying a hard deployability budget: <=0.6 ms end-to-end latency and an approximate ~1.4 Tb/s (~171 GB/s) sustained input data rate (computed from ~5,700 hits/window x 15 features/hit over the 2 us window -- see README.md Section 1 for the full derivation and its assumptions). Its two metrics are the background hit rejection factor (>~2x or more) and the signal track retention rate (~100%). The baseline solution is a Kalman Filter tracking algorithm: the rule-based (non-AI) central tracking in EICrecon (ACTS seeding + Combinatorial Kalman Filter, public, LGPL-3.0); it does not meet the latency/bandwidth constraints and serves as the accuracy baseline. It is UNCONFIRMED: not yet built or run for this benchmark, so no reference results exist. HEPTv2 (an LSH point transformer adapted from CMS L1-trigger HEPT; slide-deck results per-hit AUROC 0.992, Double Majority 0.597, 211x fewer bit-operations) is an AI comparison entry. This is a working draft: canonical dataset splits are not yet frozen -- see README.md for the full, evidence-cited gap list.
System constraints: Streaming data path: bandwidth and latency constraints (diagram)
@misc{eic_streaming_tracking_benchmark_2026,
title = {EIC Streaming Tracking Benchmark (working draft)},
author = {{Agentic AI for EIC (Purdue/FNAL/LANL/MIT/NJIT)}},
year = {2026},
howpublished = {DOE Genesis Mission Phase I project materials, Indico category 21857},
url = {https://indico.cern.ch/category/21857},
note = {No formal citation exists yet; cite the Indico category and the reference-solution talk by J. Schulte (26 Aug 2026) until one is published.}
}
| Category | Rating | Notes |
|---|---|---|
| Software | 1.00 / 5 | Reference (EICrecon) is public but UNCONFIRMED/not run; no benchmark run recipe, not containerized, no pinned environment file. |
| Specification | 5.00 / 5 | None. |
| Dataset | 2.00 / 5 | Not yet accessible via an open protocol; preprocessing script not published; canonical splits not finalized. |
| Metrics | 5.00 / 5 | None. |
| Reference Solution | 3.00 / 5 | Baseline (EICrecon Kalman Filter tracking) is public and open-licensed but UNCONFIRMED: not run, no benchmark results, not documented as a run recipe; Robustness metric not evaluated. |
| Documentation | 4.00 / 5 | No academic paper about the benchmark exists yet. |
| Average | 3.33 / 5 |