acceptodds
Under review as a conference paper at ICLR 2027

PotencyGNN: Amortized Epistemic Potency with Dual-Feasible Certificates

Abstract

Auxiliary observations can reduce causal uncertainty under latent confounding even when effects remain only partially identified. We formalize this gain through realized observational epistemic potency: the reduction in sharp identified-set width obtained when moving from the treatment–outcome marginal distribution to the full observational distribution while keeping the causal graph, query, and model class fixed. Exact computation requires four response-function linear programs and becomes rapidly expensive as parent counts grow. We introduce PotencyGNN, a query-conditioned relational graph neural network with separate parent, child, and bidirected message-passing channels, together with a dual-certificate network that provides certified lower bounds on potency. On screened core-disjoint three-to-five- and exact-six-variable benchmarks, PotencyGNN attains held-out and , remaining competitive with the evaluated graph transformer and outperforming the other baselines. Under matched CPU inputs, PotencyGNN inference is faster than cached exact evaluation for individual queries and achieves higher batched throughput. Across certificate evaluations, all satisfy the numerical validity criteria, with yielding certified potency above the reporting threshold.

Primary area:
causal reasoning
Keywords:
causal inference, causal bounds, latent confounding, structural causal models, graph neural networks, amortized inference, dual certificates, epistemic potency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_d3p7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper defines realized observational epistemic potency as the reduction in sharp-bound width obtained from auxiliary observational variables. They introduce Potency GNN, a query-conditioned relational graph neural network, together with a dual-certificate network that provides certified lower bounds on potency. This method outperforms several baselines, and provides speedups over exact LP solving.

Critical strengths

The GNN architecture seems well formulated with their ablation results showing the need for their relation specific message passing. Using a separate dualcertificate network to provide conservative potency lower bounds, allows them to not solely rely on the point prediction of the GNN. As far as I know, this is a novel application of these existing techniques.

Critical weaknesses

As stated in the abstract: Exact computation requires four responsefunction linear programs and becomes rapidly expensive as parent counts grow, then claiming PotencyGNN makes the problem more scalable. However in the experimental setup, K is set for K ≤ 256 training, and in the results (section 5.5 and appendix H), when K is outside the training range, the prediction accuracy deteriorates significantly. So the speedups are significant but the quality in their predictions drop as well. If i understand correctly the method is performing best (accuracy wise) when the exact LP is still easily solvable.(roughly 10 ms for K=256 from figure 2). The experimental setup is on graphs that aren’t significant in size "We use ADMGs with three to five observed binary variables, at most two observed parents per variable in C, and K ≤ 256 joint response types; ", so if the motivation is scalability I would like to see results on larger graphs.

Questions for authors

Regarding the amortization analysis, the authors state "for PotencyGNN (Table 5), the training cost is amortized after approximately 21,700 subsequent queries when exact labels are already available, or 47,800 queries when label generation is included." Are there practical applications involving graphs of this size, that would require such a large number of queries?

1 comment

Thank you for the assessment and for stating the main limitation precisely. Results marked "new" were obtained during the discussion and are in the revised PDF.

W1. Your reading of the submitted results is correct, and the revision responds in two ways. The claim is corrected. In the original stress test, models trained with at most 256 response types reach R² of 0.580 and 0.668 at 512 and 1,024 and fall to −9.579 at 262,144 (Section 5.5). The revised abstract states that exact computation takes milliseconds per query on the primary benchmark, that its cost grows with the number of response types, and that PotencyGNN loses accuracy beyond 256 and that retraining improves it. Section 1 states the aim as a lower cost for each additional computation when many are needed, not solving problems too large for exact computation.

Retraining improves accuracy. New: adding cores with 512 to 2,048 response types to training raises held-out R² at 512, 1,024 and 2,048 from 0.795, 0.744, 0.652 to 0.872, 0.832, 0.859, while R² at up to 256 is 0.865 on the original held-out cores (Table 13). On held-out cores like those used for retraining, accuracy at 2,048 is thus close to that within the original training range. On stress-test graphs from other causal structures (overlapping cores left out, predictions averaged over five seeds), R² rises from 0.571, 0.742, −0.184 to 0.769, 0.800, 0.490; the last graphs have a three-parent confounded variable, unlike any training graph. The added cores also differ structurally, the analysis was not prespecified, and it stops at 2,048 (Appendix H). Cost at these sizes: cached exact computation takes 11.7, 21.9 and 28.5 ms per query, 17.4, 32.2 and 43.4 times as long as PotencyGNN (Table 14). Exact computation remains feasible there, so the gain is in cost per query, and the saving grows with response-space size: the estimated 402.2 s for labeling the added instances and one fit, assuming the earlier labels exist, is recovered after 36,441, 18,916 and 14,443 queries.

