acceptodds
Under review as a conference paper at ICLR 2027

From Heatmaps to Reliance Fields: Diagnosing Visual Explanations

Abstract

Visual explanation methods are widely consumed under the assumption that spatial heatmaps directly expose model reliance. However, what an explanation highlights, what functionally supports a prediction under intervention, how evidence propagates across internal representations, and what is semantically meaningful are fundamentally non-interchangeable. We prove in Theorem 1 that finite-order local attribution signals do not uniquely determine finite-intervention reliance. This conflation obscures critical failure modes, including highlighted-but-unused evidence, hidden reliance, and potential shortcut cues. To overcome this limitation, we introduce Reliance Uncertainty Fields (RUF), which formulate latent functional reliance as a layer-wise posterior distribution over an Energy-Based Factor Graph. RUF couples local preference, finite-intervention influence, representation flow, stability, and optional semantic references into a unified probabilistic field. Building on RUF, we propose RIFT (Reliance Inference over Flow-based Traces), a diagnostic framework that computes posterior reliance means, marginal entropy (uncertainty), preference–function disagreement, flow attenuation, and shortcut candidates. RIFT reframes visual explanation from generating plausible heatmaps to inferring operational functional reliance under a specified intervention family and explicit uncertainty.

Primary area:
learning theory
Keywords:
ML: Deep Learning Theory & Learning Theory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6kcMOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper distinguishes visual attribution from functional reliance and introduces Reliance Uncertainty Fields (RUF), an energy-based model over layer-wise representation graphs. Its RIFT framework combines attribution preferences, finite-intervention responses, representation flow, stability, and optional semantic references to infer reliance scores and marginal entropy. These quantities support diagnostics of hidden reliance, highlighted-but-unused evidence, attenuation, and shortcut candidates. A non-identifiability theorem motivates supplementing finite-order local attribution with intervention evidence. Experiments across classification, corruption, dense prediction, medical imaging, and tracking report improved prediction of held-out intervention responses and mismatch detection, while sparse probing reduces the cost of collecting evidence.

Critical strengths

1、The distinction between attribution preference and intervention-defined reliance is well motivated, and the local non-identifiability result provides a clear theoretical rationale for collecting additional evidence.

2、RIFT integrates complementary evidence across representation layers and produces explicit diagnostic scores, offering a useful framework for investigating hidden reliance and highlighted-but-unused evidence beyond conventional heatmaps.

3、The evaluation spans multiple vision domains and backbone families, includes component ablations, and separates semantic-free results from reference-assisted diagnostics, helping assess the framework’s breadth and contributions.

Critical weaknesses

1、In Section 3, marginal entropy depends only on the aggregated posterior mean; conflicting and agreeing evidence can yield identical entropy, so its interpretation as evidence conflict is insufficiently justified.

2、Flow uses feature-change magnitude without class-specific relevance, while outgoing normalization largely removes absolute propagation strength. These choices weaken the interpretation of coupling and attenuation as prediction-supporting evidence flow.

3、The projection ablation changes node-level functional predictiveness from 0.86 to 0.78 (Table 1), but the manuscript does not clarify whether input evidence or the evaluation domain also changes.

4、Equation (6) rewards stability independently of response magnitude, allowing consistently zero intervention effects to increase inferred reliance. This conflates evidence reliability with positive evidence of functional support.

Questions for authors

1、How should entropy distinguish conflicting from agreeing evidence when both produce the same unary score and graph messages? Please clarify the intended interpretation and provide a controlled comparison separating evidence conflict from reliance magnitude.

2、How does the flow potential exclude feature changes irrelevant to the target prediction? If two nodes have proportional outgoing disturbances with very different total magnitudes, what enables the attenuation diagnostic to distinguish their propagation strength?

3、What exactly changes in “RIFT w/o RF-aware projection”? Please specify whether preference aggregation, posterior inference, node selection, or the evaluation domain changes.

2 comments
AuthorsOpenReview ↗

Official Comment (1/2)

CW1 / Q1 (Conflicting and agreeing evidence ...)

