PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning to Trigger: Reinforcement Learning at the Large Hadron Collider

Zixin Ding, Shaghayegh Emami, Giovanna Salvi, Cecilia Tosciri, Abhijith Gandrakota, Jennifer Ngadiuba, Nhan Tran, Christian Herwig, David W. Miller, Yuxin Chen

arXiv:2606.23993v3cs.LGcs.AIhep-ex

TL;DR

The paper addresses whether trigger thresholds can be retuned online as detector conditions and background composition drift, without sacrificing signal efficiency under strict rate constraints. It trains streaming reinforcement-learning controllers with feasibility-aware GFPO variants and evaluates them on simulated and CMS collision data, reporting improved in-tolerance operation and signal efficiency, including simulation-to-real transfer.

  • Problem

    Static, hand-tuned LHC trigger menus require costly expert re-optimization and can become miscalibrated as pileup and background composition drift.

  • Method

    A reinforcement-learning policy observes streaming summaries and updates thresholds, while GFPO-F and GFPO-FR filter rollouts by background-rate feasibility and signal efficiency.

  • Results

    The framework yields consistent gains in rate stability and signal efficiency on Monte Carlo and CMS data, with sim-to-real transfer and no deployment fine-tuning.

  • Takeaways & Limitations

    Feasibility-first filtering is presented as the operative mechanism across LHC triggers and non-physics streaming benchmarks under distribution shifts.

  • Takeaways & Limitations

    CPO’s constraint guarantee concerns expected cost after a parameter-space step rather than the feasibility of candidates actually proposed by the policy.

Abstract

from arXiv · show

High-throughput scientific facilities such as the Large Hadron Collider depend on real-time event filtering (\textit{triggering}) under tight constraints on bandwidth, latency, and storage. In practice, trigger menus are largely static and hand-tuned and can become suboptimal as detector conditions, pileup, and background composition drift over time. We cast online threshold tuning as a sequential decision-making problem: a reinforcement learning agent ingests streaming summaries of recent rates and signal-sensitive features and updates trigger thresholds to maximize signal efficiency while tracking a target background rate within a tolerance band. We adapt Group-Filtered Policy Optimization (GFPO) to streaming control and introduce two variants (GFPO-F, GFPO-FR) that enforce background rate feasibility during training. On a benchmark that emulates realistic collider operation, we study two representative triggers: a total transverse energy ($H_{T}$) trigger sensitive to pileup variation, and an anomaly-detection (AD) trigger based on reconstruction loss for rare or non-standard signatures. On Monte Carlo streams, our agent increases the fraction of in-tolerance time intervals by 48\% ($H_T$) and 28\% (AD), with a cumulative gain of up to 2\% in signal efficiency on those in-tolerance intervals. Transferring from simulation to \emph{real} collision data (CMS Run 283408), the same agent, without fine-tuning, achieves a 56\% ($H_T$) and 28\% (AD) in-tolerance improvement over baselines, with further signal-efficiency gain on both triggers. To our knowledge, this is the \emph{first} demonstration of RL-based trigger control on real Large Hadron Collider collision data. Code is available at https://github.com/Zixind/GFPO_LHC (see repo for details).

1 Introduction

The paper frames online trigger-threshold tuning as streaming reinforcement learning under shifting conditions and strict rate constraints. It introduces sequence-aware policies and GFPO variants, demonstrating transfer from simulation to real CMS collision data.

  • Motivation: Static, hand-tuned trigger menus require repeated expert re-optimization as pileup and background composition drift.This workflow can become miscalibrated under distribution shift, reducing efficiency for rare-phenomena detection.
  • Approach: Online threshold tuning is formulated as sequential decision-making, with streaming summaries guiding threshold updates to maximize long-horizon physics utility under background-rate constraints.The observations include recent rates and distributional statistics, allowing the policy to model time-dependent effects.
  • Evaluation: The framework evaluates feasibility through the InBand fraction and performance through signal efficiency, while a DQN controller can adapt thresholds in real time.InBand is the proportion of events whose trigger rate remains within the target bandwidth tolerance.
  • Approach: A sequence-based observation model preserves distributional context that can distinguish global luminosity shifts from local score-distribution changes near threshold.A recurrent encoder compresses the K-event window and physics-informed features into a fixed-size state vector.
  • Contributions: GFPO-F and GFPO-FR filter rollouts by feasibility and signal efficiency, improving stability without degrading efficiency on simulated events and CMS collision data.The variants transfer from simulation to CMS data with negligible performance loss and require no gradient-based deployment fine-tuning.
  • Contributions: The framework extends beyond the LHC to online anomaly-detection benchmarks, where sequence-based representations with GFPO-F and GFPO-FR improve streaming performance under distribution shifts.This extension is presented as evidence that the approach applies beyond particle physics.
  • Contributions: The work demonstrates an end-to-end RL-trained trigger policy using real CMS collision data, learning threshold updates from streaming experience rather than a global cost function.The demonstration uses CMS Runs 283408 and 283876.