W2. Prediction is evaluated up to six observed variables: models trained on six-variable cores reach held-out R² of 0.779 on six-variable cores absent from training (Table 1, Appendix B). Beyond six, neither the definition of potency nor PotencyGNN is tied to graph size. The linear programs are defined for any number of variables, and their size is set by the response types of the confounded set, not by the number of variables; under parental closure and the support assumption, a larger graph reduces to its causal core without changing the bounds or potency (Section 5.1, Appendix C; checked on graphs with up to seven variables); and PotencyGNN applies the same maps at every node, so it takes larger graphs without new parameters. Cores with more than six variables are not tested, but two results point the same way: training on six-variable cores gave the 0.779 above, and retraining on five-variable cores with larger response spaces raised R² at 2,048 from 0.652 to 0.859. In both cases PotencyGNN learned the larger setting once it was represented in training, which suggests that larger cores can be learned in the same way once their exact labels are computed (Appendix G). The certificate network accepts at most five observed variables, because its inputs have 33 constraint rows (Appendix F).

Question. Yes, on graphs of this size one application would reach such counts directly: quantifying the sampling uncertainty of potency. Potency is computed from an estimated distribution and does not measure sampling uncertainty (Section 3); a bootstrap interval recomputes it once per resample. Even with 1,000 resamples, one tenth of the 10,000 replicates the paper uses for its model-comparison intervals (Appendix D), an interval needs 1,000 computations per query, so 22 queries exceed the 21,700 break-even and 48 the 47,800 that includes label generation. Screening candidate queries and comparing auxiliary sets multiply the counts further (Section 1). For scale, the empirical evaluation has 19 queries on five real datasets (Section 5.6), and the published-network study enumerates 1,811 candidates on subgraphs of at most five variables (Appendix E). No evaluated study is large enough to recover the training cost (Section 5.4); Table 5 gives the per-query saving for any study size, and at larger response spaces the break-even is 14,443 to 36,441 queries (Table 14).

Strengths. Thank you for noting both. Removing all graph edges at the same parameter count lowers held-out R² from 0.829 to 0.277, and symmetrizing the directed aggregates lowers it to 0.764 (Appendix D). The certificate is separate from the point prediction and, given the assumed graph, the support assumption, known P(V) and full-program feasibility, cannot overstate potency: across 79,200 held-out evaluations, all certificates satisfy the numerical validity criteria, and the 43.4% that give a positive bound have precision 1.0 (Section 5.7).

Reviewer_HAwyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposed a framework to study the epistemic potency: the reduction in the identified set width when moving from the marginal distribution to the full observational distribution. Computing this quantity requires solving four linear programs. The paper introduces PotencyGNN to estimate the potency efficiently and provide a certified lower bound.

Critical strengths
  • The paper proposed an amortized method to estimate the potency with a GNN. However, I'm not particularly familiar with that area to judge the significance and novelty of the proposed framework.
Critical weaknesses
  • The paper is a bit hard to follow. I think the overall presentation would improve if the technical terms were clearly explained. For example, the introduction section mentioned "auxiliary variable", "base regime'' and "full regime''. But those are not explained until later sections. And even in Section 3, the definition of the auxiliary variable is still a bit vague.

  • The motivation of the proposed work is unclear. First of all, the paper assumes all variables are binary. In that case, solving 4 linear programs usually does not pose a computational challenge. For the "base regime" with marginal distributions, it even has a closed-form solution. Also, I don't understand the motivation for the epistemic potency. In Maringgele and Etesami (2026), the potency is used to determine the optimal experiments for reducing the bounds. However, in this case, since we already have access to the auxiliary variables, what's the purpose of computing the potency instead of directly estimating the sharp bounds?