We use separate readouts for marginal uncertainty and source disagreement. Our entropy is a function of the fused Bernoulli mean, so equal means necessarily give equal entropy. We do not use entropy to reconstruct which sources produced that mean. In D.7, Eq. (81), we separately define preference–function disagreement as DJS(ppref∥pfunc)D_{\mathrm{JS}}(p_{\mathrm{pref}}\Vert p_{\mathrm{func}}), and Eq. (82) compares preference and RUF views. These source-conditioned distributions retain a distinction that is absent from the fused mean alone. When referring to uncertainty, we mean model-relative marginal uncertainty; source conflict itself is assessed through the separate disagreement readout.

For the controlled comparison, consider the unary-only case with graph messages and other contributions set to zero, using βpref=1\beta_{\mathrm{pref}}=1, βfunc=2\beta_{\mathrm{func}}=2. The views are ppref(R=1)=σ(zpref)p_{\mathrm{pref}}(R=1)=\sigma(z_{\mathrm{pref}}) and pfunc(R=1)=σ(2zfunc)p_{\mathrm{func}}(R=1)=\sigma(2z_{\mathrm{func}}). With (zpref,zfunc)=(1,0.5)(z_{\mathrm{pref}},z_{\mathrm{func}})=(1,0.5), both views equal 0.731059 and JS divergence is zero. With (zpref,zfunc)=(0,1)(z_{\mathrm{pref}},z_{\mathrm{func}})=(0,1), the views are 0.5 and 0.880797 and JS divergence is 0.089541 nats. Both cases have η=2\eta=2, fused mean 0.880797, and entropy 0.365334 nats. Thus our existing disagreement statistic distinguishes the weighted source views while reliance magnitude and marginal uncertainty are fixed. This is an algebraic comparison, not a new experiment or an interpretation of zero as negative binary evidence. We also evaluate uncertainty against prediction error: Table 1 reports correlation 0.56 for RIFT versus 0.42 for RP+FI. Our framework therefore reports the fused belief, its uncertainty, and source-view disagreement as distinct quantities.

CW2 / Q2 (Flow uses feature-change magnitude ...)

We use flow edges to represent feature propagation, and functional evidence and class-conditioned beliefs to represent target information. Eq. (7) combines these through a soft compatibility factor. As stated in K.3, a feature disturbance can propagate without changing the target score; we do not treat an edge's existence as a class-specific causal certificate. Our model also does not guarantee that every target-irrelevant edge is excluded. We evaluate the value of combining these signals through the graph comparisons in Table 7 and the attenuation diagnosis in Table 10. Empirical-flow graphs reach functional predictiveness 0.87 for CNNs and 0.86 for ViTs, and attenuation AUROC is 0.84 for RIFT versus 0.68 for RP+FI and 0.60 for FI-only.

We retain propagation magnitude separately from the normalized pattern. D.4 defines Ai→jA_{i\to j} in Eq. (63), outgoing mass Mflow(i)=∑jΦi→jcM_{\mathrm{flow}}(i)=\sum_j\Phi_{i\to j}^c in Eq. (64), and flow entropy in Eq. (65); we also record mass in E.5. For disturbances (1,3)(1,3) and (10,30)(10,30), neglecting the small normalization constant, both patterns are A=(0.25,0.75)A=(0.25,0.75) but the masses are 4 and 40. At local belief 0.8 and downstream beliefs (0.6,0.8)(0.6,0.8), both have downstream support 0.75 and attenuation 0.05. This follows from our definition of attenuation as the posterior-support gap in Appendix G. It does not measure absolute feature-amplitude decay. The separate mass readout distinguishes the total defined disturbance; we do not multiply it into Eq. (7) or claim that it makes attenuation scale-sensitive. The proportional example therefore yields equal support attenuation and different flow mass, exactly the two quantities our definitions separate.

AuthorsOpenReview ↗

Official Comment (2/2)

CW3 / Q3 (The projection ablation changes ...)

We distinguish input-side preference aggregation from output-side projection. In D.2, Eq. (48), we aggregate an image-space preference map into node evidence using zpref,ic=Norm⁡(∑pKi(p)AEc(x;p))z_{\mathrm{pref},i}^c=\operatorname{Norm}(\sum_p K_i(p)A_E^c(x;p)). When this input route is used, the kernel can change the unary evidence. In D.9, Eq. (87), we instead map inferred means to image coordinates using Cprojc(p)=∑iwimicKi(p)C_{\mathrm{proj}}^c(p)=\sum_i w_i m_i^cK_i(p). We infer the posterior on the representation graph, as stated in E.7, and define primary functional predictiveness over held-out representation nodes in Section 4. At fixed node means and evaluation nodes, changing only the output rendering cannot change that Spearman correlation.

