acceptodds
Under review as a conference paper at ICLR 2027

A Knowledge-Informed Pretrained Model for Causal Discovery

Abstract

Causal discovery has been widely studied, yet many existing methods rely on restrictive assumptions and occupy one of two extremes: they either rely on strong auxiliary supervision, such as costly interventional signals or partial ground-truth, or adopt purely data-driven paradigms without guidance, which hinders practical deployment. Motivated by real-world scenarios where coarse domain knowledge is available, we propose a knowledge-informed pretrained model for causal discovery that integrates weak prior knowledge as a middle ground between the two extremes. Our model adopts a dual-source encoder-decoder architecture to jointly process observational data and structural knowledge while preserving their distinct semantics. We design a diverse pretraining dataset and a curriculum learning strategy that smoothly adapts the model in a knowledge-informed way to varying prior strengths across mechanisms, graph densities, and variable scales. Extensive experiments on in-distribution, out-of-distribution, and real-world datasets demonstrate consistent improvements over existing baselines, while exhibiting robust performance across diverse settings and promising practical applicability.

Primary area:
causal reasoning
Keywords:
causal discovery

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_fXNtOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Kode is a knowledge-informed pretrained model for causal discovery. It encodes plausible, impossible, and unknown relations in a structural-prior matrix and combines this knowledge with observational data through a dual-source encoder and graph decoder. Curriculum learning and ground-truth anchors are used to improve the use of structural knowledge.

Critical strengths

1.Supporting incomplete structural knowledge through a soft input is practically useful. The dual-source architecture gives a coherent way to combine knowledge and observational data.

2.The study includes synthetic mechanisms, different graph sizes, industrial datasets, and ablations of the training strategy. Appendix D.4 also includes prior-aware PC, GES, and NOTEARS controls.

Critical weaknesses

1.The comparisons do not cleanly isolate knowledge integration. Kode and the pretrained baselines use different training distributions and graph sizes. The prior-aware controls in Appendix D.4 help, but cover one synthetic setting and use different forms of knowledge. The gains cannot yet be attributed cleanly to the proposed knowledge interface.

2.The meaning of the structural prior needs clarification. Section 3.3 defines entries through possible direct or indirect influence, while ground-truth anchors use adjacency information. An absent direct edge is not an absent causal path. How these meanings are distinguished affects what the model is actually learning from the prior.

3.Robustness to imperfect knowledge is not sufficiently established. Masked ground-truth priors and fully random priors are useful controls, but leave a gap between correct knowledge and realistic mixtures of correct, incorrect, and contradictory expert statements.

Questions for authors

1.What remains of the gain when a matched pretrained backbone receives the same training distribution, and conventional baselines receive the same usable forbidden relations?

2.How are adjacency-based anchors distinguished from reachability-based priors, and how does performance change under partial corruption of otherwise informative knowledge?

6 comments

We address the weaknesses identified by the reviewers and respond to their questions below.
1.The comparisons do not cleanly isolate knowledge integration. Kode and the pretrained baselines use different training distributions and graph sizes.  The prior-aware controls in Appendix D.4 help, but cover one synthetic setting and use different forms of knowledge. The gains cannot yet be attributed cleanly to the proposed knowledge interface.
We first clarify that Kode and all baselines, whether pretrained or not, were evaluated on data from the same distributions and at the same graph sizes. For methods that require pretraining, the training data were also drawn from the same distributions and covered the same graph sizes. Kode additionally receives prior knowledge as input. We therefore attribute its overall advantage to the combination of coarse-grained prior knowledge, the architecture, and the training strategy designed to use it. We have revised the wording in the appendix that gave rise to this misunderstanding.

We agree that Appendix D.4 alone does not isolate the contribution of the proposed knowledge interface. The “0% knowledge” results in Figures 4, 6, and 13--16, Table 1, and Appendix D.4 show that Kode retains an advantage even when no prior knowledge is supplied. These results indicate that the architecture and training strategy also contribute to its performance.

The baselines in Appendix D.4 can use only fine-grained knowledge of direct causal edges, which limits their applicability when such knowledge is unavailable. The purpose of this comparison is to assess whether Kode can also use precise prior knowledge of direct edges. The results show an advantage for Kode, with an important qualification. When the entire ground-truth graph is supplied as hard constraints, those constraints fully determine the output of PC, NOTEARS, and GES, so their advantage in this setting is expected. Kode takes a knowledge-informed approach and uses both data and prior knowledge to predict the graph. Its predictions are not required to reproduce the prior, so providing the complete ground-truth graph does not guarantee exact recovery.