Questions for authors
  • Could the author please provide some intuition and a practical scenario for when the epistemic potency score is useful? It would also be helpful to demonstrate the challenge with the traditional method using 4 linear programs to estimate the potency score, and show improvement with the proposed framework.
3 comments

Thank you for the direct comments on presentation and motivation. Results marked "new" were obtained during the discussion and are in the revised PDF.

The paper contributes a definition of realized observational potency, with the mechanisms of the variables outside the confounded set held fixed; a graph neural network that predicts potency without solving the linear programs; a separate network whose outputs are repaired into certified lower bounds on potency; and their evaluation on causal cores absent from training.

W1. Agreed. The first paragraph of the revised introduction now defines each term. Auxiliary variables are the observed variables other than treatment and outcome, for example an instrument (Section 3: all observed variables except X and Y). The base regime requires a model to reproduce the observed treatment–outcome distribution; the full regime, the joint distribution of all observed variables. The confounded set is the one set of variables, containing treatment and outcome, to which latent confounding is confined; a variable's mechanism is its conditional distribution given its parents. The second paragraph defines response types, the deterministic rules by which variables in the confounded set respond to their parents. Both regimes share the graph, the query, the model class and the mechanisms outside the confounded set. Potency is the resulting reduction in the width of the identified interval.

W2, first part. Agreed that four programs are cheap here: 6.776 ms per query on the primary benchmark. The revision states the aim as lower cost per query when many computations are needed, not problems too large for exact computation (Section 1). New: if a directed treatment–outcome path lies inside the confounded set, the base interval is the closed form given by the treatment–outcome distribution alone, so two programs suffice. This sufficient condition covers the instrumental-variable graph and 181 of 440 cores (confirmed on 1,810 fresh instances). On the other 259 cores the fixed mechanisms narrowed the base interval on all 2,590 fresh instances, by at least 0.021, so the base programs are still needed. On the 181 cores, two programs lower exact latency by a median of 44% per instance, and the benchmark median falls 17%; applying the separately timed reductions to Table 5, PotencyGNN remains an estimated 6.8–7.1 times faster (8.22 against four programs), and break-even rises from about 21,700 to 25,600–27,000 queries (Section 5.4, Appendix G).

W2, second part. With the full distribution and one query, the full sharp bounds are the right output, and potency does not replace them. It answers a different question: how much of the identification comes from the constraints linking the auxiliary variables to treatment and outcome. In Appendix A, the treatment–outcome constraints give [0.340, 0.835]; adding those involving the instrument gives [0.590, 0.670]. One reports [0.590, 0.670]; potency, 0.495 − 0.080 = 0.415, says how much of that precision is owed to the instrument. When potency is zero, the base analysis gives the same width, which the full interval alone does not show; of 400 published-network queries, 144 have potency above 10⁻⁶ and 256 are near zero, at or below it (Section 5.6). Maringgele and Etesami (2026) score experiments before their outcomes are observed; potency is computed from data already collected. Both regimes hold the structural assumptions fixed, so potency does not measure sensitivity to them. New: the tightening relative to a base regime with free mechanisms splits exactly into the part from fixing the mechanisms outside the confounded set and potency (means 0.184 and 0.154 on 1,320 instances; Appendix G).

Question. Intuition: potency is how much of the identification is owed to the auxiliary variables; above, the instrument narrows the width from 0.495 to 0.080. Scenario: one query is cheap; the cost comes from repetition. Potency is computed from an estimated distribution and does not measure sampling uncertainty (Section 3); a bootstrap interval recomputes it once per resample, so with 1,000 resamples 22 queries exceed the 21,700 break-even. Screening candidate queries and comparing auxiliary sets multiply the counts further (Section 1); the published-network study alone enumerates 1,811 candidates (Appendix E). PotencyGNN's error is measured on held-out cores and empirical data. Challenge and improvement: exact cost grows with the number of response types. New: retraining raises held-out R² at 512, 1,024 and 2,048 response types from 0.795, 0.744, 0.652 to 0.872, 0.832, 0.859, where exact computation is 17.4–43.4 times slower; on stress-test graphs from other causal structures (overlapping cores left out), from 0.571, 0.742, −0.184 to 0.769, 0.800, 0.490, the last with a three-parent confounded variable unseen in training (Appendix H; not prespecified; added cores also differ structurally).

