Scoring ISAC: Benchmarking Integrated Sensing

• Scoring ISAC is a framework that uses learned score functions to estimate communication and sensing metrics in complex ISAC environments.
• It replaces fragile analytical assumptions with DSM-trained models to accurately estimate metrics like mutual information, MMSE, FIM, BCRB, and KLD.
• The framework unifies ISAC performance benchmarking by addressing nonlinear, heavy-tailed, and hardware-impaired regimes through score-based Monte Carlo estimators.

Searching arXiv for the specified paper and closely related ISAC benchmarking work. Scoring ISAC is a tutorial-style framework for benchmarking integrated sensing and communications (ISAC) systems through score-based generative modeling, introduced in “Scoring ISAC: Benchmarking Integrated Sensing and Communications via Score-Based Generative Modeling” (Chen et al., 4 Aug 2025). In this framework, score functions—gradients of log densities—are learned from data and then used to estimate classical performance metrics such as mutual information (MI), minimum mean squared error (MMSE), Fisher information (FIM), Bayesian Cramér–Rao bound (BCRB), and Kullback–Leibler divergence (KLD) under realistic ISAC conditions. The motivation is that practical ISAC environments involve hardware impairments, multipath propagation, interference, clutter, and scene priors that induce nonlinear, heavy-tailed, multimodal, and non-Gaussian statistics, making analytical evaluation of these metrics intractable or severely mismatched under simplified models (Chen et al., 4 Aug 2025).

1. Problem Setting and Rationale

Integrated sensing and communications systems share spectrum, signals, and often hardware to simultaneously deliver high-throughput communication and high-accuracy sensing. A canonical model is

$y_c = H_c x + z_c$

for communication and

$y_s = H_s x + z_s$

for sensing, where $x$ is the transmit signal, $H_c$ and $H_s$ are the communication and sensing channels, and $z_c, z_s$ are additive noises (Chen et al., 4 Aug 2025).

The central benchmarking problem addressed by Scoring ISAC arises because practical operating conditions depart sharply from linear-Gaussian assumptions. The data summary identifies hardware impairments such as PA nonlinearity, IQ imbalance, phase noise, and ADC quantization, as well as multipath, clutter, interference, and scene priors constrained to low-dimensional manifolds. Under such conditions, analytical expressions for MI, MMSE, FIM, and BCRB become difficult to derive or unreliable as performance surrogates (Chen et al., 4 Aug 2025).

These metrics are not peripheral quantities. MI is tied both to communication capacity and to expected information gain about the environment; MMSE measures estimation accuracy for quantities such as channels or target parameters; FIM and BCRB quantify fundamental estimation limits and inform waveform, beamforming, precoding, and resource allocation design (Chen et al., 4 Aug 2025). This positions Scoring ISAC as a benchmarking methodology rather than merely a generative-modeling application.

A plausible implication is that the framework is intended to replace assumption-driven performance analysis with data-driven performance analysis when the underlying distributions are unavailable in closed form. This interpretation follows from the paper’s explicit statement that the key idea is to “replace fragile analytical assumptions with data-driven score-based generative modeling” (Chen et al., 4 Aug 2025).

2. Score-Based Formulation and Metric Identities

The core object is the score function

$s(x) = \nabla_x \log p(x),$

defined for a random vector $x$ with density $p(x)$ (Chen et al., 4 Aug 2025). Scoring ISAC uses learned approximations of such scores in observation space, parameter space, and posterior space.

The learning mechanism is based on Noise Conditional Score Networks and denoising score matching. Gaussian noise is injected according to

$x_t = x + \sigma n,\quad n \sim \mathcal{N}(0,I),$

and the practical training objective is the denoising score matching surrogate

$y_s = H_s x + z_s$0

The summary states that $y_s = H_s x + z_s$1 is often chosen as $y_s = H_s x + z_s$2, and that learning $y_s = H_s x + z_s$3 at multiple $y_s = H_s x + z_s$4 yields access to score gradients across noise scales (Chen et al., 4 Aug 2025).

The framework’s importance rests on exact identities linking classical metrics to score functions. The de Bruijn identity gives

$y_s = H_s x + z_s$5

for $y_s = H_s x + z_s$6 with $y_s = H_s x + z_s$7, thereby connecting learned observation-space scores to Fisher information and, through entropy, to MI (Chen et al., 4 Aug 2025). For additive Gaussian channels, the I–MMSE relationship is

$y_s = H_s x + z_s$8

which allows MI to be recovered from MMSE once MMSE itself is estimated from learned scores (Chen et al., 4 Aug 2025).

For estimation-theoretic limits, the parameter-space Fisher information is

$y_s = H_s x + z_s$9

and the Bayesian FIM satisfies

$x$0

with

$x$1

The corresponding BCRB is

$x$2

The posterior form

$x$3

shows that learning the conditional or posterior score alone suffices for BCRB estimation (Chen et al., 4 Aug 2025).

The same summary also records the small-perturbation KLD curvature identity

$x$4

which links local KLD geometry to Fisher information (Chen et al., 4 Aug 2025).

