acceptodds
Under review as a conference paper at ICLR 2027

Defeasible and Deontic Probabilistic Circuits: Semantic Topology for Belief Revision and Normative Audit

Abstract

A model that learns from what people do can predict what they will do next, but it cannot tell whether what they do is allowed or not. Nor does it have a principled way to retract a conclusion when an exception arrives without disturbing everything else it believes. To address this, we introduce Defeasible Probabilistic Circuits (DPCs) for structured belief revision and Deontic Probabilistic Circuits (DePCs) for normative auditing. These models designate selected gated sum nodes of a conditional probabilistic circuit as defeasible or deontic (audited) nodes, preserving exact, linear-time inference. To support defeasible reasoning—where previous conclusions can be overridden by new evidence—a DPC structures a node so one branch is the default and the rest are prioritized exceptions. It compiles any learned softmax gate exactly into this ordered-rule form without retraining. A new exception can then be inserted without altering higher-priority weights, accompanied by an evidence-specific certificate bounding its effect. To support deontic reasoning—evaluating what is allowed versus forbidden—a DePC designates branches as either compliant or violating. It pairs the learned descriptive distribution with a policy-supplied normative reference, exactly computing severity-weighted departures and using a gate-wise audit budget to localize policy violations. Across two physical-reasoning and two clinical environments, DPCs and DePCs improve update and action ranking, respectively, over comparably sized conditional circuits. Although trained solely from observations and without ranking supervision, they also outperform GPT-5.6 Sol, Qwen3.8-27B, and Qwen3-32B under zero-shot evaluation.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Probabilistic Circuits, Defeasible Reasoning, Deontic Reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YbxxOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DPCs and DePCs with the goal of modeling defeasible and deontic (probabilistic) logic within a probabilistic circuit framework.

Critical strengths

The introduction of defeasibility and deontic logic to the PC literature is a needed one. There has been little work on belief revision in circuits. This work makes a step in this direction.

Critical weaknesses

The theoretical results of the paper are hard to follow. I believe that most (if not all of it) is AI generated, reflected by the confusing writing at times. There is an excessive use of new variables. It is best to show a few important results in a clear human-readable way (and defer the rest to the appendix) than to instead cram all results in the paper and very briefly and hurriedly speed through them.

My main concern is with the description of how to construct the defeasible and deontic gates. The paper teases an almost-answer to this in Section 4.2, but just like in the defeasible case, it does not concretely show an algorithm to identify the gates to edit, from what I understand.

Questions for authors
  1. Line 282-285: it is unclear what the exact statement of intractability is being made here. What exactly (in terms of probabilistic or logic queries) do the authors mean by auditing? Could one not have an sd-DNNF representing the policy and (assuming a compatible PC representing the learned distribution and deontic/defeasible logic) the model count of the disjunction of the two representing a measure of compliance to the policy?
  2. There is a lack of description on how GPT and Qwen were prompted and how questions were queried. How were autoregressive models evaluated?
  3. There is no discussion on how one selects which sum nodes to turn into defeasible or deontic gates (and why that is compliant with the policy or how this successfully represents an exception). This is arguably the most important contribution, yet there is no discussion of it. Could the authors explain how this is done concretely?

Minor issues (no need to answer during rebuttal):

  1. Typesetting of DePCs. Sometimes as DePCs and sometimes as \textsc{DePC}s.
  2. It takes a while to get used to the naming. I would suggest DefPCs and DeoPCs instead. This way the reader immediately understands which framework the term refers to.
  3. Line 250: Notation Sc\mathbf{S}^c is undefined.
  4. Theorem 8 states a certificate, however the statement never says what this certificate is. One has to infer that βs(e)\beta_s(\mathbf{e}) is the certificate. It would be better to state it clearly.
  5. Line 258: What is KK? It is left undefined. Is it KsK_s when ss is the root node?
  6. There is very little elaboration on the intuition behind the theoretical results. How is one supposed to interpret the (very many) different variables defined and the equations that are derived henceforth? The statement of results are very technical and without a proper explanation of how they link to the overarching story. This makes the math hard to follow. There are too many variables being defined. It is better to repeat one-self rather than define many variables and rely on the memory of the reader.
  7. Line 372: What is CSPN? Missing citation to Shao et al. 2019?
  8. Line 374: What is Vanilla PC? Missing citation or just the same structure as CSPN but where weights are not conditional?
  9. Are the datasets evaluated tabular? In which case it is expected for PCs trained on these datasets to perform better than LLMs, which were only prompted with no fine-tuning.
3 comments

Thank you for the careful review, and for the point that belief revision in circuits has received little attention. We answer every point below, starting with your main concern. All changes in the revised PDF are marked in blue.