Thanks for the author's response. Sorry, I still don't understand the paper's main motivation. In particular, if the auxiliary variables are observed, why would one want to understand their contribution to identification, and why would that be time-sensitive (compared to the linear programming approach)?

Maybe there are applications that require solving partial identification as a subprocess, where the speed improvement could be important. But since the paper's main contribution is relatively narrow in those aspects, I think it should position itself around applications like that rather than as a general framework. That would make it clearer what the exact challenge is and how the proposed method fits. Otherwise, it would be hard to imagine an application like that and assume the proposed method would seamlessly improve efficiency without affecting other aspects.

Thank you for the follow-up and for the concrete suggestion. We agree that a specific application makes the challenge and the role of PotencyGNN clearer, and we have added it to Section 1, with the full experiment in Appendix E. It addresses both of your questions.

Application: deciding which variables to measure in a follow-up study. An analyst holds a causal network and a dataset and plans a replication or a study at a new site. For each treatment–outcome query, the analyst needs to know how much of the identification rests on the auxiliary variables, and potency measures this as the narrowing of the bounds they provide. In the instrumental-variable example of Appendix A, as when Z is randomized assignment in a trial with noncompliance, the treatment–outcome distribution bounds the effect to [0.340, 0.835]; adding Z narrows it to [0.590, 0.670]. Potency, 0.495 − 0.080 = 0.415, is 84% of the base width: the joint constraints of Z with X and Y remove most of it, which is evidence for measuring Z again, and with only the P(X,Y) constraints the bounds are six times wider. When potency is zero, the treatment–outcome constraints alone yield the same bounds, and the auxiliary variables do not narrow the bounds for that query. The full interval cannot reveal either fact; only the comparison of the two regimes can.

Why the contribution of observed variables matters. Observed auxiliary variables are not always useful, and how much they help differs sharply across queries. Of 400 queries on published networks (Section 5.6), the auxiliary variables narrow the bounds by more than 0.05 for 40, by a smaller positive amount for 104, and by at most 10⁻⁶ for the remaining 256. Adding observed variables never widens the bounds (Section 3), but for most queries it barely narrows them; only potency shows which case applies.

Why it is time-sensitive, and how the cost grows with the number of variables. The decision is made for every candidate query and auxiliary set in the network, and each candidate is recomputed under resampling to check that differences between candidates persist across resamples (Section 1). Without further restrictions, nn observed variables give n(n−1)n(n-1) ordered treatment–outcome pairs, which grows quadratically, and up to 2n−22^{n-2} candidate auxiliary sets per pair, which grows exponentially; with BB resamples, up to B n(n−1) 2n−2B\,n(n-1)\,2^{n-2} exact computations are needed. For n=5n=5 this gives at most 160 candidates; for n=10n=10, at most 23,040. The cost of each exact computation grows as well. Its linear programs have one unknown per joint response type, K=∏Vi∈C22kiK=\prod_{V_i\in C}2^{2^{k_i}} in total, where kik_i is the number of observed parents of the confounded variable ViV_i (Section 3), so KK grows doubly exponentially with these parent counts. As KK rises from 16 to 262,144, cached exact computation slows from 3.57 ms to 4,265.83 ms, while PotencyGNN stays near 1 ms (Section 5.5); within its training range (K≤256K\le256) its R2R^2 is 0.761–0.870. Even with subgraphs restricted to five variables, computing all bounds in the application below means bounding 263,867 resamples (1,351 published-network queries, 200 resamples each, less those violating Assumption S).

The screen keeps nearly all decisions while computing fewer bounds. We ran the application on the 1,351 supported published-network queries, with 200 resamples of size 1,000 per query drawn from the network models, and the rule "select if the median score exceeds 0.05". Where sampling variation made the full linear programs infeasible (47.5% of resamples satisfying Assumption S), we computed the full bounds over the closest distributions the model can produce; on those resamples the score is not the potency of Section 3. The resulting reference agrees with the original-distribution criterion, potency above 0.05, on 1,347 of the 1,350 queries it classifies. PotencyGNN then gates which queries have their bounds computed; the others are not selected. For the two training regimes, this screen selects 95 and 94 of the 98 reference selections, adds none, and agrees on 1,347 and 1,346 of the 1,350 decisions, while computing bounds on 45% and 41% of the 263,867 valid resamples; these count bounded resamples, not run time. Of the 96 reference selections whose original potency exceeds 0.05, it recovers 94; the two it misses have original potency 0.054 and 0.055 and reference medians 0.050 and 0.052, just above the threshold, and its other misses have original potency 0.020 and 0.025. This falls short of the criterion we fixed before the run: at most two disagreements and no missed selection with original potency above 0.05. PotencyGNN alone errs toward over-selection, and the linear programs remove the additional candidates it flags. Appendix E reports the full results and their limits, including that restricting the reference to the smaller projections lowers its selections from 98 to 86.