Our central contribution is the use of coarse-grained prior knowledge together with an architecture and training strategy designed to exploit it. In the revised manuscript, we make clear that the reported gains reflect their combined contribution. The current experiments do not establish an independent advantage of the knowledge interface alone.

2.The meaning of the structural prior needs clarification. Section 3.3 defines entries through possible direct or indirect influence, while ground-truth anchors use adjacency information. An absent direct edge is not an absent causal path. How these meanings are distinguished affects what the model is actually learning from the prior.
We thank the reviewer for suggesting this clarification. First, we use a common encoding for both types of prior. We assign 11 to a direct edge in an anchor adjacency matrix or to a direct or indirect relation in a reachability matrix. In either case, we assign 00 to the entry for the reverse direction. We also assign 00 to absent relations and diagonal entries, and −1-1 to all remaining entries whose status is unknown. Thus, 00 means “no direct edge” in an adjacency prior and “no directed path” in a reachability prior.

Second, adjacency and reachability have different meanings. Kode does not use a separate type label to distinguish the two priors. It treats both as relational information and processes them jointly with the observational data. Since reachability can be derived from adjacency, the data and supervision provided during large-scale pretraining allow Kode to learn from both sources of information. At inference time, this helps the model distinguish, to some extent, between direct connections and indirect effects and predict the causal structure more accurately. As a knowledge-informed model, Kode combines the prior with the data when predicting the graph. The shared encoding therefore does not require the model to interpret every positive entry as a direct edge or to reproduce the prior as its output.

3.Robustness to imperfect knowledge is not sufficiently established. Masked ground-truth priors and fully random priors are useful controls, but leave a gap between correct knowledge and realistic mixtures of correct, incorrect, and contradictory expert statements.
We thank the reviewer for the comment. We agree that masked ground-truth priors and fully random priors do not adequately capture the mix of correct, incorrect, and contradictory statements that can occur in expert knowledge. We use coarse-grained relations to make prior knowledge easier to obtain and to reduce errors introduced by overly specific judgments. For example, suppose the true causal structure is A -> B -> C. Even without knowing the direct connections, an expert can provide the coarse-grained statement “A may influence C” to guide learning. This statement leaves open whether there is a direct causal edge from A to C. Similarly, the topology of a production line or the order of its operations may allow an expert to conclude that “C cannot influence A” and encode C -> A as a forbidden edge without knowing the complete causal structure. Experts can often make such coarse-grained judgments with greater confidence. In practice, users can also select high-confidence statements for inclusion.

When expert statements are encoded using {0,1,−1}\{0, 1, -1\}, explicit contradictions can be identified, for example, when one statement affirms a relation and another denies that same relation. Such conflicts can be referred to experts for review. If a conflict remains unresolved, the relation can be marked as unknown -1 to limit the amount of unreliable information supplied to the model.

To address this gap in our experiments, we have added robustness tests that mix correct and incorrect prior entries in controlled proportions, as detailed in Appendix D.5. The results show that the model can still make effective use of prior knowledge over a range of the error rates tested. We will limit our robustness claims to the conditions supported by these experiments and investigate more complex contradictions and systematic biases in expert knowledge in future work.

Q1:What remains of the gain when a matched pretrained backbone receives the same training distribution, and conventional baselines receive the same usable forbidden relations?
We thank the reviewer for the question. The ablation results in Figures 9 and 10 show that the pretrained backbone (BCNP), when given prior knowledge without Kode's accompanying architecture and training strategy, still performs worse than the full Kode model. This indicates that the gains cannot be attributed solely to the knowledge input. Figure 5 also shows that Kode adapts better to graphs with different numbers of variables. This ability reflects the combined effects of the training strategy, padding with a very small value, and knowledge integration, and is one of the contributions of this work.

For the traditional baselines, Appendix D.4 and the two rightmost panels of Figure 11 compare performance when the methods are given the same usable forbidden relations.

Besides, Kode outperforms these baselines even in the “0% knowledge” setting, where Kode itself receives no prior knowledge. This result highlights Kode's performance and indicates that its training also contributes to its advantage.

