Towards a Self-Driving Trigger at the LHC: Adaptive Response in Real Time

Motivation

• Collider filtering must operate under extreme throughput, bandwidth, computation, and storage requirements.The stated scales include 10^6–10^9 events per second, 10^12–10^15 bits per second, and 10^15–10^18 bytes per year.Source: §1 · pp. 2–3

Self-driving trigger

• The self-driving trigger continuously learns what to collect, responds to changing observations, and explains save-or-discard decisions through hardware- and resource-aware updates.The framework builds on reinforcement learning, explainable AI, and control optimization.Source: §1 · p. 3

Motivation

• Static trigger menus require extensive manual intervention and remain inflexible under changing instrumentation, data-taking conditions, and newly emerging features.Such updates can take weeks or months, potentially reducing sensitivity to rare or interesting data.Source: §1 · pp. 3–4

Self-driving trigger

• A cost-aware agent interacts with the hardware-based L1 trigger in a feedback loop, using recent decisions, rates, correlations, computation costs, and labeled Monte Carlo events to prescribe updated rules.The agent updates thresholds and bandwidth allocations for subsequent event batches while estimating efficiency for selected signatures.Source: §1 · p. 4

Benchmark

• The benchmark evaluates adaptive control with CMS Open Data and Monte Carlo simulations using both a traditional energetic-event trigger and an unsupervised neural-network anomaly detector.The task focuses on selecting events with interesting jet activity while covering both known-signal selection and non-standard event topologies.Source: §1 · p. 5

Results

• The demonstrated adaptive system can prioritize known-signal efficiency, anomaly sensitivity, or computational burden according to operational objectives.The authors present this as a proof of concept for a robust alternative to the static status quo.Source: §1 · p. 5

• CMS Open Data studies use 16.4 fb−1 of 13 TeV proton-proton collisions from Run 2 periods G and H, with reconstruction matching 2016 pileup, beam-energy, and detector-calibration conditions.Source: §2 · pp. 5–6
• The benchmark uses HT and anomaly-detection triggers to represent heuristic high-energy selection and searches for poorly modeled processes.HT selects large scalar sums of jet transverse momenta, while AD targets events deviating from known backgrounds.Source: §2 · p. 6
• The framework validates adaptive trigger control with CMS collision data and simulation in a reproducible evaluation setup.CMS Open Data provide detector-condition variations, while simulated samples support interpretation and strategy design.Source: §2 · pp. 5–6
• The simulation library includes simulated minimum-bias collision events, top-quark pair production as a representative Standard Model process, and exotic Standard Model Higgs decays to new pseudo-scalar particles.Source: §2 · p. 7
• Collision data show that NPV and HT evolve during a run, with HT shifting toward lower values as pileup decreases over the fill.The observed HT changes in scale and shape demonstrate its time dependence and motivate adaptive trigger control.Source: §2 · pp. 7–8, Figure 2 · p. 8, Figure 3 · p. 8
• Simulated events are ordered by NPV and smeared with Poisson fluctuations to create time-dependent trends resembling those observed in collision data.This construction reproduces pileup evolution and the associated decrease in hadronic activity over an LHC fill.Source: Figure 4 · p. 9, §2 · p. 9
• The benchmark configures the L1 menu to pass 99.75 percentile of background events, while a controller dynamically adjusts HT and AD thresholds as a function of time.Simulated minimum-bias and real ZeroBias events form the background streams; t¯t and h →4b samples are reserved for evaluating the evolving menu.Source: §2 · pp. 9–10
• With simulated backgrounds, signal events are ordered using the same NPV-based time indexing.Source: §2 · p. 10
• With real-data backgrounds, each data event is randomly matched to signal events sharing the same NPV.Source: §2 · p. 10
• The benchmark progresses from simulated-background development and validation to real-data evaluation under real experiment conditions.The initial stage develops framework and controller strategies, while the later stage evaluates methods using matched signal samples.Source: §2 · p. 10

3. FEATURES OF THE TRIGGER DATA AND ALGORITHMS

• The benchmark compares an HT trigger based on a simple energy threshold with an AD trigger that selects events using machine-learned anomaly scores.HT targets energetic hadronic events but suffers pileup dependence, whereas AD is designed to capture rare or previously unknown collision features with minimal model bias.Source: §3 · pp. 10–11
• Both algorithms use up to eight L1 jets described by η, ϕ, pT, and NPV, enabling direct comparison between supervised and data-driven trigger approaches.The feature distributions include HT, jet multiplicity, jet pT, and Hmiss_T across real and simulated samples.Source: §3 · p. 11