2 Related Work

The related work positions LHC trigger control as rate-constrained reinforcement learning and distinguishes it from anomaly-detection thresholding and prior selective-update methods. The paper adapts rejection-style filtering to enforce rate feasibility under distribution shift.

  • Reinforcement learning for scientific decision making: LHC trigger control must maintain background acceptance within a strict tolerance while maximizing signal efficiency under pileup and detector drift.This rate-constrained setting distinguishes it from other scientific RL applications such as self-driving laboratories and accelerator tuning.
  • Anomaly detection under drift: Unlike active anomaly detection, this work adapts an autoencoder score threshold to satisfy rate constraints under drift rather than learning query policies under a label budget.It treats trigger thresholds as real-time control variables in a feedback loop.
  • Anomaly detection under drift: DSPOT is included as a baseline, but it has no mechanism for optimizing signal efficiency.
  • Selective updates: The paper adapts GFPO-style rejection filtering to streaming trigger control by filtering candidate actions according to rate feasibility under distribution shift.GFPO-F retains candidates with the smallest rate deviation, while GFPO-FR first selects feasible candidates.

3 RL formulation in Adaptive Thresholding

The paper formulates adaptive trigger thresholding as an MDP with streaming observations, discrete updates, and rewards balancing rate feasibility, signal efficiency, and control stability. It compares DQN, GRPO, and feasibility-filtered policy optimization under non-stationary trigger conditions.

  • 3 RL formulation in Adaptive Thresholding: The controller updates a scalar threshold independently for HT and AD triggers using streaming event summaries and observed rates.The state includes recent-event features, the current threshold, observed rate, and the previous threshold change.
  • 3 RL formulation in Adaptive Thresholding: The MDP uses discrete threshold increments, with each action applied to the next event batch while the data-stream distribution may drift with beam conditions.The transition dynamics are not assumed known to the agent.
  • 3 RL formulation in Adaptive Thresholding: The reward combines background-rate control, signal efficiency for t¯t and h →4b, and a penalty for threshold movement.The target operating band is defined by a tolerance around the target rate, and the controller is bounded by a maximum per-step update.
  • 3.1 Warmup: DQN for single trigger: The safety shield limits threshold changes to configurations compatible with detector bandwidth, computing, and safety constraints.The paper gives an approximately 100 kHz ATLAS Level-1 rate limit as an operational example.
  • 3.1 Warmup: DQN for single trigger: DQN improves held-out MC InBand from 0.52 to 1.00 for HT and from 0.73 to 0.99 for AD over PID, with comparable or higher signal efficiency.Removing online updates sharply degrades the frozen DQN-F ablation, especially for HT.
  • 3.1 Warmup: DQN for single trigger: Non-stationary and heteroskedastic rate sensitivity makes small threshold changes capable of producing large rate excursions, motivating on-policy methods without replay buffers.The paper connects this regime to stale-target problems in critic-based learning and to memory constraints in hardware implementations.
  • 3.2 Critic Free Policy Optimization: GRPO removes the learned critic through within-group reward normalization, while GFPO variants filter candidate rollouts by feasibility before policy updating.GFPO-F ranks candidates by rate deviation, whereas GFPO-FR first selects feasible candidates and then ranks them by signal utility.

4 Group-Filtered Policy Optimization

GFPO adapts group-filtered policy optimization to streaming trigger control by enforcing feasibility before advantage normalization. GFPO-F prioritizes rate proximity, while GFPO-FR prioritizes feasible candidates and then signal efficiency, with both addressing zero-feasibility failures in GRPO.