W2 and Q3: how the defeasible and deontic gates are constructed and selected.

There are three separate questions here, and we answer each one.

(a) Which sum nodes are gates?

Gate placement is fixed when the circuit is built. We build four balanced region trees over the 32 variables and place an update-dependent sum node at every internal region (31 per tree). Every one of these nodes is a designated node, so no selection step is needed. Appendix E.2 of the revision describes this.

(b) Which branch has which role, and why does it represent a default or an exception?

The first child of every gate is the designated branch: the default in a DPC and the compliant branch in a DePC. During training, its logit starts with a positive bias, and a regularizer encourages the gate to route to it under the neutral input. Thus, the designated branch carries the behavior under the neutral input, and the other branches receive mass when an update moves the gate away from it. This is the sense in which they are exceptions. No label tells the model which observations are exceptions or violations. Note that the role itself is a specification and is not learned: Proposition 4 shows that which behavior is compliant cannot be recovered from behavior alone. In a DePC, the normative reading routes every gate to its compliant branch, and the descriptive reading uses the learned weights.

(c) Which gate should be edited when a new rule arrives?

The results in the paper give a selection rule, which we now state explicitly. Suppose the new rule concerns a set of variables V.

The candidate gates are the designated gates whose scope contains V. In a region tree, these are the smallest region that contains V and its ancestors. By Corollary 7, an edit at a gate with scope S cannot change the distribution of the variables outside S. Thus, the candidate with the smallest scope leaves the largest set of variables provably untouched, and it is the default choice. Before any edit is applied, one upward and one backward pass yield the certificate of Theorem 8 for every candidate gate at once (Appendix B). The user can therefore compare the candidates on the contexts that must be protected, and choose the gate with the smallest certificate. Theorem 6 then inserts the rule at the chosen priority position. The new branch takes mass only from lower-priority rules. With binary applicabilities, the gate selects the highest-priority applicable exception, which is the behavior that prioritized default reasoning asks for. For an audit, the question is which gates to inspect, and Theorem 12 answers it. Each term of the audit budget belongs to one gate, and the gates with the largest terms are where behavior departs from the policy under the given evidence.

Q1: lines 282-285 and the sd-DNNF construction.

Thank you for this suggestion. By auditing we mean two queries: how much the expected severity under the learned model exceeds the expected severity under the policy (Lemma 9), and which gates account for the difference (Theorem 12).

When a policy is a logical constraint on X, compiled into an sd-DNNF that is compatible with the PC, the probability that the constraint holds is tractable, and it is a useful measure of compliance. Our setting differs in two ways. First, a DePC labels branches of sum nodes and not assignments to X. A compliant branch and a violating branch can produce the same outcome. Thus, compliance is an event on traces, and it cannot be written as a constraint over X unless the latent routing is made explicit. Second, the audit gap measures how much worse the outcomes are, and not only whether the policy holds.

The statement at lines 282-285 was too strong. We meant that exact distances and divergences between two circuits are tractable only for particular measures and under structural conditions such as determinism (Vergari et al., 2021), which we do not assume. We did not mean that policy-satisfaction queries are intractable. The revision rewrites the sentence and adds your construction as a remark.

Q2: how the LLMs were prompted and evaluated.

Each model receives the observed context and the eight candidate descriptions, as in the examples of Appendix G. In the physical task, it is asked to rank the updates by their effect on the hypothesis. In the clinical task, it is asked to rank the actions by their deviation from the applicable protocol. The model returns a complete ordering of the eight candidates in JSON, and we score it with the same metrics as the circuits. No gold scores or rankings appear in the prompt. Appendix E.2 gives the decoding settings.

W1: clarity of the theory.

We agree that the theory was harder to follow than it should be. As our AI-use statement says, we used AI tools to help draft the text and to help develop and check the proofs. We reviewed every proof ourselves, and we take responsibility for them. In the revision, the main text points to the proofs and to each appendix, the symbols you could not find (S^c and K) are defined where they first appear, and the explanation after each main result is rewritten in plainer language. In plain terms, the main results say the following.

