Precision physics measurements increasingly hinge on differences rather than absolutes. A typical question is "how much does the measured energy spectrum move if the beam spot shifts by a millimetre, or if the energy scale is off by one percent?" Today such questions are answered by running a detector simulation twice, once nominal and once varied, with independent random events, and subtracting the two. Each run carries its own counting noise, so the difference is buried under statistical fluctuations that shrink only as one over the square root of the sample size. This proposal treats a generative simulator as a source of paired events instead. The generator is a diffusion or flow-matching model, meaning a network that turns random noise into realistic detector showers step by step. We generate each nominal shower and its varied counterpart from the same starting noise, so the two match everywhere except where the physical change acts. The subtraction then cancels most of the noise.
The mechanism comes from image editing. There, the attention maps of a generator (its internal record of which spatial locations influence which) are copied from a source image into an edited one, so the layout stays fixed while the content changes. Physics surrogates have so far used attention only to feed in conditioning variables. Our claim is narrower and testable. Coupled generation can cut the variance of systematic-shift estimates by a large factor at fixed simulation cost, in exactly the tail regions where counting statistics limit sensitivity. A two-level estimator removes any bias that attention copying introduces. The open question we settle empirically is whether attention injection adds anything beyond the simpler step of sharing the starting noise.
Core Claim
Let $f$ be an analysis observable, such as total deposited energy, a layer energy fraction, or the indicator that a tower exceeds a threshold with probability around $10^{-4}$. Let $c_s$ and $c_t$ be the nominal and varied conditions. The shift is $\Delta = \mathbb{E}[f \mid c_t] - \mathbb{E}[f \mid c_s]$. From $N$ pairs, its estimator has variance
$$ \mathrm{Var}(\hat\Delta) = \frac{\sigma_s^2 + \sigma_t^2 - 2\rho\,\sigma_s\sigma_t}{N}, $$
where $\sigma_s$ and $\sigma_t$ are the per-event standard deviations of $f$ under each condition and $\rho$ is the correlation between paired events. Independent simulation has $\rho = 0$. If a coupling drives $\rho$ toward 1, the result is equivalent to multiplying the sample size by roughly $1/(1-\rho)$. That is the only lever that escapes the $1/\sqrt{N}$ wall without more events.
Two coupling mechanisms are available, and they behave differently. The first is shared-noise coupling: feed the same starting noise $x_T$ through the deterministic sampling path, an ordinary differential equation (ODE), of a conditional flow model under $c_s$ and under $c_t$. Each output is then exactly distributed as the model's $p(\cdot \mid c)$, so marginals are preserved, but correlation is whatever the learned flow happens to give. The second is attention injection: additionally copy self-attention maps from the source branch into the target branch over a fraction $\lambda$ of steps and layers. This should raise $\rho$, but it distorts the target marginal.
We make injection safe with a two-level decomposition borrowed from multilevel Monte Carlo:
$$ \Delta = \mathbb{E}\big[f(y_t^{\lambda}) - f(y_s)\big] + \mathbb{E}\big[f(y_t^{0}) - f(y_t^{\lambda})\big]. $$
Here $y_s$ is the source sample, $y_t^{0}$ is the target sample with shared noise only (unbiased under the model), and $y_t^{\lambda}$ is the target sample with injection. The first term is estimated on many cheap, highly correlated pairs. The second term corrects the injection bias, and because $y_t^{0}$ and $y_t^{\lambda}$ come from the same noise, it should also have low variance and need fewer pairs. The result estimates the same quantity as unbiased shared-noise coupling. Injection is worth using only if it lowers total variance at equal compute.
Where This Sits
Three bodies of prior work feed the proposal, each in a different role.
The image-editing literature supplies the method. Prompt-to-Prompt shows that overwriting a target image's cross-attention maps with the source's pins where each concept appears in the image. Zero-shot Image-to-Image Translation stabilizes composition with an attention-consistency loss. The diffusion editing survey separates two roles: self-attention carries geometry, while cross-attention carries semantics. Its finding that resampling noise causes structural drift is exactly the decorrelation we want to suppress. MAG-Edit and COCO-Counterfactuals run parallel source and target branches to produce minimal counterfactual pairs. None of these works asks whether the edited image is a fair sample from the edited condition, because editing has no notion of marginal correctness. That requirement is new here, and the two-level estimator exists to meet it.
The physics surrogate literature supplies the testbed and the reference models. CaloChallenge 2022 compared 31 generators on calorimeter showers and found diffusion and flow-matching models most faithful on the large datasets. CaloDREAM uses a vision-transformer shape network trained with flow matching, sampled by a fast ODE solver in 4–8 steps. This makes it the natural host for injection, because it already has self-attention over voxel patches and deterministic sampling. The ALICE zero-degree calorimeter diffusion model conditions a U-Net through cross-attention on nine particle variables, including vertex position. The sPHENIX whole-event model shows that pixel-space diffusion stays accurate where event probabilities reach $10^{-5}$, while GANs deviate by factors of two. All of these works produce independent samples per condition and evaluate only per-condition fidelity. None measures or exploits correlation across conditions.
