acceptodds
Under review as a conference paper at ICLR 2027

GPARA: Graph-Posterior-Aligned Refinement and Active Acquisition for Grounding Diffusion Priors

Abstract

Active grounding of a frozen diffusion prior requires jointly determining where new measurements should be taken and how they should be used to refine the current reconstruction. Posterior-ensemble-based methods can estimate acquisition utility from generated samples, but require repeated ensemble generation as observations accumulate and capture posterior geometry only through empirical statistics. This paper proposes GPARA, which learns a context-dependent graph surrogate over diffusion prediction residuals, inducing an explicitly reusable posterior response operator that propagates measurement innovations to unobserved variables and evaluates candidate measurements through weighted posterior-risk reduction. Under the matched surrogate, we show that the same response operator also determines expected one-step acquisition benefit and yields an analytic ranking consistent with expected reconstruction improvement. A bounded learned residual calibrates the analytic utility to account for surrogate mismatch, while a small prior ensemble is generated once and reconditioned to update risk weights without repeated diffusion posterior sampling during acquisition. Experiments on two reconstruction tasks spanning wireless sensing and computer vision show consistent improvements in refinement and active acquisition over the evaluated baselines. Ablations further support the complementary roles of step-wise graph refinement, adaptive risk weighting, and analytically anchored calibration.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Diffusion models, active acquisition, posterior inference, inverse problems, active sampling, uncertainty quantification

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_FurEOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a way to improve a diffusion model’s prediction using a few real measurements, learning how errors at different locations are related so that each measurement can correct a wider area. It uses those relationships to choose where to measure next, improving radio-map reconstruction and depth estimation in its experiments.

Critical strengths
  1. GPARA introduces a potentially novel use of learned graph-based error correlations to connect diffusion refinement with active measurement selection.
  2. In experiments, the paper shows that using the same graph for both measurement selection and refinement helps both tasks than controlled baselines.
Critical weaknesses
  1. The benefit of learning the graph is not strongly established. In Table 4(a), learned edge weights improve RMSE over a uniform graph by only about 2% on both tasks, despite requiring an additional trained network. Moreover, the paper does not compare uniform and learned graphs in the full active-acquisition pipeline, so it's unclear whether graph learning provides sufficient benefit to justify its added complexity.
Questions for authors

How does the paper's algorithm work with other diffusion priors such as image generation for inverse tasks such as inpainting?

3 comments

Response (W1): Thanks for your comment. We have added a uniform-graph variant of the complete active-acquisition pipeline in revised Table 5(b). The uniform variant achieves RMSEs of 0.0146 on RFR and 0.243 on DC, compared with 0.0142 and 0.236 for learned graph, corresponding to reductions of 2.7% and 2.9%. Thus, the additional gains persist, while remain modest.

Regarding the additional complexity, we also compared computational overhead of the learned conductances in Table 8. The graph predictors are lightweight, with only 0.769M and 0.433M parameters for RFR and DC, respectively. They are trained offline once per task with the diffusion prior frozen and reused across deployment instances. Graph learning introduces negligible deployment overhead: acquisition time increases only 0.08 s on DC, while peak memory remains virtually unchanged. RFR shows no increase in either acquisition time or memory. Reconstruction time increases by only 0.01 s for both tasks.

Thus, our results distinguish the contributions of the framework: substantial gains arise from structured posterior propagation and repeated refinement, while graph conductance learning provides additional adaptation to the target task, offering further reconstruction improvements at marginal additional deployment cost.

Table 8: Online computational overhead of learned graph conductances relative to uniform graphs.

TaskGraphParamsAcquisition time (s/img)Reconstruction time (s/img)Peak memory (MiB)
RFRUniform05.03±0.845.03 \pm 0.840.136144.3
RFRLearned0.769M5.03±0.725.03 \pm 0.720.126144.3
DCUniform06.65±0.226.65 \pm 0.220.186677.6
DCLearned0.433M6.73±0.186.73 \pm 0.180.197677.9

Response (Q1): Thank you for your question. GPARA is designed as a general framework for grounding pretrained diffusion priors with sparse measurements. Many conditional diffusion models incorporate prior information through semantic embeddings (e.g., CLIP features), visual features, or other contextual representations. GPARA leverages such information to learn a structured graph over prediction residuals, capturing dependencies that guide how measurement innovations propagate across the target. The resulting graph-posterior response supports both active acquisition and observation-consistent refinement during diffusion sampling.