Reviewer_HgsuOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies partial recognition problems under potential confounding. The author proposes PotencyGNN, which uses a relational graph neural network that distinguishes between parent nodes, child nodes, and bidirectional edges for amortization calculation; Additionally, train a dual certificate network to provide a provable lower bound of effectiveness through dual feasibility repair. The experiment is based on a composite benchmark of 3 to 6 binary variables, evaluating generalization, inference speed, certificate validity, and migration to real datasets while leaving the causal core.

Critical strengths
  1. The design of dual certificates separates validity from prediction accuracy, and the validator rounds outward under dual accuracy. All 79200 evaluations are valid, which is a solid engineering aspect.
  2. The experimental design and reporting methods are highly self-critical, including datasets divided by causal core, coverage experiments under equal label budget, exposure state audits, and truthful reporting of external push failures, which are worthy of recognition.
Critical weaknesses
  1. The actual motivation for amortization is weak: within the scope covered by the paper, accurate calculation takes only about 7 milliseconds per time step, the value of 8x acceleration is limited, and the training cost requires more than 20000 queries to recoup. However, when accurately calculating the truly expensive large response space area, the model's predictions have become ineffective, and the method happens to be unavailable where it is most needed.
  2. The scope of application is very narrow, limited to binary variables, single bidirectional groups, each group variable can have up to two parent nodes, the response type does not exceed 256, and it needs to meet the positive support assumption of the parent configuration of the external group. The author did not provide a path for generalizing to multivalued variables or multiple confounding components.
  3. In baseline comparison, the optimized GRIT is better than PotencyGNN on the benchmark of 3-5 variables, and the difference on the benchmark of 6 variables is mostly insignificant. The ablation experiment also showed that the query unit features and global context have almost no effect, making it difficult to establish the contribution of the architecture itself.
  4. Certificate inference still requires solving two basic linear programming problems, which are of the same order of magnitude as precise calculations, and only about 43% of queries can obtain a positive lower bound. The input dimension of the certificate network is also fixed at the current scale, and its actual filtering value and scalability are unclear.
  5. The persuasiveness of real data evaluation is insufficient: all positive validity queries share the same instrumental variable structure, and there are closed-form formulas that can be accurately reproduced; The reference values queried by Sachs are all zero. The ROC-AUC with positive efficacy detected on published networks is only about 0.58 to 0.68, indicating weak screening ability.
  6. The basic interval also fixes the distribution of out-of-group conditions, not just the distribution of processing and results. The author only examined the impact of this setting in a single example, so it is still unclear to what extent the effectiveness reflects the information brought by the auxiliary variable.
Questions for authors
  1. In what practical scenario is it necessary to repeatedly calculate the effectiveness of the same small-scale problem tens of thousands of times to make amortization truly meaningful?
  2. What impact does baseline screening have on label distribution and reserved set R ², as it only retains at least one causal core with efficacy exceeding the threshold? What are the results on unfiltered data?
  3. If the basic interval does not fix the distribution of out-of-group conditions, but allows these distributions to change freely, how much difference will there be in the definition and value of effectiveness?
  4. Can the certificate network be extended to constraint rows and response types that were not seen during training? If not, how do you plan to handle it when expanding to a larger image?
  5. Can a verifiable upper bound on the validity be provided simultaneously, so that the certificate can be used not only to confirm tightening, but also to exclude invalid auxiliary variables?
2 comments

Thank you for the detailed review. Results marked "new" were obtained during the discussion and are in the revised PDF.