3.1. Anomaly Detection Algorithm

• The AD model is a compact fully connected autoencoder that reconstructs a 25-dimensional event representation and uses mean squared reconstruction error as its anomaly score.It learns background structure from real collision data or minimum-bias simulation, while NPV helps model pileup-dependent detector response.Source: §3.1 · p. 12
• Models with smaller latent dimensions show higher signal efficiency and larger exclusive contributions relative to the HT trigger, motivating adoption of the d = 2 configuration as the baseline.The d = 2 configuration is selected for consistent performance, lower resource requirements, and resilience to noise.Source: §3.1 · p. 14
• At comparable L1 rate reduction, the AD trigger recovers t¯t and h →4b signal efficiency below the HT threshold, where HT rejects essentially all signal events.Operating AD alongside HT retains sensitivity at lower energy scales, demonstrating complementarity to energy-sum selection.Source: §3.1 · p. 14, Figure 8 · p. 15
• Signal efficiency improves in the lower pT kinematic region when using the anomaly-detection algorithm.Source: §3.1 · p. 15
• Improved separation in the high-score tail translates directly into higher signal acceptance.The high-score tail is identified as the most relevant region for signal efficiency.Source: §3.1 · p. 15
• The anomaly-score distributions show clear separation between background and the t̄t and h →4b signal samples.This separation is reported for models trained on simulation and real data.Source: Figure 9 · p. 15, §3.1 · p. 15

• The fixed-menu study evaluates HT and anomaly-detection triggers by tracking their rates and signal efficiencies across evolving pileup conditions.Thresholds determine whether events are saved or rejected, using either HT or an anomaly-detection score.Source: §4 · p. 16
• Significant rate drifts occur under a fixed menu as detector conditions evolve.The menu thresholds are fixed using the first few data batches and then applied without control mechanisms.Source: §4 · p. 16
• The rate target is at most 100 kHz from background events under an assumed 40 MHz collision rate.This target is treated as the maximum tolerable rate given bandwidth, storage, and available-resource constraints.Source: §4 · p. 16
• HT rates decrease significantly as NPV decreases over time, reflecting HT sensitivity to changing pileup conditions.Fewer interactions per bunch crossing alter event characteristics and the trigger response.Source: §4 · pp. 16–17
• Static menus increasingly miss valuable signal events as detector and beam conditions evolve without adaptation.The observed efficiency loss exposes a practical limitation of time-invariant trigger thresholds.Source: §4 · p. 17, Figure 11 · p. 17

Results

• Static thresholds degrade performance as pileup decreases, whereas the adaptive controller stabilizes background rates within tolerance and mitigates drift.The corresponding background-rate behavior is shown for HT and AD triggers under PD control.Source: §5 · pp. 18–19, Figure 13 · p. 19

Feedback control

• The controller adjusts trigger thresholds in response to deviations from the target event rate, enabling dynamic regulation across changing conditions.Observed deviations in one data batch inform threshold adjustments in the subsequent batch.Source: §5 · pp. 17–18

PID formulation

• A PID controller extends linear feedback with proportional, derivative, and integral corrections, while the regulated parameter is the threshold of either the HT or AD trigger.The toy model targets a fixed background rate of 100 kHz.Source: §5 · p. 18

Controller selection

• Comparisons across controller gains find no significant benefit from integral feedback, making the simpler PD loop sufficient for dynamic menu regulation.The additional improvement from derivative feedback is relatively modest.Source: §5 · p. 18

6. MULTI TRIGGER CONTROL FRAMEWORK

• PD-controlled menus maintain higher and more stable signal efficiencies than fixed-threshold strategies, improving accumulated signal events by up to 70%, especially for pileup-dependent triggers.The section contrasts PD control with fixed thresholds while motivating a more general multi-trigger strategy.Source: §6 · pp. 19–20
• The proposed multi-trigger framework minimizes a dynamically assessed global cost function whose weighted constraint terms balance competing objectives without generalized PID gain tuning.Constraints can represent requirements such as background-rate limits, with weights encoding relative importance or acceptable deviation scales.Source: §6 · p. 21