Our direct projection experiment is the image-space comparison in J.10, Table 12. Bilinear resizing gives pointing-game accuracy 0.62, IoU@10% 0.31, and image-calibration loss 0.184; analytic RF gives 0.69, 0.37, and 0.142; effective RF gives 0.72, 0.40, and 0.128. These results support the localization and image-calibration benefit. For the Table 1 row, we have not specified the run-level aggregation and selection settings needed to identify whether the comparison also changed an input-side operation. The existence of Eq. (48) does not establish that it was toggled in that run. We therefore do not attribute the 0.86-to-0.78 node-level difference to an output-only rendering change, or assert an undocumented change in evaluation domain. We use Table 12 as the direct evidence for the projection contribution and keep that result separate from the configuration question concerning the Table 1 row.

CW4 (Equation (6) rewards stability ...)

We use stability as an auxiliary predictor of reproducibility across baselines, separately from functional magnitude. In Eq. (6), the two contributions are additive, so a consistently zero response can receive positive stability support. We do not identify invariance itself with a measured nonzero functional effect. The empirical rationale for including the channel is our component comparison: in Table 1, full RIFT improves functional predictiveness from 0.82 without stability to 0.86, RCE from 0.101 to 0.083, and uncertainty–error correlation from 0.43 to 0.56. We select the coefficients on validation data as described in D.6. These results support the auxiliary feature's predictive value in the evaluated setting.

The direct factor effect can be stated exactly. Holding other factor inputs and neighboring states fixed, a normalized stability increase s∈[0,1]s\in[0,1] multiplies conditional odds by exp⁡(βstabs)\exp(\beta_{\mathrm{stab}}s), at most e0.5e^{0.5} under our defaults. A normalized functional increment u∈[0,1]u\in[0,1] gives exp⁡(2u)\exp(2u). This is a conditional comparison, not a global bound after graph feedback, and the smaller stability coefficient does not eliminate the stable-zero effect. Our functional-variance penalty in D.5 is zero at zero functional variance; the preference-variance penalty concerns a different source. Neither acts as a zero-response detector. We consequently distinguish the component's measured predictive benefit from a guarantee that every stable-zero node is correctly classified. The ablation supports including reproducibility in the joint model, without changing what the observed functional response means.

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

RIFT represents functional reliance as a latent binary field over internal representation units. Local attribution and intervention evidence enter the field alongside stability terms and directed flow potentials. The marginals and entropies are used to diagnose missed reliance, unused highlights, and related failures. The evaluation tests these diagnostics against held-out intervention responses across several vision domains.

Critical strengths

The formulation makes its assumptions explicit. Reliance depends on the intervention family, replacement baseline, and representation graph, which gives model auditors concrete choices to examine. It also separates semantic alignment from functional reliance. That distinction is useful when a model legitimately uses context outside an object mask.

I find the use of held-out intervention responses to evaluate the diagnostics appropriate to the paper's aim. The held-out-node protocol and semantic-free evaluation help assess functional reliance even when a universal ground-truth explanation is unavailable. The additional sensitivity checks provide support for these choices.

Critical weaknesses

I am concerned that the field may not distinguish measured non-reliance from a lack of positive evidence. In Equation 6, a zero functional response adds zero to the unary score; it provides no evidence against reliance. Stable zero responses can even receive a positive stability contribution. Equation 7 then rewards joint reliance at neighboring nodes whenever an intervention changes downstream representations.

Consider two nodes with zero unary scores, one directed edge of weight 1, and the default flow and cut coefficients of 1.5 and 0.5. The four configuration weights are 1, 1, exp(-0.5), and exp(1.5). This gives exact reliance marginals of approximately 0.72 and 0.77 despite the lack of positive unary evidence. This calculation concerns the stated model; I have not reproduced the reported experiments.

For the probabilities and their entropy to support the proposed diagnostics, inference needs to account for repeated, stable zero output effects. I could not find how it rules out reliance when downstream representations change but the target output does not. The discussion in Appendix K.3 acknowledges this possibility, which reinforces the need to explain how the formulation handles it.