For example, in image inpainting, the known pixels serve as observations. We need to learn a graph network offline through posterior refinement under the diffusion backbone as shown in Appendx C.1-C.2. At each reverse step t, GPARA obtains the predicted clean image with the fixed diffusion backbone, computes the graph-posterior residual correction, and injects the corrected estimate into the reverse diffusion transition, following the refinement mechanism in Eq. (9)-(13). Here, only Stage II in Algorithm 1 is needed because the observed locations are predefined.

Thus, the framework is transferable at the methodological level, although different priors and observation models may require task-specific adaptation.

Thanks for the additional experiments and responses to my questions. I will keep my score given that the gain of a learned graph versus non-learned seems very modest and requires additional training.

Thank you for your consideration. We agree that graph learning provides incremental rather than dominant gains. We would like to clarify that learning graph conductances is an optional instantiation rather than a prerequisite for GPARA. The same graph-residual posterior-response framework accommodates learned, fixed or other graph constructions, allowing the choice to be tailored to the target task. Accordingly, a uniform graph can still be a viable alternative without additional training, while learned conductances provide further improvements when beneficial.

Reviewer_VLZVOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

GPARA combines sparse measurement acquisition with reconstruction using a frozen conditional diffusion prior. A graph-Gaussian residual surrogate supplies a reusable posterior response for propagating measurement innovations and estimating weighted risk reduction from prospective measurements. A bounded learned residual calibrates the analytic acquisition score. A small cached prior ensemble and matrix-free solves avoid repeated diffusion posterior sampling during acquisition. Experiments on radio-field reconstruction and RGB-guided depth completion separate fixed-observation refinement from active acquisition and include controls with a shared reconstruction backend. The empirical framework is convincing, although the learned graph's incremental benefit is modest and the theoretical ranking result does not directly characterize the deployed pipeline.

Critical strengths

S1. The experiments separate measurement selection from reconstruction quality

Table 2 (p. 8) compares GPARA with AdaSense using the same graph reconstruction backend. RMSE decreases from 0.018 to 0.014 on radio fields and from 0.268 to 0.236 on depth completion, supporting an acquisition benefit beyond changing the reconstructor. Table 4(b) (p. 9) also fixes the backend while varying local, analytic, weighted and learned acquisition scores. The calibrated score outperforms the directly learned scorer over complete acquisition, which supports retaining an analytic anchor in this system.

S2. The paper checks sampling-budget sensitivity and accounts for numerical costs

Appendix A, Table 5 (p. 14), shows that several competing solvers improve substantially with more reverse steps. At 500 steps, DPS reaches 0.018 RFR RMSE versus 0.016 for GPARA, whose result saturates near 20 steps. This makes the few-step advantage interpretable rather than hiding the stronger baseline. Table 3 (p. 9) reports both NFE and actual acquisition time: GPARA reduces NFEs to 20, but is candidly slower than ADS and only about 1.71 times faster than AdaSense with the graph backend.

Critical weaknesses

W1. The incremental role of learning the graph is smaller than the overall improvement

With step-wise injection fixed, the uniform graph already reaches RMSE 0.0195 and 0.267, versus 0.0191 and 0.262 for the learned graph (Table 4(a), p. 9). Thus, much of the improvement comes from structured propagation and repeated correction, while learning conductances adds roughly two percent. The posterior and utility identities are meaningful within the surrogate, but the paper itself notes the connection to AdaSense's LMMSE criterion (p. 4). The distinct contribution should therefore be attributed carefully among the learned graph, the reusable implementation and the acquisition calibration.