  • GFPO filters candidate actions by rate feasibility before advantage normalization, adapting the method to streaming trigger control under distribution shift.The approach must choose which candidates to keep and how to handle steps with too few feasible candidates.
  • GFPO-F keeps the K candidates with the smallest absolute rate error, including infeasible candidates when necessary.This biases learning toward the tolerance-band boundary and stabilizes training when feasible candidates are scarce.
  • GFPO-FR selects feasible candidates first and ranks them by signal efficiency, padding with closest-to-feasible out-of-band candidates when fewer than K are available.Its design seeks signal efficiency within the safe region, whereas GFPO-F trades signal for tighter rate control.
  • HT diagnostics: GFPO-F concentrates HT background rates at µ = 100.0 kHz with σ = 2.0 kHz, while GFPO-FR has µ = 106.1 kHz and σ = 3.4 kHz inside the [90, 110] kHz band.The baselines spread well outside the tolerance band.
  • Failure mode: 30.8% of HT steps contain no rate-feasible candidate in the sampled GRPO group, causing group-relative normalization to reinforce an out-of-band action.The average feasible fraction is 0.58, so zero-feasibility groups provide no in-band learning signal.
  • Trade-off: GFPO-FR reduces zero-feasibility groups to 8.42% versus 9.26% for GFPO-F, trading tighter rate control for signal efficiency.GFPO-FR reaches 97.9% InBand on AD versus 100% for GFPO-F, while its HT rate-deviation MAE is 0.016 versus 0.004.

5 Beyond HEP: Anomaly Detection Benchmarks

The paper tests whether rate-constrained threshold control under distribution drift generalizes beyond particle physics. On UNSW-NB15, GFPO transfers both the failure mode and its resolution, while on unconstrained NAB its components remain competitive.

  • Rate-constrained threshold control under drifting score distributions is presented as a general problem beyond LHC triggering.The setting involves scalar-score thresholding, background drift, and a hard rate budget.
  • Experimental setup: UNSW-NB15 preserves the rate-constrained protocol, while NAB removes rate regulation to isolate sequence-based state and adaptive thresholding.The common protocol updates thresholds once per temporal chunk and evaluates InBand fraction, rate error, and signal detection metrics where applicable.
  • UNSW-NB15: GFPO-F and GFPO-FR reach nearly 4% and 4.9% in-band frozen FAR on UNSW-NB15, compared with 2.5% for PID.GRPO, L-GRPO, DQN, and DSPOT collapse to 0% in-band FAR, while GFPO methods dominate signal TPR on Backdoors.
  • NAB: GFPO-F and GFPO-FR attain the highest NAB F1 scores, 0.215 and 0.216, approximately 3% above Constant-opt at 0.184.DQN, GRPO, and L-GRPO underperform Constant-opt despite using the same reward.

6 Conclusion

The paper frames feasibility-first filtering as the key to stable, efficient trigger control under non-stationarity, while identifying important scope limits for single-threshold operation and CMS Open Data evaluation.

  • 6 Conclusion: GFPO feasibility-first filtering improves rate stability and signal efficiency across Monte Carlo and CMS collision data.The same failure mode and resolution also appear on UNSW-NB15 and NAB.
  • 6 Conclusion: Single-trigger control couples signal efficiencies, with t¯t exceeding 95% in-band efficiency while h →4b reaches only 29-36% under the same cut.The reward mixing parameter α provides only partial control over this coupling.
  • 6 Conclusion: The HEP evaluation is restricted to CMS Open Data, using Run 283408 for deployment and Run 283876 for autoencoder training.The passage identifies generalization to other LHC experiments such as ATLAS as an unresolved scope boundary.
  • 6 Conclusion: At larger menu scale, the proposed extension is feasibility filtering over a joint threshold vector under a shared rate budget.The single-trigger case is presented as a conservative estimate of the gap at trigger-menu scale.

Appendices

The appendices position the method against scientific RL, sequence modeling, anomaly detection, extreme-value thresholds, and selective-update methods, then specify feasibility-aware GFPO filtering for streaming trigger control.

  • RL for scientific decision making: Unlike scientific RL settings targeting a desired physical state, this work maintains a fixed background acceptance rate within tolerance.The paper gives 0.25% as an example target background rate.
  • Anomaly detection under drift and threshold control: The Anomaly Transformer is excluded because its association-discrepancy assumption requires regularly sampled, autocorrelated time series, unlike independent LHC bunch crossings.The paper instead adapts an autoencoder detector’s operating threshold online.
  • Anomaly detection under drift and threshold control: DSPOT responds statistically to score drift but tracks score quantiles rather than closing the loop on measured background acceptance and signal efficiency.It is included by using its adaptive threshold z_q as the per-chunk trigger cut.
  • Data filtering and selective updates in RL: GFPO-F retains candidate threshold updates with the smallest rate deviation, whereas GFPO-FR first selects candidates inside an expanded tolerance band and then ranks them by signal utility.Both filters turn candidate selection into a soft realization of the rate constraint.
  • Choice of policy optimizer: PPO’s learned-critic and multi-epoch assumptions are mismatched to drifting background distributions and rate-threshold sensitivities, motivating GRPO-style updates.The deployment setting also makes trajectory clustering incompatible with the per-chunk update budget.