  • Proposition 5. Any trained gate can be rewritten as an ordered list of rules: use exception 1 if it applies, otherwise exception 2, and so on, and otherwise the default. The applicability a_i is the probability that exception i fires given that no higher-priority exception did. Nothing about the model changes.
  • Theorem 6. A new exception inserted at position t takes its mass, delta, only from the rules below position t.
  • Theorem 8. Under evidence e, write r_e for the probability mass of e that the edit leaves untouched, v_e for the mass it removes, and u_e for the mass it adds. Then no conditional prediction can move by more than beta_s(e), the larger of v_e/(r_e+v_e) and u_e/(r_e+u_e). In words, a prediction can move only as much as the edit changes the mass consistent with the evidence.
  • Theorem 12. The divergence between behavior under the policy and behavior under the model is at most a sum over gates. Each term is the probability that the gate is reached under the policy, times how differently the policy and the model route at that gate.

Minor issues.

  1. Typesetting of DePC. Fixed. DPC and DePC are now typeset the same way everywhere.
  2. Naming (DefPC, DeoPC). We see the point. The abstract on OpenReview already uses DPC and DePC, so we kept these names during the discussion period to keep the record consistent.
  3. Line 250. S^c is the set of variables outside the scope S of the edited node, that is, X \ S. It is now defined before it is used.
  4. Theorem 8. The certificate is beta_s(e). The theorem now says so.
  5. Line 258. Your reading is correct. K is the part of the root distribution that the edit leaves unchanged, K = A_s K_s + B_s, where A_s and B_s come from the affine form at line 244. When s is the root, A_s = 1 and B_s = 0, and thus K = K_s. The definition is now in the text.
  6. Intuition and number of variables. Please see our answer to W1. Each main result is now followed by a plain explanation.
  7. Line 372. CSPN is the conditional sum-product network of Shao et al. (2020). The citation is now given at that line.
  8. Line 374. It is the latter. The Vanilla PC has the same structure with sum weights that do not depend on the update. It models the joint distribution over (X, U) and is conditioned on u at query time.
  9. Tabular data. Yes, all variables are categorical. We agree that a zero-shot LLM is not a like-for-like competitor for a model trained on the environment, and we present the LLMs as reference points.

Thank you for responding to my concerns. Having read through the author response, I will maintain my score.

My main concerns were

  1. The presentation and writing, which reads awkwardly and sometimes confusingly. It appears to be heavily AI generated. I believe the revised PDF addresses some of the clarity concerns, but the presentation and writing continue to be confusing and clearly AI generated. There is hardly any intuition behind the explanation, and the text overflows with unnecessary technical jargon that is common in AI generated text. The rebuttal is also seemingly heavily AI generated. I recommend a thorough rewrite of the paper with human words.
  2. Clarity on the algorithms. I asked for a clear algorithm on the construction, learning and belief revision. No algorithm is given in the revised PDF. The rebuttal answer to this question is confusing and I believe completely AI generated.
  3. How the autoregressive models were queried. The rebuttal does not address this concern.

I believe none of these concerns have been sufficiently addressed.

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

The paper extends conditional probabilistic circuits with semantic branch roles at selected sum nodes. In the Defeasible construction one child is a default and the others are exceptions carrying a supplied priority order, and new exceptions can be inserted without rescaling existing branches. In the Deontic construction children are labelled compliant or violating, and the same circuit is read twice, once with learned gates for the descriptive distribution and once against a policy-supplied reference with violating branches masked and renormalized. The theoretical results include an exact, closed-form compilation of any learned softmax gate into an ordered-rule form with identical predictions, a priority-preserving insertion theorem with a sharp conditional total-variation bound, an exact one-pass audit gap in expectation, and a gate-wise bound relating that gap to posterior-weighted local KL terms. Evaluation uses four synthetic environments covering physical update ranking and clinical action auditing, against CSPN, a vanilla PC, and three LLMs, plus two independent simulators for generalization.

Critical strengths

The theory is the strongest part of the submission and is independently useful. Exact compilation of a learned softmax gate into stick-breaking-like ordered rules in closed form, with a proof that predictions are preserved, is a clean result that a practitioner could apply directly. The insertion theorem with its sharp conditional bound is better than the usual soft-update alternative, which rescales every existing branch, and the gate-wise audit bound is an elegant use of the latent-trace view of circuits, with Proposition 13 correctly identifying what the joint trace-and-outcome divergence measures relative to the outcome-only quantity. The generalization benchmarks are a real attempt to escape the self-generated-train-and-test pattern, and using a stochastic Bayesian network without circuit structure for one of them is the right control.