3. Practical Estimators and Training Procedure

Scoring ISAC turns the preceding identities into Monte Carlo estimators driven by learned score networks. In observation space, the trace of Fisher information is approximated from samples $x$5 by

$x$6

which then supports entropy and MI estimation via the de Bruijn identity (Chen et al., 4 Aug 2025).

For Gaussian denoising models, Tweedie’s formula gives the MMSE estimator

$x$7

so with a learned score $x$8 one forms

$x$9

and estimates

$H_c$0

by Monte Carlo before integrating over SNR to obtain MI (Chen et al., 4 Aug 2025).

For BCRB estimation, the framework supports two implementations. In the posterior approach, one learns $H_c$1 and computes

$H_c$2

In the measurement-prior approach, one separately learns a prior score $H_c$3 and a measurement score $H_c$4, obtaining

$H_c$5

and then

$H_c$6

When the likelihood is unknown, the measurement score may be trained via Fisher Score Matching using

$H_c$7

(Chen et al., 4 Aug 2025).

For general MI and KLD estimation beyond standard AWGN settings, the paper introduces a DSM-based path-integral estimator over a virtual AWGN channel $H_c$8. The identity reported is

$H_c$9

with a corresponding conditional form, leading to

$H_s$0

Replacing the exact gradients with DSM-trained scores yields a practical estimator for MI, and the same integrand supports KLD estimation as well (Chen et al., 4 Aug 2025).

The paper’s benchmarking protocol is explicitly staged: dataset creation, score training, metric estimation, validation against analytical or high-accuracy Monte Carlo references where available, and reporting with point estimates, confidence intervals from bootstrap, calibration checks based on denoising residuals, and score-norm diagnostics (Chen et al., 4 Aug 2025).

4. ISAC-Specific Training Considerations

The framework is tailored to both communication and sensing data generation. For communication, the summary specifies generating $H_s$1 under realistic waveforms including OFDM and DFT-s-OFDM, together with hardware impairments, interference, and multipath channels. For sensing, it specifies generating $H_s$2 under near-/far-field geometry, array response, clutter, target RCS models such as Swerling I, motion constraints, and interference (Chen et al., 4 Aug 2025). Conditioning variables can include $H_s$3, SNR, bandwidth, array geometry, waveform index, and known parts of the measurement model.

Complex-valued data are handled by stacking real and imaginary parts, a step the summary calls “realification” (Chen et al., 4 Aug 2025). Noise schedules are geometrically discretized between $H_s$4 and $H_s$5, and classifier-free conditioning is used by dropping $H_s$6 with probability $H_s$7 so that one network can output unconditional and conditional or posterior scores (Chen et al., 4 Aug 2025).

Architecturally, the paper states that U-Nets or MLPs suffice in low- to moderate-dimensional ISAC examples, while 1D or 2D CNNs can exploit array structure. Calibration is emphasized: denoising consistency requires $H_s$8 to predict $H_s$9 on average, and the summary specifically lists EMA weights, spectral normalization, and consistency regularizers as stabilizing techniques (Chen et al., 4 Aug 2025). Heavy-tailed and multimodal statistics are treated as a central use case rather than an edge case; VE SDE training with large $z_c, z_s$0 is identified as helping global coverage in such settings (Chen et al., 4 Aug 2025).

This training guidance aligns with broader ISAC benchmarking efforts that emphasize the mismatch between deterministic or Gaussianized objectives and realistic signal statistics. For example, “Random ISAC Signals Deserve Dedicated Precoding” (Lu et al., 2023) formalizes ergodic linear minimum mean square error under finite-frame random signaling, while “Modeling and Analysis of Near-Field ISAC” (Zhao et al., 2023) shows that accurate electromagnetic modeling can materially alter attainable sensing and communication rates. Scoring ISAC extends this general trend by shifting the benchmarking object from explicitly modeled moments or rates to learned score fields (Chen et al., 4 Aug 2025).

5. Proof-of-Concept Experiments

The proof-of-concept validation in Scoring ISAC focuses on target detection and target localization (Chen et al., 4 Aug 2025). In the detection experiment, the binary hypotheses are

$z_c, z_s$1

with $z_c, z_s$2 (Chen et al., 4 Aug 2025). The Neyman–Pearson likelihood-ratio test is

$z_c, z_s$3

When both hypotheses are Gaussian, the analytical relation

$z_c, z_s$4

holds, together with the asymptotic law

$z_c, z_s$5

With learned scores, the likelihood ratio is approximated by a line integral of score differences,

$z_c, z_s$6

computed in practice along a diffusion path or straight-line path with trapezoid-rule discretization (Chen et al., 4 Aug 2025).

Two cases are reported. When the reflection is known, $z_c, z_s$7, analytical KLD and analytical $z_c, z_s$8 match the score-based KLD and the detection performance curve, and $z_c, z_s$9 grows with KLD as predicted by theory. When the reflection is unknown, $s(x) = \nabla_x \log p(x),$0, no closed form is available; a DSM-trained class-conditional score network estimates KLD numerically, and empirical plots show a clear positive correlation between KLD and $s(x) = \nabla_x \log p(x),$1 at fixed $s(x) = \nabla_x \log p(x),$2 (Chen et al., 4 Aug 2025).