B Theoretical Justification

The theoretical analysis models each stream chunk as a changing MDP and contrasts critic-dependent PPO bias with within-chunk, affine-equivariant GRPO estimation under non-stationarity.

  • Streaming non-stationary setting: Each chunk induces a distinct MDP because rewards and transition kernels shift as luminosity, pileup, and detector conditions evolve.The analysis summarizes the gap between successive chunks with a scalar drift measure.
  • Streaming non-stationary setting: Observed monotone drift in background HT and AD-score distributions across MC and CMS runs satisfies the drift pattern used in the theoretical analysis.The figures show running distributions shifting over run time.
  • Critic drift biases the PPO advantage: PPO’s advantage bias scales with critic mismatch, which is O(δ_k/(1−γ)) when the critic was fit on past chunks and drift is δ_k.Bootstrapping and multi-epoch trajectory reuse propagate and compound this bias across chunks.
  • Group baseline as a within-chunk control variate: GRPO’s empirical per-state estimator has bias O(G−1) and conditional mean-squared error O(G−1), depending only on within-chunk statistics.No information from earlier chunks enters the gradient estimator.
  • Per-chunk affine equivariance: Per-chunk affine rescaling of rewards preserves the GRPO learning signal, while PPO has no analogous equivariance because its critic is not rescaled for the current chunk.With ε = 0, the GRPO difference vanishes identically; ε > 0 introduces a bounded residual.
  • Discussion and limitations: The analysis concludes that PPO bias is O(δ_k), whereas GRPO bias against its preconditioned target is O(G−1) and drift-independent.The two bounds measure different targets, so the paper draws no direct comparative conclusion from them.

D.1 Dataset Summary

The dataset summary covers independent MC and CMS collision streams for HT and AD triggers, defines oracle and in-band evaluation conventions, and reports signal-overlap and hyperparameter-sweep context.

  • Datasets: The study uses t¯t and h →4b signals shared across MC and real CMS collision data, with evaluation on CMS Run 283408 for deployment.For the AD trigger, the autoencoder is trained on Run 283876 and evaluated on 1.99M background events from Run 283408.
  • Score drift: Background scores drift downward over each run; MC moves more strongly toward zero, while CMS retains a heavy upper tail and is globally higher-scoring than MC.Figures 6 and 7 summarize running means, medians, and central 5-95% bands.
  • Signal/background overlap: t¯t is predominantly above the oracle cut, whereas h →4b overlaps heavily with background and therefore has a lower threshold-based efficiency ceiling.For h →4b, each b-quark carries approximately mH/4 ≈31 GeV, placing many scores in the background bulk.
  • Oracle cut: The oracle cut r+ is the tightest threshold inside the upper tolerance edge, and its signal efficiency is the maximum achievable by a deterministic threshold rule.Results exclude the first 10 calibration chunks and report events after the specified MC and CMS cutoffs.
  • Oracle signal efficiencies: On CMS Run 283408 h →4b, GFPO-F and GFPO-FR reach 42.533 and 44.477 in-band efficiency, respectively, versus 41.028 for the oracle.The in-band chunks concentrate in the high-efficiency tail of the run.
  • Hyperparameters in reward design: Across reward-weight configurations, GFPO-F and GFPO-FR form a tight upper-right cluster, while baselines span a pronounced in-band-rate versus signal-efficiency trade-off.The sweep covers 25 configurations per trigger for λ1 and λ2.

D.4 Standard anomaly detection metrics for LHC

Standard anomaly-detection metrics capture signal-selection performance but omit the background-rate constraints central to LHC trigger deployment. The reported results therefore emphasize operational feasibility alongside signal efficiency.

  • Temporal drift: The h →4b signal efficiency drifts from approximately 11% in early high-pileup chunks to approximately 59% in late low-pileup chunks.Inband chunks concentrate in the run’s high-efficiency tail, affecting aggregate efficiency comparisons.
  • MC results: GFPO-FR ranks in the top two for 9 of 14 HT metrics and 7 of 14 AD metrics on the MC benchmark.It also achieves the highest InBand rate among methods.
  • CMS results: GFPO-FR ranks in the top two for 7 of 14 HT metrics and 6 of 14 AD metrics on CMS Run 283408 with test-time training.These conventional metrics focus on signal efficiency rather than background-rate constraints.
  • Zero-shot transfer: In zero-shot CMS transfer, baseline methods can obtain more top-two signal-efficiency metrics than GFPO-F, but those metrics do not enforce background-rate constraints.MC-trained policies are frozen and deployed without fine-tuning in this setting.