6.1. Case 1: Background-Signal Tradeoff

• In Case 1, an ideal controller holds the combined HT and AD background rate near its target while prioritizing signal acquisition and maintaining high efficiencies for h →4b and t¯t.The cost balances background-rate stability against total signal efficiency, with σb = 4 kHz and σs = 0.05 in the example configuration.Source: §6.1 · p. 23

6.2. Case 2: Specific Path Bandwidth Allocation

• A 50 kHz (p = 0.5) exclusive allocation to the AD path heightens trigger activity for collecting both signal and background events compared with Case 1.The strategy combines exploration with continued coverage of high-value SM processes and may enforce a minimum baseline efficiency on these signals.Source: §6.2 · p. 23
• Case 2 shows that hierarchical cost priorities maintain high efficiency across both triggers while increasing the anomaly-detection contribution, which nearly fails the imposed 90% efficiency criterion.The richer objective prevents the ideal controller from satisfying every target perfectly at all times, while demonstrating the effectiveness of hierarchical priorities.Source: §6.2 · pp. 24–25

6.3. Case 3: Computational Cost Awareness

• Computational cost is treated as an optimization objective alongside physics goals because accepted events consume downstream HLT resources that could otherwise support more sophisticated trigger rules.The composite cost function balances physics performance against computational feasibility rather than minimizing computational cost alone.Source: §6.3 · p. 26
• The computational-cost benchmark uses a toy model rather than experiment-specific technical details, leaving more realistic and dynamic integration for future work.Reference costs are derived from the physics-only Case 1 configuration and can in principle be updated using historical computational-cost trends.Source: §6.3 · pp. 26–27
• The benchmark models computational load using event-level complexity and trigger-path algorithmic costs, assigning 4 units to events passed by AD and 1 unit otherwise, without double counting events passing both paths.Source: §6.3 · pp. 26–27
• The resulting computationally aware strategy maintains stable total background rates while increasing signal efficiencies over time.These outcomes are reported for the Case 3 parameter configuration shown in the supplied passage.Source: §6.3 · p. 28
• The optimization favors the HT trigger as the cheaper path, with no significant loss in total signal efficiency.The background and signal samples are dominated by the HT trigger.Source: §6.3 · p. 28
• The implemented AD model tends to collect events with lower jet multiplicities, affecting event-level cost.This could in principle compete with the trigger-level cost term.Source: §6.3 · p. 28
• The differences among the three cost-related distributions are not significant enough to favor the AD path.Consequently, the optimization remains directed toward the HT path.Source: §6.3 · p. 28

Case 3: Computational Cost Awareness

• The cost-aware controller maintains stable total background rates, improves signal efficiency over time, and allocates events toward the cheaper HT path.The results are consistent with HT being cheaper than the anomaly-detection path, which becomes dominant in event collection.Source: §7 · p. 29

Controller design

• A global grid search can approximate the optimal upcoming-batch parameters but becomes impractical as the number of trigger paths grows.The ideal controller assumes complete knowledge of the upcoming batch, establishing an upper bound rather than a deployable strategy.Source: §7 · pp. 28–29

Local controller

• Because optimal parameters evolve smoothly and remain localized, the practical controller uses a local search with memory to follow trends while limiting abrupt shifts.The implementation searches a 10 × 10 window, averages optimal local points from five previous batches, and uses a ±20-unit range.Source: §7 · pp. 29–30

Evaluation

• The simple controller follows historical trends while preventing abrupt parameter changes, producing smooth evolution through the action space.Figures 21–24 compare its cost evolution, trigger rates, and efficiencies with ideal-controller behavior across the benchmark cases.Source: §7 · p. 30

Benchmark comparison

• The realistic controller preserves rates, efficiencies, and trigger-path decomposition largely unchanged from the ideal controller across the three benchmark cases.Cumulative signal efficiency remains close to the ideal case in the corresponding benchmark panels.Source: §7 · p. 30

8. CONTROLLER PERFORMANCE WITH DATA

• The real-data study uses Run 283876 for anomaly-detector training and Run 283408 for evaluation, with 1.99 million events processed in batches of 2 × 10^4.The evaluation stream updates approximately every 12 minutes and is a proxy for evolving conditions rather than a faithful model of 40 MHz online throughput.Source: §8 · pp. 32–33