The localization experiments use parameter vectors $s(x) = \nabla_x \log p(x),$3 or $s(x) = \nabla_x \log p(x),$4 with measurement model

$s(x) = \nabla_x \log p(x),$5

For $s(x) = \nabla_x \log p(x),$6 and Gaussian prior $s(x) = \nabla_x \log p(x),$7, the analytical expressions are

$s(x) = \nabla_x \log p(x),$8

with

$s(x) = \nabla_x \log p(x),$9

and

$x$0

The learned BCRB tracks the analytical BCRB closely for both $x$1 and $x$2; the summary adds that CRB without a prior is looser at low SNR and that average bounds slightly improve in the exponential-reflection case because $x$3 increases average SNR (Chen et al., 4 Aug 2025).

These experiments are deliberately proof-of-concept rather than large-scale benchmarks. Their role is to validate that score-based estimators reproduce known analytical results in tractable settings and extend to cases where closed forms are absent (Chen et al., 4 Aug 2025).

6. Benchmarking Role, Advantages, and Limitations

The paper identifies three principal advantages. First, score-based estimators handle nonlinear, multimodal, heavy-tailed, and hardware-impaired regimes where closed forms are unavailable. Second, the same learned score functions can support MI, MMSE, FIM, BCRB, and KLD estimation. Third, the framework unifies communication and sensing by using the same engine for observation-space metrics such as MMSE and MI and parameter-space metrics such as FIM and BCRB (Chen et al., 4 Aug 2025).

This benchmarking orientation situates Scoring ISAC alongside other ISAC performance frameworks but with a different methodological emphasis. “MIMO-ISAC: Performance Analysis and Rate Region Characterization” (Ouyang et al., 2022) analyzes sensing rate, communication rate, high-SNR slope, and rate-region inclusion under tractable Gaussian models. “Downlink and Uplink NOMA-ISAC with Signal Alignment” (Zhao et al., 2023) similarly studies sensing rate, communication rate, outage probability, high-SNR slope, and diversity order through structured analytical models. By contrast, Scoring ISAC is designed for cases where such explicit analysis becomes unreliable or impossible (Chen et al., 4 Aug 2025).

The paper also makes a specific comparative claim against variational MI or MMSE estimators: score-based estimators exploit exact identities such as I–MMSE, de Bruijn, Tweedie, and posterior Fisher identities, avoid estimating partition functions, and are trained via DSM or ISM, which the summary describes as stable and scalable (Chen et al., 4 Aug 2025).

The limitations are equally explicit. Accurate score learning may have high sample complexity, especially in high dimensions or under severe out-of-distribution shifts. Score magnitude miscalibration induces biased metric estimates. Unknown-likelihood settings can make Fisher Score Matching sensitive to inaccuracies in the prior score. Numerical integration over $x$4 requires careful quadrature and variance reduction (Chen et al., 4 Aug 2025).

The summary also gives practical guidance rather than guarantees: DSM with 10–20 $x$5 levels often suffices for metric estimation; classifier-free guidance can share weights between unconditional and conditional branches; and for linear inverse problems, score-based turbo message passing can accelerate MMSE or posterior mean estimation by alternating LMMSE updates and score-based denoising (Chen et al., 4 Aug 2025). Since these points are framed as practical tips in the source summary, they should be understood as reported guidance rather than universal prescriptions.

7. Best Practices and Open Directions

The stated best practices include EMA of weights, gradient clipping, and spectral norm regularization; validation of score quality via denoising residuals across $x$6; uncertainty quantification through bootstrap over datasets, ensembles of score networks, and jackknife over $x$7 grids; and conditioning on SNR, bandwidth, array geometry, waveform index, and known channel components for ISAC-specific models (Chen et al., 4 Aug 2025). Reproducibility recommendations include releasing datasets with controlled impairments, scripts for DSM and FSM training, and fixed random seeds and $x$8 schedules (Chen et al., 4 Aug 2025).

The open problems listed in the summary delineate the present boundary of the framework. These include learned measurement models for unknown high-dimensional sensing likelihoods such as massive MIMO and near-field holographic arrays; Pareto optimization of communication MI versus sensing BCRB in multi-user and multi-target scenes; dynamic or sequential score models for time-varying channels and targets; native complex-valued diffusion models to avoid realification; and finite-sample error bounds, optimal $x$9 schedules, and tightness comparisons with variational estimators (Chen et al., 4 Aug 2025).

A plausible implication is that Scoring ISAC is best viewed not as a replacement for classical analysis in all settings, but as a complementary benchmarking layer for settings in which analytical benchmarks cease to be trustworthy. That interpretation is consistent with the paper’s claim that the framework “replicates and extends traditional analyses in more complex, realistic settings” (Chen et al., 4 Aug 2025).