W1. Agreed: exact computation takes 6.776 ms per query on the primary benchmark. The revision states the aim as lower cost per query under repetition, not problems too large for exact computation (Section 1). New: retraining raises held-out R² at 512, 1,024 and 2,048 response types from 0.795, 0.744, 0.652 to 0.872, 0.832, 0.859, where exact computation is 17–43 times slower; on stress-test graphs from other causal structures (overlapping cores left out), from 0.571, 0.742, −0.184 to 0.769, 0.800, 0.490, the last with a three-parent confounded variable unseen in training (Appendix H; not prespecified; added cores also differ structurally). New: two programs suffice on 181 of 440 cores; the estimated speedup then falls from 8.22 to 6.8–7.1 times and break-even rises from about 21,700 to 25,600–27,000 queries (Appendix G).

W2. At most 256 response types and two parents are benchmark settings; retraining reaches 2,048. Binary variables, one confounded set and the support assumption define where exactness is stated; support is checkable from P(V) (Appendix E excludes the 156 of 1,811 candidates that fail it). New: Appendix G gives a path for each extension. Multivalued variables keep the bounds linear. Several independent confounded sets generally give polynomial programs (Duarte et al., 2024); merging them gives valid, not necessarily sharp, intervals and the potency of the relaxed model. Implementation and evaluation remain future work.

W3. Tuned GRIT is more accurate on 3–5 variables (0.843 vs 0.829); PotencyGNN is highest on six (0.779 vs 0.754; interval excludes zero on one split), with 67,841 parameters vs 121,097–670,733 and no tuning (Table 1). Removing edges lowers R² to 0.277, symmetrizing to 0.764. The query-pair profile and global context carry information other inputs lack (Section 4); removing either alone changes R² by −0.006 or +0.004, which shows neither is individually necessary, not that both can be dropped. Results use the prespecified model (Section 5.4).

W4. The certificate adds a guarantee, not a speedup: 4.195 vs 6.776 ms per query, excluding outward verification; the two base programs remain (Table 5). 43.4% of 79,200 held-out evaluations get a certificate above 10⁻⁵, at precision 1.0 (recall 0.444; 97.8% have potency above it); ranking by the bound recovers 0.487 of the top tenth (0.10 at random). A nonpositive value is inconclusive: it cannot show a set is uninformative. The bound holds for any network output, given the assumed graph, the support assumption, known P(V) and full-program feasibility; it has no statistical coverage. The 33 rows limit it to five variables (Appendix F).

W5. We agree with the three limits. Without empirical fitting, PotencyGNN has the lowest dataset-macro MAE of ten models (0.0125 vs 0.0205 for tuned GRIT); with the IV core held out (split 7) the comparison is mixed. On published networks ROC-AUC is 0.682/0.582 at the 10⁻⁶ cutoff but 0.934/0.933 for potency above 0.05 (Section 5.6, Appendix E).

W6, Q3. New: freeing the mechanisms outside the confounded set in the base regime adds to potency exactly the tightening from fixing them (1,320 instances: mean 0.338 = 0.184 from fixing + 0.154 potency). The two coincide when a directed treatment–outcome path lies in the confounded set (181 of 440 cores; sufficient, not necessary). Both are reported (Section 3, Appendix G).

Q1. Potency does not measure sampling uncertainty (Section 3). A bootstrap interval with 1,000 resamples needs 1,000 computations per query, so 22 queries exceed the 21,700 break-even. PotencyGNN's error is measured on held-out cores and empirical data; no evaluated study is large enough to recover the training cost (Section 5.4).

Q2. New: the 291 rejected cores have zero potency in all screening draws (3 of 17,460 fresh labels positive). Training on all cores vs screened ones gives held-out R² 0.901 vs 0.879 over all 731, and 0.849 vs 0.851 on retained cores (interval includes zero); MAE on rejected cores falls from 0.021 to 0.002 (Appendix D).

Q4. Validity needs no training: any output is repaired through the normalization row, for any number of rows and columns (Appendix F). No parameter depends on the number of response types. The trained network's rows are fixed; larger graphs need a row-flexible encoder, not yet implemented; tightness is unmeasured.