The statistical-limits literature motivates and constrains the work. RESuM attacks Poisson noise in rare-event detector design by converting binary hits into smooth per-event scores, cutting design cost by 96.7%. Our approach is complementary: RESuM smooths each condition's estimate, while we correlate the estimates across conditions. The counting-limited analyses in the file (Telescope Array anisotropy, KATRIN neutrino interactions) show why tail-region shifts matter: discovery significance lives in the tails, where simulated statistics are thinnest.
The flavor-physics direction is the application regime, not the mechanism. Collider experiments now measure semileptonic rates at percent precision: BESIII D→Kℓν on 7.93 fb⁻¹ and BABAR |V_cb| on 460 million B pairs. Once lattice form-factor uncertainties shrink (Besson argues they currently dominate the error budget), the next floor is detector systematics. Those systematics are estimated by varied-simulation subtraction, whose own statistical noise is the target of this work. Lattice practice also offers the guiding principle: quantities like F_K/F_π are formed as ratios on the same gauge configurations so that common fluctuations cancel. We bring that "same ensemble" logic to detector simulation.
We deliberately do not pursue the file's Gap 3, injecting attention into lattice gauge fields across momentum transfer. There is no evidence that layout-copying respects the gauge-field probability distribution, and normalizing-flow samplers for lattice gauge fields (stochastic flows in SU(3)) already face hard exactness requirements that injection would violate.
Research Program
The first step is to measure how much correlation is achievable. We retrain or reuse a CaloDREAM-style flow-matching shape network on CaloChallenge Dataset 2 (6,480 voxels, 200,000 Geant4 electron showers, conditioned on incident energy). We also train a cross-attention conditioned diffusion U-Net for a nine-variable, ALICE-style setup. For a grid of condition changes, we measure the paired correlation $\rho$ for each observable under three settings: independent noise, shared noise, and shared noise plus self-attention injection. The condition changes are incident-energy shifts of 0.5–5% in Dataset 2, and vertex displacements and momentum-direction tilts in the nine-variable setup. Injection strength $\lambda$ is swept over which layers are injected and over which noise levels; editing work suggests early, high-noise steps fix layout. Observables cover the bulk (total energy, longitudinal centroid) and the tails (fraction of showers with a voxel above a threshold set at exceedance probabilities $10^{-3}$ and $10^{-4}$).
The second step quantifies the bias of injection and tests the correction. For each $\lambda$ we report two things. One is the target-marginal distortion, measured as classifier AUC between $y_t^{\lambda}$ and $y_t^{0}$. The other is the second-term correction size. We then allocate pairs between the two levels to minimize total variance at a fixed GPU budget, using the standard multilevel allocation rule: sample counts proportional to the level's standard deviation divided by the square root of its cost.
The third step is validation against Geant4. A fast estimator of the surrogate's shift is useless if the surrogate's shift is wrong. For a few condition changes we therefore build a high-statistics Geant4 reference $\Delta$. For energy shifts, this comes from regressing on the continuous incident-energy distribution in the full dataset. Where we can run Geant4 ourselves, it comes from dedicated large independent samples. We then check that each coupled estimator's error bars cover the reference, and we report surrogate model error separately from estimator variance.
Baselines
The naive reference is independent sampling from the same trained surrogate: separate random noise for nominal and varied conditions, with the same number of samples and the same model. This is what a physicist would do today with a drop-in fast simulator. It isolates the gain due purely to coupling.
The first-principles reference is Geant4 run independently under both conditions at equal CPU-equivalent cost. We also run Geant4 with common random-number seeds, the physicist's existing variance-reduction trick. We expect this coupling to decorrelate quickly, because shower branching is chaotic and a small geometry or energy change desynchronizes the random stream. Measuring that decay of $\rho$ is itself a result. Beating Geant4 means a smaller $\hat\Delta$ uncertainty at equal wall-clock time, with the Geant4 reference inside the error bars.
The strongest existing generator is CaloDREAM, which reports classifier AUC of about 0.53 on Dataset 2 and 17 ms per H100 forward pass (CaloDREAM). We use its architecture and solver settings for all coupled variants, so differences come only from the coupling. For the nine-variable setting, the reference is the pixel-space DDPM with mean Wasserstein distance 1.2 over five channels (ALICE ZDC). No reported result was found on Lacuna for systematic-shift variance on any of these benchmarks, so all variance numbers are new measurements.
The key ablation is shared-noise coupling without injection. This is the strongest competitor, because it is unbiased and requires no architectural change. All coupled variants get the same trained weights, the same ODE solver and step count, and the same total GPU seconds. Injection counts as beating shared-noise coupling only if the two-level estimator reduces variance by at least a factor of 1.5, with the correction included and at matched compute, on the tail observables. A second ablation injects cross-attention instead of self-attention in the U-Net model, to test whether the geometry-versus-semantics split from image editing carries over to showers.
Why not just…?
Why not just reweight the nominal sample? Event-by-event reweighting gives perfectly correlated nominal and varied estimates, and it is standard for physics-model variations such as form-factor shapes. It works when the variation changes event probabilities but not detector response. It cannot express a vertex displacement or an energy-scale shift that moves energy between voxels. Where reweighting does apply, for incident-energy shifts within Dataset 2's continuous energy range, we include it as an extra baseline and expect it to win in the bulk. The open question is whether it degrades in the tails, where few nominal events carry large weights.