E.1.3 L-GRPO

L-GRPO replaces a fixed rate-penalty weight with an online dual variable, but this adaptation does not resolve GRPO’s zero-feasibility failure under distribution shift. The failure arises because reward weighting cannot change the candidate-action proposal distribution.

  • Method: L-GRPO updates λt by projected gradient ascent on a Lagrangian after observing the realized background rate at each chunk boundary.λt increases when the rate violates the tolerance band and decreases when the constraint is satisfied.
  • Zero-feasibility: Approximately 20% of MC steps and approximately 61% of CMS steps are zero-feasible for both GRPO and L-GRPO.Their per-chunk trajectories are nearly indistinguishable.
  • Failure mechanism: Because λt changes reward weights but not the proposal distribution, it cannot reduce the fraction of steps with zero feasible candidates.This leaves the underlying candidate-sampling failure unchanged.
  • CMS behavior: On CMS data, λt remains near 0.25 while the in-band rate oscillates around 54% without systematic recovery.When no feasible action is sampled, normalization favors the least-infeasible out-of-band candidate.
  • Comparison: L-GRPO remains systematically below the tolerance band on CMS, whereas GRPO and both GFPO variants remain near the target rate.This comparison motivates feasibility-first candidate selection in GFPO.
  • Group-size analysis: Zero-feasibility persists at 24–34% of steps even with group size G = 256, so increasing G alone cannot recover constraint satisfaction.The failure is attributed to the policy distribution under the constraint rather than estimator sample size.

E.3 Sequential Network Architectures

The sequential architecture represents recent event and control context so policies can respond to non-stationary threshold-to-rate mappings. Across architectures, signal efficiency is nearly unchanged, while InBand rate distinguishes the recurrent choices.

  • Motivation: The trigger-control problem is partially observed and non-stationary because pileup, detector conditions, and score distributions change the threshold-to-rate mapping.A scalar controller state cannot expose the recent event-level distribution near the cut.
  • State representation: The state is a fixed-length sequence of the most recent K background events with event-level, threshold-geometry, and chunk-level control features.The feature groups include scores and kinematics, distance-to-cut and pass indicators, rate error, drift, last action, and feasibility.
  • Architectures: All five RL agents use the same single-layer RNN and linear projection head when trained on MC and deployed on CMS Run 283408.The compared recurrent cells include LSTM, GRU, vanilla tanh RNN, and vanilla ReLU RNN.
  • Results: On HT, signal efficiencies vary by under 0.1% across recurrent cells, while InBand rates differ substantially, including 0.849 for GRU and 0.789 for vanilla RNN.The reported t¯t and h →4b overall efficiencies are nearly architecture-invariant.
  • Choice: The GRU is adopted because trigger rate scales with instantaneous luminosity, making rate stability operationally critical.The architecture choice therefore prioritizes InBand rate over small signal-efficiency differences.
  • Controller behavior: DQN tracks the tolerance band more consistently than constant and PID controllers, while DQN-F further suppresses out-of-band swings and recovers faster after pileup transitions.The same qualitative pattern appears for both HT and AD triggers.

F.2 GRPO MC training

On Monte Carlo streams, RL controllers maintain background rates near the 100 kHz target and achieve higher signal efficiency than static and PID baselines. GFPO variants also transfer effectively to CMS data without deployment-time fine-tuning.

  • MC background-rate control: RL controllers concentrate background rates near the 100 kHz target, while Constant decays and PID frequently leaves the tolerance band.GRPO, GFPO-F, and GFPO-FR show this behavior on both HT and AD triggers.
  • MC background-rate control: GFPO-F and GFPO-FR achieve the highest in-band background-rate occupancy among established baselines.GFPO-FR operates nearer the band edge for signal than GFPO, at 106.1 kHz versus 100.0 kHz.
  • MC-to-CMS transfer: GFPO-F and GFPO-FR provide the strongest overall stability–efficiency tradeoff on real CMS collision data.Across methods and triggers, their InBand rate and signal-efficiency results remain consistently near the top.
  • MC signal efficiency: RL methods provide higher signal efficiency over time for both t¯t and h →4b than PID and constant menus.GFPO-FR reaches roughly 48% HT and 40% AD efficiency for h →4b in late chunks, versus roughly 24% for Constant.
  • MC-to-CMS transfer: Frozen policies trained on MC achieve nearly the same CMS background-rate behavior as policies updated online during deployment.The comparison uses CMS Run 283408, with no online updates for the frozen policies.