W2. Actual acquisition ranking is only weakly explained by the matched-model result

Proposition 2 and Corollary 1 (p. 5) assume the matched Gaussian surrogate and the unclipped update; deployment includes finite solves, clipping and batching (Appendices B.5–B.7, pp. 17–19). In the depth diagnostic, the calibrated utility has Spearman correlation 0.15 ± 0.24 with actual one-step benefit, below 0.23 ± 0.27 for the direct learned scorer (Table 6, p. 14). The calibrated policy still wins in full-budget RMSE, and the authors discuss why random-prefix and closed-loop states differ (pp. 14–15). Nonetheless, the diagnostic leaves the claimed acquisition–refinement alignment in the actual pipeline incompletely validated.

Overall assessment

3 / 4. Weak acceptance. The shared posterior response is useful in practice, and fixed-backend acquisition controls plus the larger-step comparison support a substantive contribution on both tasks. My reservations concern attribution and how far the surrogate theory explains the deployed system, rather than a demonstrated failure of the main empirical result.

Questions for authors

Q1. How stable are the learned-versus-uniform graph differences in Table 4(a) (p. 9) across graph-training seeds? Please clarify which sources of randomness the five independent seeds in Tables 1–2 cover. This would help establish how much of the contribution requires learning the conductances.

Q2. Can the one-step ranking diagnostic be evaluated on prefixes visited by the deployed policies, using the same approximate solves and batching as deployment? Table 6 and the discussion (pp. 14–15) suggest that state distribution may explain the gap between weak one-step alignment and stronger final RMSE; evidence on those states would strengthen that explanation.

Q3. What is the end-to-end accuracy–time comparison when cached-prior generation, acquisition and final reconstruction are all included? Table 3 (p. 9) measures acquisition only, while Table 5 (p. 14) changes reconstruction steps without reporting total wall time. A matched-time comparison would clarify the practical benefit of the reusable response.

2 comments

Response (W1 & Q1): Thanks for your question. To assess training-seed stability, we retrained five DC graph predictors independently under different seeds, and evaluated them under the same conditions as Table 5(a). The relative RMSE reduction over the uniform graph is 1.86%±0.011%, supporting a small but stable benefit across the tested DC training runs.

Regarding the original five seeds, Table 1 varies the random observation masks, simulated measurement noise, and reconstruction sampling randomness, including the initial and reverse-step diffusion noise. In Table 2, the runs additionally vary the randomness used for acquisition, including the cached diffusion-prior samples, the conditional posterior samples used by sampling-based baselines, and the stochastic probes used to estimate graph utilities.

Furthermore, we also added a uniform-graph variant of the complete active-acquisition pipeline in revised Table 5(b). The uniform variant achieves RMSEs of 0.0146 on RFR and 0.243 on DC, compared with 0.0142 and 0.236 for learned graph, corresponding to reductions of 2.7% and 2.9%. These results confirm additional gains in both refinement and acquisition, although its contribution remains incremental. Accordingly, structured propagation and repeated refinement account for the main reconstruction gains, while learned conductances provide additional task-dependent improvements.

Response (Q3): Thanks for your suggestion. We have added an end-to-end accuracy-time comparison in Table 4, including cached-prior generation, acquisition, and final reconstruction. Time limits and RMSE targets are enforced per image before averaging, and time-to-target results are reported only when all images meet the threshold.

It is observed that although GPARA takes longer than ADS to acquire the same number of measurements from Table 3, it achieves the lowest mean RMSE under every matched end-to-end time budget, including a 32.8% reduction relative to ADS at 8 s. Compared with AdaSense + Graph, the strongest baseline, its RMSE reduction increases from 5.7% at 2 s to 19.6% at 8 s. Conversely, GPARA reaches every evaluated RMSE target with the shortest mean total runtime, achieving 1.48-2.24× speedups over AdaSense + Graph. These results demonstrate improved accuracy at matched runtime and reduced total runtime at matched accuracy, complementing the acquisition-only comparison in Table 3.

Table 4: End-to-end accuracy–time comparison under fixed time budgets and target RMSE levels.