Questions for authors

Which term distinguishes a reliably measured zero intervention effect from missing or uninformative functional evidence? How does RIFT behave when interventions change downstream features but leave the target score unchanged? Please address the two-node example with a controlled negative case, or explain which assumption excludes it. This would help establish whether graph coupling can create apparent reliance on its own.

3 comments
AuthorsOpenReview ↗

Official Comment (1/2)

CW / Q1 (Which term distinguishes ...)

In our formulation, a zero unary score is neutral evidence for the latent state. Keeping the proposed setting ηi=ηj=0\eta_i=\eta_j=0, Eq. (6) gives ψi(0)=ψi(1)=ψj(0)=ψj(1)=1\psi_i(0)=\psi_i(1)=\psi_j(0)=\psi_j(1)=1. If a measured zero and a missing or uninformative entry produce the same factor inputs, this equation assigns the same potential. We do not include a separate measurement-status indicator or a negative contribution triggered specifically by an observed zero. Under the stipulated inputs, the local factor therefore does not distinguish those two records.

We define the quantity being inferred in Definition 1 and Section 3: the reliance state is latent, observations provide noisy evidence, and RUF is a joint belief conditioned on the specified energy and graph. A neutral local factor can consequently have a non-neutral marginal once the pairwise factor is included. We use “posterior” in this model-relative sense, as stated on p. 4, lines 197–202. This explains why the marginal need not equal the observed intervention response. It does not make a measured zero negative evidence, and we do not claim that this definition alone ensures low reliance for the all-zero case. For that case, the factor values and resulting marginals are exactly those calculated in our response to Q3.

CW / Q2 (How does RIFT behave ...)

We measure target-score response and downstream feature disturbance separately. Retaining zero target responses, zero unary scores, and a unit edge, Eq. (7) gives ψij(0,0)=ψij(0,1)=1\psi_{ij}(0,0)=\psi_{ij}(0,1)=1, ψij(1,0)=e−0.5\psi_{ij}(1,0)=e^{-0.5}, and ψij(1,1)=e1.5\psi_{ij}(1,1)=e^{1.5}. The observed zero target effects do not switch off that edge, because we construct its weight from feature distance in D.4. As discussed in K.3, feature propagation is not itself evidence of a measured nonzero target effect. Nevertheless, our pairwise factor favors joint latent reliance in this configuration.

The conditional odds show precisely how it acts. With the unary scores fixed at zero, p(Ri=1∣Rj=0)/p(Ri=0∣Rj=0)=e−0.5p(R_i=1\mid R_j=0)/p(R_i=0\mid R_j=0)=e^{-0.5}, whereas p(Ri=1∣Rj=1)/p(Ri=0∣Rj=1)=e1.5p(R_i=1\mid R_j=1)/p(R_i=0\mid R_j=1)=e^{1.5}. These conditions refer to the latent neighboring state. Observing a zero target-score response does not set that state to zero. Our cut term therefore penalizes the upstream-reliant/downstream-nonreliant configuration; it is not a measured-zero rejection rule. Accordingly, RIFT can return elevated reliance marginals in the target-null case you describe. We interpret those values as estimates under the stated structural compatibility model, and we do not claim that Eq. (7) provides a class-relevance gate that excludes every target-null edge.

AuthorsOpenReview ↗

Official Comment (2/2)

CW / Q3 (Please address the two-node example ...)

For your two-node setting, we use ηi=ηj=0\eta_i=\eta_j=0, Ai→j=1A_{i\to j}=1, λflow=1.5\lambda_{\mathrm{flow}}=1.5, and λcut=0.5\lambda_{\mathrm{cut}}=0.5, with partition function Z=2+e−0.5+e1.5\mathcal Z=2+e^{-0.5}+e^{1.5}. The exact Gibbs marginals are

mi=e−0.5+e1.52+e−0.5+e1.5=0.717842,mj=1+e1.52+e−0.5+e1.5=0.773352.m_i=\frac{e^{-0.5}+e^{1.5}}{2+e^{-0.5}+e^{1.5}}=0.717842,\qquad m_j=\frac{1+e^{1.5}}{2+e^{-0.5}+e^{1.5}}=0.773352.