F.5 CPO degeneracy on the CMS dataset

The CPO procedure makes executed actions effectively independent of the learned policy because candidate selection is screened by feasibility and cost. Consequently, multiple trained checkpoints can produce identical CMS trajectories, while test-time training does not improve recovery from violations.

  • CPO degeneracy: CPO produces zero per-seed standard deviation because its executed trajectory is screened from the policy.The CPO entries agree to floating-point precision even though trained checkpoints have different policy parameters.
  • CPO action screening: CPO executes the candidate with minimum cost when no sampled action satisfies the strict feasibility threshold.Its action set contains five discrete threshold changes, and candidates are evaluated using reward and cost.
  • CPO action screening: The executed CPO action is determined by data-dependent costs rather than policy preferences in the operative branch.The policy only determines which actions enter the sampled group; the selected action otherwise coincides with the full-action-set optimizer with high probability.
  • Test-time training: Test-time training does not improve CMS recovery because the recovery branch moves parameters along the cost gradient when violations exceed the budget.The recovery step dominates on roughly half of CMS update chunks for the AD trigger.
  • Test-time training: CMS test-time-training trajectories are shown as per-step background rates for HT and AD triggers.The policies are trained on MC and updated online during the real-data run.

G Ablation study on noisy anomaly scores

Changing the autoencoder latent dimension alters anomaly-score distributions but leaves RL controller behavior broadly stable. Feasibility-first filtering prioritizes rate compliance, while relaxed filtering can improve conditional efficiency at the cost of more violations.

  • Latent-dimension ablation: Changing latent dimension alters the anomaly-score distribution, but qualitative controller behavior remains stable.The ablation retrains autoencoders with d ∈ {4, 6, 8, 10, 12, 14, 16} under distribution shift.
  • Latent-dimension ablation: RL controllers achieve InBand ≈0.76–1.00 across latent dimensions, with low MAE and P95|e|.PID achieves substantially lower InBand fractions of about 0.30–0.46 for d ∈ {4, 6, 8}.
  • Feasibility–efficiency trade-off: GFPO-F achieves near-perfect background constraint satisfaction, whereas GFPO-FR can improve conditional signal efficiency with more violations.This exposes the expected feasibility–efficiency trade-off among filtered objectives.
  • Interpretation: The controllers primarily use rate feedback over windows rather than relying on precise absolute anomaly-score calibration.This supports robustness to moderate changes in autoencoder capacity and score quality.

H.1 UNSW-NB15

On UNSW-NB15, GFPO-F and GFPO-FR preserve the FAR budget while improving detection of both easy and hard attack classes, especially rare Backdoors. Interpretation is bounded by chunk-level FAR quantization and a structural ceiling on achievable in-band rate.

  • Benchmark setup: The UNSW-NB15 FAR target is 0.5% with tolerance 0.05%, matching the LHC setting’s relative tolerance rather than its absolute rate.The benchmark uses a 21-bin threshold-delta action grid and a length-8 history of distributional and detection features.
  • Benchmark limitations: Chunk-level FAR quantization makes the LHC tolerance band unreachable and limits the chosen UNSW band to one realizable false-positive count.With 447 negatives per chunk, FAR values change in steps of approximately 0.224%.
  • Benchmark limitations: The achievable in-band fraction is capped at approximately 47.6% because 43 of 82 Period 2 chunks contain only attack traffic.Those chunks have FAR ≡ 0 regardless of threshold and cannot fall within the target band.
  • Per-class detection: GFPO-F and GFPO-FR achieve approximately 22% Exploits TPR and approximately 12% Backdoors TPR while satisfying the FAR budget.Their Backdoors TPR is roughly twice that of the next-best learned baseline, and test-time training widens the gap.
  • Per-class detection: GFPO variants outperform baselines on hard Backdoors while matching the best Exploits TPR.Constant and PID reach about 22% Exploits TPR but collapse to 6–8% on Backdoors and remain outside the FAR band.
  • FAR trade-off: PPO recovers some TPR only by abandoning the FAR budget, with mean absolute FAR error two orders of magnitude worse than the proposed methods.This contrasts with GFPO-F and GFPO-FR, which satisfy the FAR constraint.
Loading 2606.23993v3…