MethodRMSE @ ≤2 s (m)RMSE @ ≤4 s (m)RMSE @ ≤6 s (m)RMSE @ ≤8 s (m)Time to ≤0.35 m (s)Time to ≤0.40 m (s)Time to ≤0.45 m (s)Time to ≤0.50 m (s)
ADS0.3470.3440.3430.341———8.71
AdaSense + DPS0.3500.3490.3480.347———32.1
AdaSense + DDNM0.3480.3430.3390.33638.7816.656.561.97
AdaSense + FPS0.3480.3440.3400.33640.6417.837.182.29
AdaSense + DiffPIR0.3470.3420.3380.33438.1316.036.572.21
AdaSense + Graph0.3320.3140.2990.2854.752.241.040.43
GPARA (Ours)0.3130.2800.2530.2292.121.010.490.29

Response (W2 & Q2): Thanks for your suggestion. We extended the diagnostic to prefixes visited by both the calibrated and direct learned policies in Table 7.

The results reveal prefix-dependent behavior rather than uniform one-step superiority. Calibration improves Spearman correlation over the deployed analytic score in six task-prefix combinations and reduces mean regret in five. However, the direct scorer has higher mean correlation on random and calibrated-policy prefixes, whereas calibration has higher correlation on direct-policy prefixes; their regret comparisons are also mixed. Thus, policy-generated states confirm the importance of the observation-state distribution.

We additionally complete deployed acquisition trajectories in Fig. 8 (with the results summarized below). The direct policy has lower mean RMSE at 48 observations, but calibration achieves lower mean RMSE at every reported budget from 144 observations onward. Its advantage increases to approximately 0.03 m at 440 observations, where it outperforms the direct policy on all 20 images. These results distinguish local one-step ranking from cumulative closed-loop performance: a locally better measurement choice does not necessarily yield a better complete acquisition trajectory. We have revised the discussion to distinguish prefix-dependent single-measurement diagnostics from full-policy reconstruction performance, rather than attributing the latter solely to random-prefix mismatch.

Table 7: Alignment between acquisition scores and empirical one-step reconstruction benefit under different prefixes.

PrefixesScoreRFR Spearman ↑RFR Regret (×10−5\times10^{-5}) ↓DC Spearman ↑DC Regret (×10−4\times10^{-4}) ↓
RandomHigh-accuracy analytic0.42±0.260.42 \pm 0.261.26±1.451.26 \pm 1.450.19±0.270.19 \pm 0.279.23±15.609.23 \pm 15.60
RandomDeployed analytic0.35±0.270.35 \pm 0.271.14±1.351.14 \pm 1.350.11±0.250.11 \pm 0.259.51±15.529.51 \pm 15.52
RandomCalibrated utility0.40±0.250.40 \pm 0.251.06±1.251.06 \pm 1.250.15±0.240.15 \pm 0.243.60±4.97\mathbf{3.60 \pm 4.97}
RandomDirect learned scorer0.44±0.29\mathbf{0.44 \pm 0.29}0.94±1.28\mathbf{0.94 \pm 1.28}0.23±0.27\mathbf{0.23 \pm 0.27}6.08±14.756.08 \pm 14.75
Calibrated-policyHigh-accuracy analytic0.18±0.200.18 \pm 0.200.68±0.910.68 \pm 0.910.08±0.230.08 \pm 0.232.67±2.092.67 \pm 2.09
Calibrated-policyDeployed analytic0.10±0.170.10 \pm 0.170.67±0.870.67 \pm 0.87−0.05±0.18-0.05 \pm 0.182.90±2.012.90 \pm 2.01
Calibrated-policyCalibrated utility0.16±0.180.16 \pm 0.180.64±0.79\mathbf{0.64 \pm 0.79}0.02±0.150.02 \pm 0.153.12±2.613.12 \pm 2.61
Calibrated-policyDirect learned scorer0.22±0.23\mathbf{0.22 \pm 0.23}0.71±0.930.71 \pm 0.930.19±0.25\mathbf{0.19 \pm 0.25}2.51±2.17\mathbf{2.51 \pm 2.17}
Direct learnedHigh-accuracy analytic0.10±0.200.10 \pm 0.200.83±0.770.83 \pm 0.770.12±0.22\mathbf{0.12 \pm 0.22}3.92±3.323.92 \pm 3.32
Direct learnedDeployed analytic0.09±0.180.09 \pm 0.180.78±0.770.78 \pm 0.770.06±0.150.06 \pm 0.153.99±2.873.99 \pm 2.87
Direct learnedCalibrated utility0.14±0.17\mathbf{0.14 \pm 0.17}0.77±0.750.77 \pm 0.750.10±0.150.10 \pm 0.152.45±1.94\mathbf{2.45 \pm 1.94}
Direct learnedDirect learned scorer0.06±0.230.06 \pm 0.230.67±0.78\mathbf{0.67 \pm 0.78}0.08±0.170.08 \pm 0.174.34±3.544.34 \pm 3.54