Critical weaknesses

The empirical case is much weaker than the theory and rests entirely on synthetic data with a single training seed. The gain over CSPN on the clinical main result is a strict Kendall tau of 0.0112 with a bootstrap lower bound near 0.002, and exact-ranking accuracy remains at 3.4 percent against 2.6 percent, so the eight-way ranking task is effectively unsolved on both sides. DPC loses to CSPN on top-1 accuracy in the physical main table, 55.83 against 56.85, and DePC loses on the clinical hard stratum, 57.74 against 58.18; neither loss appears in the main text. The reported intervals resample queries within a single fitted model, so they say nothing about whether the ordering would survive retraining, and no other PC family is compared against, which leaves the lineage thin. Second, there is no architectural ablation. The entire claim is that branch roles, priority order, and the normative second reading cause the improvement, yet the only comparison is assembled-model against assembled-model. There is no DPC without role labels, no DPC with shuffled priorities, no DePC without the normative read, and no sensitivity to the number of semantic branches or to the severity weights in Equation 15, which the authors chose. The LLM baselines are set up so that a small trained circuit wins by construction: zero-shot, no chain-of-thought, no few-shot exemplars, a no-think mode for one model, and a short token budget, on a domain the LLMs were never trained for.

Questions for authors

Could the authors report multiple training seeds and give run-level rather than query-level intervals? What is the result with the branch-role labels removed, or with the priority order shuffled, and what is the sensitivity to the number of semantic branches and to the severity weights? DPC loses to CSPN on top-1 accuracy in the physical setting and DePC loses on the clinical hard stratum. Is there a regime where DPC or DePC wins on top-1, and does the priority mechanism hurt the argmax while helping the full ranking? For the LLM comparison, what happens with chain-of-thought prompting and few-shot exemplars, and would fine-tuning the LLMs on the same observations close the gap? How do the results compare to other tractable probabilistic model families such as RAT-SPNs, HCLT circuits, or deep probabilistic circuits? Finally, given that priorities, compliance policies, and severity scales are all external inputs, how much of the DePC advantage over CSPN comes from the semantic labelling itself rather than from the additional side information?

1 comment

Thank you for the detailed review. Your questions repeat the weaknesses, so we answer both together.

  1. Seeds and run-level intervals.

Over five additional seeds, DPC improves strict tau over CSPN by +0.1249 [0.0825, 0.1674] and is ahead in all five runs. DePC reaches 0.6721 against 0.6481, a paired gain of +0.0239 [0.0136, 0.0342], with +7.44 points in top-1 and +1.06 points in exact ranking. The revision reports this in the Appendix.

  1. The losses, and top-1.

The physical top-1 loss was in the submitted main text (L416-417). The clinical hard-stratum loss was only in Appendix F, and we moved it into Section 5.2. The physical top-1 deficit does not persist across seeds: DPC improves top-1 by 6.20 points on average. DePC is also ahead on top-1 in the original clinical run (50.09% versus 44.07%) and DPC on the physical generalization benchmark (79.54% versus 76.94%). So we see no trade-off between the argmax and the full ranking. The priority mechanism cannot hurt the argmax, because by Proposition 5 it changes no prediction. We agree that exact ranking is unsolved for every model: with 8! orderings and many nearly tied gold scores it is a very strict metric, and tie-aware tau and pairwise accuracy are more informative.

  1. Branch roles removed.

The role structure enters training in one way. The first child of every gate is the designated branch. Its logit starts with a positive bias, and a regularizer encourages the gate to route to it under the neutral input. A DPC of the same size without this structure reaches a strict tau of 0.4927, against 0.5927 for DPC, a paired difference of +0.100 [0.064, 0.136] over five seeds. The submission did not describe this mechanism clearly; Section 5.1 and Appendix E.2 now do.

  1. Priority order shuffled.

By Proposition 5, any ordering of the exceptions compiles to exactly the same weights, so shuffling the order changes no prediction and no ranking. The trained models also impose no order among the exception branches. The order matters only for later edits (Theorem 6), which the ranking benchmarks do not exercise.