These values arise entirely from the pairwise factor under neutral unary support. Our mean-field update also retains the effect: ten synchronous Eq. (8) updates from 0.5, without early stopping, give approximately 0.757615 and 0.819830. These are direct calculations from our equations, not additional empirical results. We do not impose an assumption in Eqs. (6)–(7) that excludes this graph. As we state in Section 3 and K.12, our posterior is conditional on the chosen evidence model and graph, so the calculated values are consistent with the distribution we define. This model-relative interpretation explains the inference; it does not establish its functional accuracy for these null nodes. We call the marginal the reliance readout in D.12 and relate the latent state to prediction support in Definition 1. We therefore do not claim that the all-zero example is resolved merely because the Gibbs distribution is well defined. Under the unchanged conditions, our current equations permit the reported increase and do not guarantee rejection of the null nodes.

I appreciate the effort you put into the rebuttal and the new experiments. I agree the math checks out under your setup: even with zero measured target effects and neutral local scores, a single flow edge yields exact reliance probabilities of 0.718 and 0.773. Where I still get stuck is what an auditor is actually supposed to conclude from those numbers. If you intervene on a node repeatedly and the target score never moves, it still gets the same local evidence as a node you never measured at all, and feature propagation by itself can then push its reliance score up anyway. Since an auditor cannot tell those two situations apart, it undercuts the hidden-reliance diagnosis. I am keeping my score at 2 (weak rejection).

Reviewer_79e9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that attribution maps shouldn’t be read as functional reliance. It proves a non-identifiability result (Theorem 1) and proposes RIFT, a mean-field MRF over binary reliance variables on a multi-layer representation graph. The MRF combines attribution, intervention logit drops, stability, and optional semantic masks, with pairwise flow terms between layers. It reports gains over CAM/gradient baselines across datasets and backbones.

Critical strengths
  1. The distinction between visual attribution and actual model reliance is well motivated.

  2. Combining intervention evidence with representation flow is interesting and provides more insight than standard heatmaps.

Critical weaknesses
  1. The paper would benefit from a clearer explanation of why the posterior mean should be interpreted probabilistically rather than as a graph-regularized evidence score.

  2. The evaluation is closely tied to the evidence used by the method. Functional predictiveness is primarily evaluated against held-out intervention responses, while intervention measurements are also a central input to RIFT. The authors address direct leakage through disjoint probing and evaluation nodes, which is appreciated. However, the training and evaluation signals still share the same intervention-based definition of reliance.

  3. CAM and gradient baselines never observe feature interventions, so Table 1’s margin over them is expected by construction.

  4. The method is computationally expensive compared to standard attribution approaches, which may limit its practical use.

Questions for authors

See [1] - [3] in weakness above.

2 comments
AuthorsOpenReview ↗

Official Comment (1/2)

CW1 / Q1 (The paper would benefit ...)

We use “posterior mean” because RUF is defined as a joint conditional distribution, and the reported values are marginals of its variational approximation. In Section 3, Eq. (3), we define p(Rc∣Zc,G,x)=Z−1exp⁡[−E(Rc;Zc,G,x)]p(R^c\mid Z^c,G,x)=\mathcal Z^{-1}\exp[-E(R^c;Z^c,G,x)] over binary configurations, with the factors specified in Eqs. (5)–(7). Finite coefficients on a finite graph give a normalizable distribution. We approximate it using the factorized distribution in Eq. (8) and Appendix D.6, with mic=qi(Ric=1)m_i^c=q_i(R_i^c=1). Graph regularization and probabilistic inference are compatible here: the graph determines how evidence at connected nodes changes the marginal. A pointwise weighted evidence score does not capture that joint dependence.

By probability, we mean belief under this specified evidence model. We state this explicitly in Section 3, p. 4, lines 197–202, where we distinguish RUF from a posterior over a uniquely observable ground-truth explanation. We select the coefficients on validation data as described in D.6 and examine their sensitivity in Table 4. We then evaluate the resulting predictions against withheld intervention responses. In Table 1, RIFT achieves functional predictiveness 0.86 and RCE 0.083, compared with 0.81 and 0.108 for RP+FI. As defined in Appendix G, RCE measures numerical alignment with normalized held-out responses; it is not a frequency-calibration test against observed binary reliance labels. Our use of probabilistic terminology thus follows the explicit joint model and marginal inference, while our experimental claims concern the prediction and diagnostic performance of those model-relative beliefs.

CW2 / Q2 (The evaluation is closely tied ...)