Figure 8: Mean RMSE across 20 images as each policy follows its own acquisition trajectory. Paired differences are Calibrated minus Direct; negative values indicate lower RMSE for Calibrated. The final column counts images on which Calibrated achieves lower RMSE.

Observed pointsCalibrated RMSE (m)Direct RMSE (m)Paired difference (m)95% CI (m)Calibrated lower
10.347150.34686+0.00028[+0.00011, +0.00047]6/20
480.328140.32517+0.00297[+0.00048, +0.00558]6/20
1440.298680.30185-0.00317[-0.00611, +0.00040]16/20
2480.274200.28589-0.01170[-0.01650, -0.00645]19/20
3440.256480.27806-0.02157[-0.02638, -0.01680]19/20
4400.241270.27110-0.02982[-0.03544, -0.02447]20/20
Reviewer_VDmLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

GPARA addresses active grounding of frozen diffusion priors with sparse measurements. It learns a graph surrogate over prediction residuals and reuses the same graph-induced posterior response for both measurement refinement and active acquisition, while keeping the diffusion backbone frozen. Its main contributions are graph-structured residual refinement that propagates local measurement innovations across coupled variables, response-aligned acquisition based on weighted posterior-risk reduction with bounded residual calibration, and a matrix-free implementation using cached samples, stochastic probing, and sparse solves. Experiments on radio field reconstruction and RGB-guided depth completion show strong gains: under fixed observations, RMSE drops by 36.7% and 23.4% relative to the strongest baselines; under a matched active budget, GPARA beats the strongest active baseline by 46.2% and 29.3%, and still beats same-backend AdaSense+Graph by 22.2% and 11.9%. Ablations support step-wise graph correction and weighted analytic utility, while efficiency results show 20 NFE versus 1240 for AdaSense variants and 1.71× faster than AdaSense+Graph, though wall-clock time remains slower than ADS. The main limitation is that the theory covers a graph-Gaussian surrogate and one-step acquisition, leaving nonlinear observations and longer-horizon sensing open.

Critical strengths

The paper’s core strength is a shared graph-posterior response that jointly drives refinement and acquisition. Proposition 2 and Corollary 1 show that weighted posterior-risk reduction aligns with expected one-step reconstruction improvement under a matched graph-Gaussian surrogate. This lifts the contribution beyond an empirical pipeline and makes the acquisition-refinement coupling principled. Validation is also persuasive because the same-backend AdaSense+Graph control and ablations isolate the proposed acquisition strategy: GPARA still improves RMSE by 22.2% and 11.9% on RFR and depth completion, while ablations support step-wise graph correction, weighted analytic utility, and bounded calibration. The matrix-free design further shows practical efficiency. Together, these results support that the gains come from the proposed mechanism rather than a stronger backbone or uncontrolled experimental advantage.

Critical weaknesses

The paper’s central theoretical claim does not cover the deployed algorithm. Proposition 2 and Corollary 1 establish alignment only for a matched graph-Gaussian surrogate and an unclipped one-step update, but deployment uses clipping, finite Jacobi/PCG solves, a biased finite-probe ratio estimator, learned bounded calibration, and batched acquisition. No error bounds or guarantees connect the analytic utility to the actual implementation. This is consequential because the paper’s main conceptual contribution is the principled coupling of acquisition and refinement; if that coupling is guaranteed only in an idealized surrogate, the central claim rests mainly on two-task empirical results rather than on theory that validates the method as run.