8.1. Single-Path Real-Time Control on Data

• PD control maintains the background rate within the target tolerance band as luminosity and pileup decrease, whereas fixed thresholds drift.Residual batch-to-batch fluctuations are attributed to finite-batch statistical uncertainty rather than systematic trends.Source: §8.1 · pp. 33, 35
• Controlled HT and AD menus preserve higher instantaneous and cumulative signal efficiency than fixed menus, especially late in the run.The comparison covers both the t¯t and h →4b benchmark processes.Source: §8.1 · pp. 33–34, Figure 26 · p. 34, Figure 27 · p. 35
• Up to 25% gain in integrated signal yield over a fixed menu while maintaining stable background rates throughout the run.The PD controller transfers from simulation to real data without retuning and compensates for the gradual decrease in pileup.Source: §8.1 · pp. 34–35

8.2. Multi-Path Real-Time Control on Data

• The multi-path controller adapts competing HT and AD thresholds under shared bandwidth, with cases covering symmetric control, fixed AD allocation, and cost-aware optimization.Case 3 jointly accounts for computational cost, signal efficiency, and bandwidth constraints.Source: §8.2 · pp. 35–36
• Across multi-path configurations, adaptive control stabilizes rates and improves signal performance while accommodating detector fluctuations, distribution shifts, and anomaly-detection volatility.The framework also achieves cost optimization and retains interpretable behavior on real data.Source: §8.2 · p. 36

9.1. Performance Summary on Simulated Events

• Adaptive trajectories move toward higher signal efficiency at lower computational cost as conditions evolve, with the Low-Computational-Cost agent achieving the most aggressive cost reduction at comparable efficiency.The total cost combines trigger-path and event-level contributions.Source: §9.1 · pp. 39–40
• Adaptive agents keep background-rate deviation small while improving signal efficiency, whereas the fixed menu drifts toward larger rate mismatch and poorer efficiency in simulation.The summary trajectory shows adaptive control maintaining background stability without sacrificing signal acceptance.Source: §9.1 · pp. 39–40
• Differences among adaptive agents reflect their cost-function priorities, demonstrating that the framework can encode experiment-specific objectives.The comparison includes standard, anomaly-focused, and low-computational-cost strategies against a fixed menu.Source: §9 · p. 37, Figure 32 · p. 39, §9.1 · p. 40

9.2. Deployment on Collision Data

• In CMS Run 283408, adaptive agents reduce total computational cost over time while maintaining rate control, with the Low-Computational-Cost agent reaching lower cost near the run’s end.This behavior is consistent with the simulated trend as pileup decreases.Source: §9.2 · pp. 40–41
• Collision-data trajectories show adaptive strategies moving toward higher average signal efficiency at reduced total cost, rather than reproducing simulation point by point.The efficiency–cost improvement persists in collision data and is not an artifact of simulated stream ordering.Source: §9.2 · p. 41
• Overall, adaptive control stabilizes background response, preserves or improves signal acceptance, and reduces computational burden across simulation and collision data.The qualitative agreement between MC and data supports robustness under realistic operational conditions.Source: §9.2 · pp. 41–42

10. CONCLUSIONS AND OUTLOOK

• The framework combines traditional and machine-learning trigger algorithms to collect promising events while controlling false positives and computational resources.Its evaluation focused on CMS collision data with hadronic jet activity relevant to rare or undiscovered phenomena.Source: §10 · p. 42
• The benchmark evaluates independent path control and generalized menu control, with rewards prioritizing signal collection, anomaly exploration, or reduced processing cost.A simple controller using modest, adiabatic menu changes performed near the theoretical ideal for each objective.Source: §10 · p. 42
• Adaptive trigger control achieved stable collection rates and improved efficiencies for candidate signals in simulated and historical CMS data.The observed trends in approximate-time simulation largely coincided with those seen in collision data.Source: §10 · pp. 42–43
• The study provides a proof-of-concept that adaptive trigger systems can be implemented relatively simply while improving substantially over static experimental menus.The authors release an open benchmark ecosystem and associated workflow and datasets to encourage further development.Source: §10 · pp. 42–43
• Full autonomy still requires decisions about trigger and data-acquisition integration, configuration bookkeeping, and incorporating occasional human intervention.The paper identifies more powerful reinforcement-learning and active-learning agents as future extensions.Source: §10 · p. 43
• Self-driving triggers offer a focused strategy for navigating complex constraints while maximizing sensitivity to interesting physics.The conclusion characterizes the approach as having operational advantages over legacy systems and being within modern experimental capabilities.Source: §10 · p. 43