  1. No normative read.

As Section 4.3 states (L351-355), a reference shared across candidates cancels from the ranking score. We confirmed this on six checkpoints: the rankings are identical with and without the normative reading. Its role is to give a single action's score a meaning relative to the policy, not to change the ordering.

  1. Number of branches and severity weights.

With B = 2, 4, and 8 semantic branches, DPC improves strict tau over the corresponding symmetric control by +0.143, +0.130, and +0.100 over five seeds. For DePC at B = 8, top-1 accuracy improves by 2.0 points and exact-ranking accuracy by 1.2 points over the symmetric control. By Proposition 10, affine changes to the severity scale change no ranking. For other reweightings we evaluated seven configurations of Equation 15, and DePC keeps its exact-ranking advantage over CSPN in all of them.

  1. LLM baselines.

We agree that they are reference points and not trained competitors. One clarification: GPT-5.6 Sol ran with medium reasoning effort and a 4,096-token budget (L1125), so it did reason before answering. Only Qwen3.8-27B ran in no-think mode with 256 tokens. We did not run few-shot prompting, chain-of-thought prompting, or fine-tuning.

  1. Other PC families.

Thank you for this suggestion. The generative PC we ran is the Vanilla PC baseline in Tables 1 and 2. It belongs to the family of generative PCs, which includes RAT-SPNs: it is a circuit with ordinary sum weights, trained on the joint over (X, U). CSPN is its conditional counterpart, and DPC and DePC add branch roles. On the main benchmarks the Vanilla PC reaches a strict tau of 0.3836 (physical) and 0.5050 (clinical), against 0.5953 for DPC and 0.6757 for DePC; parameter counts are in Table 1. We did not run HCLT. The ablation in point 3 differs from DPC only in the role structure.

  1. Semantic labelling versus side information.

DePC and CSPN are trained on the same observations and receive the same action inputs. Both are scored with the same severity utility (Appendix E.3). No label tells DePC which observations are violations. The designated branch is fixed by construction, and the regularizer uses only the identity of the neutral action, which CSPN also receives as an input. The priority order changes no probability. So the ranking gain does not come from extra information. The ablation in point 3 indicates that it comes from the structure.

Reviewer_UpfROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

The authors introduce two new types of probabilistic circuits (PC): defeasible PCs, containing nodes that imposes a priority order on child nodes, and deontic PCs, containing nodes that partition child nodes into compliant vs violating (a binary indicator). They show that these nodes allow for rules to be added to a trained model in a tractable manner (without retraining). They evaluate these circuits on two PC benchmarks and show that their model is aligned more closely with the ground-truth rankings of candidate actions.

Critical strengths
  1. Most of the theoretical part of the paper (Sections 2-3) seems to be quite thoroughly developed and explains: (i) the additional functionalities that these new circuits introduce (that a standard PC does not possess) as well as (ii) how they can be updated tractably after inference from data.
Critical weaknesses
  1. My main criticism lies with the limited evaluation and unconvincing comparisons/baselines:
  • Both datasets are (I believe) entirely synthetically generated. It is not clear to me how these datasets were generated and what the distribution over the training observations is. It is also unclear how "easy" the datasets are, which leads me to my next point:
  • The comparison is quite limited. Two of the methods (CSPN, Vanilla PC) are simply standard PCs without the functionalities introduced by the authors and the remaining comparisons are simply zero-shot LLMs. It should not be surprising that the authors achieve better results on synthetic datasets where the problem is designed, in some sense, to fit the solution.
  • Even with this major caveat, DePC is almost identical to (only barely better than) CSPN on the clinical dataset.
  • The main issue with PCs in general is that they scale far worse than other methods. If the authors were to compare to zero-shot LLMs, then perhaps comparison should also extend to, e.g. training an LLM or just a deep neural network from scratch on the same data.
  1. While a relatively minor criticism compared to the first point, I found most of the paper to be unnecessarily pedantic and the language to be needlessly obfuscatory. The language used is also a bit strange in parts. E.g. L397-403, L477-485, and just various sentences here and there throughout the whole paper. I would recommend editing the entire manuscript (manually) for clarity.
Questions for authors
  1. Can you explain how the two datasets were generated?

  2. Are there comparisons to trained LLMs/neural networks, or any other non-PC baselines?

You can generally address any of the sub-points I made in Point 1 of the Weaknesses section above.

2 comments

Thank you for the review. We answer each point below. The general response lists the new experiments.