Questions for authors

Could the authors add a brief discussion or diagnostic showing how the calibrated acquisition score behaves when the matched-surrogate assumptions are only approximately satisfied? Deployment involves clipping, finite Jacobi/PCG solves, finite-probe estimation, bounded calibration, and batching. A small additional diagnostic—for example, how ranking quality or regret changes with probe count, solve depth, or batch size—would help readers understand the practical reach of the theoretical alignment, without undermining the paper’s main message. Could the authors comment on how task-specific the learned graph predictor and residual scorer are, and whether there are signs of transferability or robustness? Since GPARA trains these components offline per task, a short discussion of sensitivity to key hyperparameters, or any preliminary evidence of cross-task transfer or few-shot adaptation, would help clarify whether the framework is intended as a general deployment-time grounding approach or as a strong two-task instantiation. Either answer would be useful and would not weaken the contribution. Could the authors clarify the intended scope of the efficiency claim? The results show clear advantages over AdaSense-style repeated-ensemble methods in NFE, memory, and wall-clock relative to AdaSense+Graph, while wall-clock remains slower than ADS. A brief statement of the intended comparison regime would help readers interpret the efficiency results fairly and appreciate the practical value of the matrix-free design.

3 comments

Response (W1 & Q1): Thank you for your question. We agree that Proposition 2 and Corollary 1 establish the acquisition-refinement alignment under the matched graph-Gaussian surrogate and an unclipped single-measurement update, rather than guaranteeing exact alignment in the complete deployed pipeline.

Table 7 shows that this empirical one-step alignment is state-dependent, and calibrated acquisition consistently improves the analytic score across the evaluated prefix distributions. We furthter add a numerical sensitivity diagnostics in Fig. 7 (with the results summarized below), varying the probe count and PCG iteration cap. It shows that increasing the probe count generally improves ranking correlation and selection regret, whereas increasing the PCG iteration cap has relatively little effect within the tested range. These results indicate that practical acquisition alignment depends not only on the current state but also on the accuracy of the estimated posterior response.

Overall, our theory provides a principled basis for coupling refinement and acquisition, while its empirical alignment under deployment depends on the observation state and numerical estimation accuracy. We do not claim exact theoretical guarantees for the complete batched pipeline; instead, our results demonstrate that the shared posterior-response mechanism remains practically useful under the evaluated deployment settings.

Figure 7: Sensitivity of calibrated acquisition scores to probe count and PCG iteration cap. Values are mean changes relative to the defaults: RFR (P,C)=(16,24)(P,C)=(16,24) and DC (4,24)(4,24). Higher Δ\DeltaSpearman and lower regret change indicate improvement.

(a) Varying probe count; PCG cap fixed at 24.

Probe countRFR ΔSpearman (×10⁻³)RFR regret change (%)DC ΔSpearman (×10⁻³)DC regret change (%)
4——0.000.00
8-28.24-1.65-33.00-1.27
160.000.00-16.65+3.24
32+17.00-16.29+20.41+2.68
64+25.68-8.59+37.62-1.66
128+41.03-22.04+50.15-3.03

(b) Varying PCG iteration cap; probe count fixed at 16 for RFR and 4 for DC.

PCG capRFR ΔSpearman (×10⁻³)RFR regret change (%)DC ΔSpearman (×10⁻³)DC regret change (%)
240.000.000.000.00
32+0.91-1.87-3.15-1.13
48+3.47-1.87+0.76-1.39
96+3.18-1.87+0.32-1.39
128+3.18-1.87+0.32-1.39

Response (Q2): Thank you for your suggestion. We clarify that GPARA is a general grounding framework rather than a universally transferable set of network weights. Its generality lies in the reusable graph-residual posterior response that connects measurement-induced refinement with prospective measurement valuation.

Specifically, the graph predictor learns context-dependent residual dependencies conditioned on the frozen diffusion prior, while the residual scorer calibrates analytic acquisition utilities against realized reconstruction benefits. Both networks are trained offline with the diffusion prior frozen. Consequently, direct weight transfer across substantially different tasks or modalities is not generally expected, even when network architectures are compatible. For related tasks with similar conditioning semantics, diffusion priors, and residual structures, pretrained components may provide useful initializations for few-shot adaptation; otherwise, task-specific training is more appropriate. Thus, GPARA provides a general grounding framework through reusable posterior inference and acquisition mechanisms, while allowing its learned components to be adapted to individual tasks.