Q5. Yes in principle: potency is at most the base outer width, already computed in the same run, minus a verified lower bound on the full width. The latter needs exactly verified feasible full-program points, which the verifier does not provide (Section 5.7); solving the full programs for them removes the cost advantage. A bound below a tolerance would certify that an auxiliary set narrows the interval by at most that tolerance. Not implemented; future work.

Thank you for the detailed response. Although most of my concerns have been addressed, I still do not think this paper reaches the ICLR acceptance threshold. I will therefore keep my original score unchanged.

Comments on the paper

1 comment

General response to all reviewers

We thank the reviewers. Hgsu and d3p7 noted that the certificate separates validity from prediction accuracy, and d3p7 that the ablations support relation-specific message passing. HAwy's comments on presentation and on the closed-form base interval led to a clearer introduction and a two-program exact baseline. The shared concern is the case for amortization. We answer it here, and each response below gives the evidence and the scope for that reviewer's points. Results called "new" were obtained during the discussion period and are in the revised PDF.

The aim (all reviewers). The aim is a lower cost for each additional potency computation when many are needed. It is not to solve instances that are too large for exact computation. On the primary benchmark, four-program exact computation with a cached graph takes 6.776 ms per query and PotencyGNN 0.824 ms (8.22 times faster). The submitted abstract did not make this clear. The revised abstract states that exact computation takes milliseconds per query on the primary benchmark, and Section 1 states the aim.

Is the surrogate accurate where exact computation is slower? (Hgsu, d3p7) New: we retrained with response spaces of up to 2,048 response types. Mean held-out R² at 512, 1,024 and 2,048 response types rises from 0.795, 0.744 and 0.652 to 0.872, 0.832 and 0.859, and exact computation is 17.4, 32.2 and 43.4 times slower than PotencyGNN at these sizes. On the paper's stress-test graphs, which differ from the retraining graphs, R² computed from predictions averaged over five seeds rises from 0.571, 0.742 and −0.184 to 0.769, 0.800 and 0.490 after excluding training-overlapping cores. The added graphs all have five variables and differ from the original training graphs in structure as well as in response-space size, so this result extends the response-space range, not the graph size (Appendix H, Tables 13 and 14).

What else is new.

  1. Presentation (HAwy). The first paragraph of Section 1 now defines auxiliary variables, the base and full regimes, the confounded set and mechanisms, and states that potency complements the sharp bounds and does not replace them.
  2. Fewer linear programs (HAwy). When a directed treatment–outcome path lies inside the confounded set (181 of 440 cores), the base interval has a closed form and two programs suffice. Across the benchmark, median exact latency falls by 17%; the estimated speedup is then 6.8–7.1 times (Appendix G).
  3. What potency measures (Hgsu). Both regimes fix the mechanisms of the variables outside the confounded set. If the base regime does not fix them, the tightening is at least as large, and the two are linked exactly: on 1,320 instances the mean total of 0.338 splits into 0.184 from the mechanisms and 0.154 of potency (Section 3, Appendix G).
  4. Benchmark screen (Hgsu). In the 291 rejected cores potency is zero in every screening draw; in fresh draws, 3 of 17,460 labels are positive. Training on all cores lowers the error on rejected cores from MAE 0.021 to 0.002, and R² on the retained cores is 0.851 with screened-only training and 0.849 with all cores (Appendix D).
  5. Certificate (Hgsu). Recall is 0.444 at precision 1.0 over 79,200 evaluations: no positive certificate exceeds the numerical reference potency. A nonpositive certificate is inconclusive (Section 5.7).

The paper's contributions and their evidence (Section 1). (1) Definition. Realized observational potency, with the mechanisms outside the confounded set held fixed, and an exact account of what fixing them contributes (Section 3, Appendix G). (2) PotencyGNN. A surrogate that predicts potency on causal cores absent from training, with held-out R² of 0.829 on the three-to-five-variable benchmark and 0.779 on the six-variable one (Table 1), at 0.824 ms per query against 6.776 ms for exact computation. (3) Certificate. A separate network gives certified lower bounds on potency; in all 79,200 held-out evaluations the numerical validity criteria hold, and no positive certificate exceeds the numerical reference potency (Section 5.7). (4) Evaluation. The main accuracy results and the certificate evaluation are measured on causal cores absent from training, and at a fixed number of training labels, covering more distinct cores improves accuracy more than adding instances of the same few (Section 5.3).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.