We evaluate whether RIFT predicts intervention responses that were unavailable during inference. Using the same intervention family for defining and evaluating that quantity does not reveal the withheld values. In our protocol, Vprobe∩Veval=∅V_{\mathrm{probe}}\cap V_{\mathrm{eval}}=\varnothing, and responses on the evaluation nodes are excluded from posterior construction, candidate selection, graph smoothing, and hyperparameter selection (Section 3, p. 5, lines 220–227). We access those responses only after inference to compute the metrics. Our prediction target therefore matches the operational definition of reliance without reusing the response being predicted.

The controls show why this prediction is nontrivial. Under the same response-based target, Table 1 gives functional predictiveness 0.81 for RP+FI, 0.75 for RIFT without flow, and 0.86 for full RIFT. In Table 9, shuffling functional evidence reduces 0.86 to 0.41, while shuffling flow reduces it to 0.66. Thus performance depends on the correspondence between the evidence, nodes, and graph, rather than following automatically from the target definition. We also evaluate replacement-baseline sensitivity in Table 5. We describe this setting as held-out node prediction and sparse spatial generalization within an image and graph. The split prevents direct response reuse; it does not make connected nodes statistically independent or establish independent cross-image generalization. Within the setting we evaluate, the shared intervention family specifies the estimand and does not make the predictive comparison circular.

AuthorsOpenReview ↗

Official Comment (2/2)

CW3 / Q3 (CAM and gradient baselines ...)

We include intervention-informed comparisons precisely to separate access to intervention evidence from the benefit of structured inference. RP+FI in Table 1 already combines attribution preference with functional-intervention evidence. Relative to RP+FI, semantic-free RIFT improves functional predictiveness from 0.81 to 0.86, RCE from 0.108 to 0.083, mismatch AUROC from 0.82 to 0.87, and uncertainty–error correlation from 0.42 to 0.56. These differences cannot be explained by RP+FI lacking intervention evidence. We also include Ablation-CAM and explicitly mark its intervention access. Our evidence for the method therefore extends beyond the margin over attribution-only CAM and gradient methods.

The diagnostic comparisons give the same distinction. In Table 10, hidden-reliance AUROC is 0.76 for FI-only, 0.82 for RP+FI, and 0.93 for RIFT; highlighted-but-unused AUROC is 0.72, 0.80, and 0.88, respectively. Both FI-only and RP+FI use intervention information, yet the full framework achieves better diagnosis. Our component ablations further test how the information is integrated: Table 1 gives functional predictiveness 0.75 without flow and 0.82 without stability, compared with 0.86 for full RIFT. We use βsem=0\beta_{\mathrm{sem}}=0 in the default comparison, so semantic masks do not supply the gain. These intervention-informed comparisons and component ablations support our contribution beyond obtaining a perturbation map; the comparison with conventional heatmaps is one part of the evaluation, not its sole justification.

CW4 (The method is computationally expensive ...)

We evaluate the additional cost together with the diagnostic information it provides. On the shared benchmark in Table 1, RIFT takes 5.3 times Grad-CAM's runtime, compared with 7.8 for Score-CAM and 5.6 for Ablation-CAM. RIFT is therefore within the reported runtime range of these intervention-based alternatives. Our timing includes attribution, candidate selection, interventions, flow estimation, posterior inference, and diagnostic-map generation; Appendix H specifies this scope and the eight-H200 environment. We do not report an inference-only time that omits evidence collection.

We also evaluate a sparse operating point instead of relying on dense probing. Our default uses 20 candidates per selected layer, eight baseline samples, and up to ten mean-field iterations (D.6 and E.8). In Table 6, this sparse setting reaches Spearman agreement 0.84 with dense RIFT and Hidden-F1 0.79, while dense probing takes 14.6 times the default sparse runtime. Here 0.84 is agreement with the dense reference, and the runtime reference is default sparse RIFT; Table 1 instead uses Grad-CAM as its reference. In D.16, we separate intervention cost O(∣C∣B)O(|\mathcal C|B), flow comparisons O(∣EC∣B)O(|\mathcal E_{\mathcal C}|B), and mean-field cost O(TMF∣E∣)O(T_{\mathrm{MF}}|E|). Tables 1 and 10 quantify the diagnostic gains at this additional cost. Our practical contribution is this measured cost–quality trade-off for reliability auditing: a controlled probing budget, substantial agreement with dense analysis, and better diagnosis than the intervention-informed alternatives we evaluate.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.