Why not just share the starting noise and skip attention injection? This is the real threat, and the proposal is built so the experiments decide it. Deterministic flow samplers may already yield $\rho$ near 1 for small shifts, in which case injection adds bias-correction cost for nothing. If that is the outcome, we will report shared-noise coupling as the method and attention injection as unnecessary. That would still be a useful, unreported result for anyone using CaloChallenge-style surrogates for systematics. We expect injection to matter mainly for larger shifts and tail observables, where flow trajectories diverge.
Why not just generate more independent surrogate events, since they are cheap? At 17 ms per shower, a 100-fold variance reduction from brute force costs 100 times the GPU time per variation. Analyses evaluate dozens of variations across many bins, and tail bins with probability around $10^{-4}$ need millions of events each. Our cost-normalized comparison directly reports how many independent events each coupled estimator is worth.
Why not just smooth the response with a RESuM-style per-event score? RESuM reduces per-condition Poisson noise and is orthogonal to coupling: one can score coupled pairs. We test the combination on the tail observables. If scoring alone closes most of the gap, coupling is a second-order gain and we will say so.
Proposed Experiments and Decisive Checks
The main experiment is on CaloChallenge Dataset 2. For energy-scale shifts of 0.5%, 1%, 2% and 5%, we report the variance-reduction factor relative to independent sampling, at equal GPU cost, for each estimator: shared noise, injection uncorrected, and injection with two-level correction. Results are given for bulk and tail observables, together with coverage of the Geant4 reference. The second experiment repeats this on the nine-variable cross-attention model with vertex and direction shifts. The third tests scale: we take the best coupling to Dataset 3 (40,500 voxels) through CaloDREAM's latent model. That paper reports latent-space low-energy artifacts, and the ALICE latent diffusion model degraded to a mean Wasserstein distance of 12.6. Those results warn that compression may break pairing in the sparse low-energy voxels, which is exactly where tail observables live.
The decisive checks are three. First, the shared-noise estimator must cover the Geant4 reference; if it does not, surrogate error rather than variance is the bottleneck and the program shifts to model fidelity. Second, the correlation curve $\rho(\text{shift size})$ under shared noise determines whether injection has any room to help. Third, at matched compute, the corrected injection estimator must beat shared noise by the factor stated above on tail observables, or the injection component is dropped.
References
- Prompt-to-Prompt Image Editing with Cross-Attention Control
- Zero-shot Image-to-Image Translation
- Diffusion Model-Based Image Editing: A Survey
- MAG-Edit: Localized Image Editing in Complex Scenarios via Mask-Based Attention-Adjusted Guidance
- COCO-Counterfactuals: Automatically Constructed Counterfactual Examples for Image-Text Pairs
- CaloChallenge 2022: A Community Challenge for Fast Calorimeter Simulation
- CaloDREAM – Detector response emulation via attentive flow matching
- Generative Diffusion Models for Fast Simulations of Particle Collisions at CERN
- Effectiveness of denoising diffusion probabilistic models for fast and high-fidelity whole-event simulation in high-energy heavy-ion experiments
- RESuM: Rare Event Surrogate Model for Physics Detector Design
- INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT
- First Constraints on General Neutrino Interactions Based on KATRIN Data
- Improved measurements of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mo>ℓ</mml:mo><mml:mo>+</mml:mo></mml:msup><mml:msub><mml:mi>ν</mml:mi><mml:mo>ℓ</mml:mo></mml:msub></mml:math> and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mover accent="true"><mml:mi>K</mml:mi><mml:mo stretchy="false">¯</mml:mo></mml:mover><mml:mn>0</mml:mn></mml:msup><mml:msup><mml:mo>ℓ</mml:mo><mml:mo>+</mml:mo></mml:msup><mml:msub><mml:mi>ν</mml:mi><mml:mo>ℓ</mml:mo></mml:msub></mml:math>
- Measurement of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mo stretchy="false">|</mml:mo><mml:msub><mml:mi>V</mml:mi><mml:mi>cb</mml:mi></mml:msub><mml:mo stretchy="false">|</mml:mo></mml:math>and the Form-Factor Slope in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mover accent="true"><mml:mi>B</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mo>→</mml:mo><mml:mi>D</mml:mi><mml:msup><mml:mi>ℓ</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mi>ℓ</mml:mi></mml:msub></mml:math>Decays in Events Tagged by a Fully Reconstructed<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi></mml:math>Meson
- Improved measurements of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>D</mml:mi></mml:math>meson semileptonic decays to<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>π</mml:mi></mml:math>and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>K</mml:mi></mml:math>mesons
- Ratio<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>F</mml:mi><mml:mi>K</mml:mi></mml:msub><mml:mo>/</mml:mo><mml:msub><mml:mi>F</mml:mi><mml:mi>π</mml:mi></mml:msub></mml:math>in QCD
- Scaling of stochastic normalizing flows in SU(3) lattice gauge theory