W1.1 and Q1: how the datasets were generated, and how hard they are.

Both main benchmarks are synthetic, for a reason. To score a ranking of updates or actions, we need the true conditional distribution under every candidate, and only a known generator provides it. We are not aware of an existing benchmark for this task (general response, point 3). The physical benchmark is sampled from a frozen ground-truth DPC that defines P_GT(X | U). It has 32 categorical variables and 109 update identities, and the neutral update makes up about 30% of the samples. The clinical benchmark is sampled from a frozen ground-truth DePC with 32 categorical variables and 31 action configurations, which are covered approximately uniformly. Each domain has 100K training and 10K validation observations. The models see only (X, U) pairs. They never see gold rankings or policy labels.

Each ranking set has 1,080 queries with eight candidates and context sizes of 16, 20, and 24. Half of the queries are sampled independently. The other half form 270 matched pairs that keep the same candidates and change 3-6 context values, so a model cannot do well by learning which candidates tend to rank high. Exact-ranking accuracy is below 5% for every model, and on the hardest physical pairs every model is between 47% and 61% pairwise accuracy (Table 4).

You are right that generators of this kind could favor our models. This is why we also built two benchmarks from independent simulators that use no circuit and no branch routing; the physical one is a hand-specified Bayesian network. All circuit models are retrained on them. DPC reaches a strict tau of 0.7272 against 0.6806 for CSPN, and DePC reaches 0.7792 against 0.7685. The revision states all of this in Section 5.1 and Appendix E.1.

W1.2, W1.4, and Q2: are the baselines sufficient.

A baseline for this problem must do what the problem requires: answer conditional queries exactly, for any hypothesis and any set of observed variables, from one trained model. Only tractable probabilistic models can do this. We therefore compare with tractable models, as is standard in this literature. Our baselines isolate each ingredient of our models in turn. The Vanilla PC is a circuit whose weights do not depend on the update (one member of the family of generative PCs; the family includes RAT-SPNs, LearnSPN, PSDDs, cutset networks, and ACs). CSPN adds dependence on the update, with the same inputs and a comparable size, but no branch roles. The ablation is our own model with the roles removed. DPC and DePC add the roles. Each step is tested: across five additional seeds, DPC and DePC beat CSPN in every run, and the ablation shows that the gain comes from the default-branch structure (general response, points 1 and 2).

A neural network trained from scratch is not a baseline for this problem. It is a discriminative predictor for one form of query. It has no exact marginals and no branches, so it cannot produce the normative reading, the audit gap, or the certificate of Theorem 8, and no existing method computes these queries. For this reason we do not report such a baseline, and we make no claim about ranking accuracy relative to unconstrained predictors. The zero-shot LLMs are in the paper as reference points for general-purpose reasoning, not as trained competitors.

On scaling, our models have fewer than one million parameters. What the circuit provides is exactness. Theorems 6, 8, and 12 rely on exact marginals and circuit flows. They give a certificate on how far a new rule can move any prediction, and an exact severity gap against a policy. A neural network can be trained to predict the same quantities, but it cannot certify them.

W1.3: DePC is only barely better than CSPN.

We agree that the gain in the submission was modest (+0.0112 in strict tau). Over five additional seeds DePC reaches a mean strict tau of 0.6721 against 0.6481 for CSPN, a paired gain of +0.0239 [0.0136, 0.0342], and it is ahead in all five runs. Top-1 accuracy improves by 7.44 points and exact-ranking accuracy by 1.06 points. On the hardest clinical pairs the two models remain comparable, and CSPN is slightly ahead in the original run; the revision now says so in the main text. We would add that ranking is not the main purpose of a DePC. Its purpose is the audit of Section 4, which CSPN cannot compute because its branches have no compliance role. The ranking results show that this structure does not cost predictive quality.

W2: writing.

We agree. We rewrote L397-403 and L477-485 in plainer language. You also noted similar sentences throughout the paper, so we went through Sections 1 to 4 and rewrote about 25 more of them. We also defined notation that was used before it was introduced, and added pointers from the main text to the proofs and appendices.

Reviewer_UpfROpenReview ↗

Response to Authors

I thank the authors for their response.

I think the authors' reply has not only not addressed my concern, it has exacerbated it. The introduction of DPCs and DePCs introduce new functionalities to a probabilistic model. But as they've mentioned, none of the baselines can do these same things (auditing, normative reading, etc., which, by the way, should be defined properly). CSPNs can't do that and neither can neural nets, because they simply do not have the functionality to do so (no branches, etc.)