We have added a hyperparameter sensitivity study in Fig. 6 (with the results summarized below), covering correction strength ρg\rho_g, graph-coupling strength λ\lambda, and acquisition-calibration bound β\beta. The results reveal task-dependent correction requirements, while stronger graph coupling consistently benefits both tasks within the evaluated range, with diminishing returns. The calibration bound exhibits consistent behavior across tasks, with the default setting providing a favorable balance between the analytic anchor and learned residual correction. Overall, these results indicate that the key hyperparameters should be tuned to the target task rather than universally shared.

Figure 6: Sensitivity to key hyperparameters: (a) graph-correction strength ρg\rho_g; (b) graph-coupling strength λ\lambda; and (c) acquisition-calibration scale β\beta. Values are the plotted RMSE percentage changes relative to the default configurations; lower values indicate better reconstruction.

(a) Graph-correction strength.

ρg/ρgdep\rho_g/\rho_g^{\mathrm{dep}}RFR RMSE change (%)DC RMSE change (%)
0.5+14.59+12.28
0.75+4.26+5.20
10.000.00
1.25+0.07-3.69
1.35+1.58—
1.5—-6.14
1.75—-7.43
1.9—-7.64

(b) Graph-coupling strength.

λ/λdep\lambda/\lambda^{\mathrm{dep}}RFR RMSE change (%)DC RMSE change (%)
0.5+2.78+1.50
0.75+0.93+0.51
10.000.00
1.25-0.56-0.31
1.5-0.93-0.52
2-1.40-0.79
2.5-1.68—
3—-1.06
3.75-2.05—
4—-1.19
5-2.23—
8—-1.39
10-2.51—
16—-1.49
20-2.65—

(c) Acquisition-calibration scale.

β\betaRFR RMSE change (%)DC RMSE change (%)
0.5+3.10+6.52
1+1.01+3.77
1.5+0.49+0.95
20.000.00
2.5-0.05+1.13
3+0.89+1.59
3.5+3.08+2.05

Response (Q3): Thanks for your question. Our efficiency claim concerns the trade-off among computational cost, memory usage, and reconstruction accuracy, rather than universally lower runtime. In Table 3, GPARA uses 20 sample-equivalent NFEs versus 1240 for AdaSense + Graph and achieves a 1.71× acquisition-stage speedup, with comparable peak memory. Although ADS is faster in acquisition, GPARA uses 54.6% less peak memory. The matrix-free design avoids dense covariance storage, but posterior computations still incur runtime overhead.

To further assess the accuracy-time trade-off, we added Table 4 to measure end-to-end runtime, including cached-prior generation, acquisition, and final reconstruction on DC. GPARA achieves the lowest mean RMSE under every matched end-to-end time budget, including a 32.8% reduction relative to ADS at 8 s. It also reaches every evaluated RMSE target with the shortest mean total runtime. Together, these results establish the practical benefit in terms of computational and memory requirements alongside reconstruction quality. GPARA reduces repeated diffusion sampling and delivers a more favorable end-to-end accuracy-time trade-off, making it particularly suitable for resource-constrained tasks.

Table 4: End-to-end accuracy–time comparison under fixed time budgets and target RMSE levels.

MethodRMSE @ ≤2 s (m)RMSE @ ≤4 s (m)RMSE @ ≤6 s (m)RMSE @ ≤8 s (m)Time to ≤0.35 m (s)Time to ≤0.40 m (s)Time to ≤0.45 m (s)Time to ≤0.50 m (s)
ADS0.3470.3440.3430.341———8.71
AdaSense + DPS0.3500.3490.3480.347———32.1
AdaSense + DDNM0.3480.3430.3390.33638.7816.656.561.97
AdaSense + FPS0.3480.3440.3400.33640.6417.837.182.29
AdaSense + DiffPIR0.3470.3420.3380.33438.1316.036.572.21
AdaSense + Graph0.3320.3140.2990.2854.752.241.040.43
GPARA (Ours)0.3130.2800.2530.2292.121.010.490.29

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.