Furthermore, the authors mention in their rebuttal that the ranking (prediction) is not what's important, it's these additional functionalities like auditing. Putting this together, the authors have essentially introduced a new problem with no existing baselines, solutions, or even a way to evaluate quantitatively.

Given all this, my natural reaction is that I should first be convinced this is an important problem to solve. The scenarios provided in the introduction are merely suggestive and do not actually describe concrete modeling problems. So I turn to the evaluation, which is why I am asking about the datasets. My concern here isn't the synthetic nature of the datasets per se, it is that: (a) there are no existing benchmarks (created by someone else other than the authors) to be found at all, and (b) evaluating only on two contrived datasets with an opaque data-generating process indicates that there may not be a meaningful problem here that is worth solving. It is possibly not difficult at all to generate two datasets that allow for DePCs/DPCs to achieve comparable predictive performance to CSPNs.

Furthermore, if functionalities like auditing are what's truly important, is there at least some way (even proxy or contrived) they can be numerically evaluated?

The combination of all of these reasons above is why I am skeptical about the paper's motivation and evaluation in general. If the authors are introducing an entirely new modeling problem with no existing baselines, I would suggest that they do some combination of the following: (a) provide stronger qualitative explanations or empirical evidence for why this is an important unsolved problem, (b) find a way to quantitatively evaluate what really matters, or (c) create synthetic datasets from real data (e.g. the MIMIC database is a good way to make healthcare-related datasets, which is one of the motivating domains the authors provide in the introduction).

P.S. while I appreciate the efforts made in the paper revision, I still think that the paper is, overall, overly dense. For such a theorem-heavy paper, it is far more important to define terms unambiguously, cut down on jargon, and present the intuition clearly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for their careful reading. We ran new experiments and uploaded a revised PDF in which every change is shown in blue. The main text is now ten pages, as the rebuttal phase allows. This note summarizes what changed. Our replies to each reviewer give the details.

  1. Training seeds.

We retrained DPC, DePC, and CSPN on the main benchmarks with five additional seeds, and we report paired run-level 95% intervals. DPC improves strict Kendall tau over CSPN by 0.1249 [0.0825, 0.1674]. DePC improves it by 0.0239 [0.0136, 0.0342], with means of 0.6721 and 0.6481. Each model is ahead of CSPN in all five runs. Top-1 accuracy improves by 6.20 points for DPC and by 7.44 points for DePC.

  1. What drives the gain.

In our implementation, the first child of every gate is the designated default branch (DPC) or compliant branch (DePC). Its logit starts with a positive bias, and a regularizer encourages the gate to route to it under the neutral input. Nothing else supervises the roles. The submission did not describe this clearly, and the revision now does (Section 5.1 and Appendix E.2). To test whether this structure is the source of the gain, we trained a DPC of the same size without it. Over five seeds its strict tau is 0.4927, against 0.5927 for DPC, a paired difference of 0.100 [0.064, 0.136].

  1. Why the data are synthetic, and where they come from.

To our knowledge, no prior work audits a probabilistic circuit against a supplied behavioral policy, or revises one under supplied priorities, with exact guarantees. The closest work certifies fairness of probabilistic circuits (Selvam et al., 2023), which asks a different question. So no benchmark existed for our task. A benchmark also has to be synthetic for a second reason: to score a ranking of updates or actions, we need the true conditional distribution under every candidate, and only a known generator provides it. The standard PC benchmarks are density-estimation datasets with no updates, actions, or norms. We therefore built four environments. The two main benchmarks are sampled from frozen ground-truth circuits: a DPC for the physical domain and a DePC for the clinical domain. The revision says this plainly (Section 5.1 and Appendix E.1). Because of this, we regard the two generalization benchmarks as the stronger test. They come from independent simulators that use no circuit and no branch routing, and every circuit model is retrained on them. There, DPC improves strict tau over CSPN from 0.6806 to 0.7272, and DePC from 0.7685 to 0.7792.

  1. Presentation.

The main text now points to the proofs and to each appendix. The notation at L250 and L258 is defined, Theorem 8 precisely describes the certificate, and the tractability statement at L282-285 is made precise as well. L397-403 and L477-485 are rewritten in plainer language, and so are about 25 other sentences in Sections 1 to 4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.