Q2:How are adjacency-based anchors distinguished from reachability-based priors, and how does performance change under partial corruption of otherwise informative knowledge?
We thank the reviewer for the question. The main purpose of introducing adjacency-based anchors is to diversify the priors seen during pretraining. Kode uses no separate type label to distinguish adjacency from reachability. It treats both as relational information and processes them jointly with the observational data. Because the reachability matrix can be derived from the adjacency matrix, large-scale pretraining with supervision from ground-truth causal graphs helps the model distinguish, to some extent, between direct connections and indirect effects. As a knowledge-informed model, Kode combines prior knowledge with data to predict the graph. A shared encoding therefore does not require it to interpret every positive entry as a direct edge or simply reproduce the prior as its output. How the model interprets each type of prior internally and how it avoids confusion arising from their different meanings remain open questions.

When an initially informative prior is partially corrupted, we distinguish between missing and incorrect information. For missing information, the experiments at different retention rates show that retaining only part of the correct knowledge can still improve performance over the no-knowledge baseline under the settings tested. For incorrect information, the fully random prior serves as a noise control. The experiments in Appendix D.5 address partial corruption more directly by mixing correct and incorrect entries in controlled proportions. The results show that the model can still benefit from the prior over a range of the error rates tested.

Reviewer_6sfwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Kode, a Knowledge-Informed Pretrained Model for Causal Discovery. The authors identify two extremes in existing approaches: relying on strong auxiliary supervision or adopting purely data-driven paradigms without guidance. Neither readily accommodates expert causal knowledge that is incomplete, coarse-grained, or potentially incorrect. Kode bridges these approaches by encoding prior knowledge and observational data separately, then jointly modeling them through alternating attention and cross-source information interactions. It also introduces soft knowledge biases in the latent space to improve robustness to inaccurate priors. Experiments show that Kode outperforms the compared baselines.

Critical strengths
  1. The paper addresses a practically relevant problem, particularly the realistic setting where expert causal knowledge is incomplete and may contain errors.

  2. The architecture is well motivated. It uses separate encoders for prior knowledge and observational data, with alternating attention and cross-source interactions to preserve source-specific information while enabling information exchange. The soft knowledge biases in the latent space offer a useful way to incorporate uncertain priors.

  3. The experiments cover varying levels of prior knowledge, prior perturbations, out-of-distribution (OOD) settings, and real-world data. Together, these evaluations provide substantial support for the method.

Critical weaknesses
  1. The paper’s central research question overlaps with prior work on incorporating partial causal or ancestral knowledge into pretrained causal models. The novelty of the individual architectural components relative to existing work is not sufficiently clear.

  2. Inaccurate priors are mainly simulated by masking or corrupting the ground-truth causal structure. This may not adequately capture systematic errors or varying confidence levels in real-world expert knowledge, so the robustness claims would benefit from further evidence.

Questions for authors
  1. In real-world settings where the ground-truth causal graph is unavailable, how should the strength of the soft knowledge biases in the latent space be chosen?

  2. Would it be useful to include comparisons with causal discovery methods that incorporate soft prior knowledge through loss regularization?

  3. When expert prior knowledge conflicts with the causal evidence from observational data, how does Kode balance the two sources of information?

3 comments

We address the weaknesses identified by the reviewers and respond to their questions below.
1.The paper’s central research question overlaps with prior work on incorporating partial causal or ancestral knowledge into pretrained causal models. The novelty of the individual architectural components relative to existing work is not sufficiently clear.
We thank the reviewer for raising this point. Our architectural contributions are twofold. At the encoding stage, we designed a dual-source encoder that jointly processes observational data and coarse-grained prior knowledge. At the decoding stage, we reintroduce the knowledge embeddings so that the prior can further inform the prediction of the causal structure. Section 4.2 cites the prior work from which the other design elements were adopted. In the revised manuscript, we have distinguished our new designs more clearly from the components adopted from prior work.

Our contributions also extend beyond the architecture. While the baselines primarily use interventional signals or partial causal graphs, we focus on how to use coarse-grained reachability information that does not directly specify causal adjacency. We designed both the architecture and the training strategy to make use of this information. We will summarize our main contributions at line 75 of the manuscript.

2.Inaccurate priors are mainly simulated by masking or corrupting the ground-truth causal structure. This may not adequately capture systematic errors or varying confidence levels in real-world expert knowledge, so the robustness claims would benefit from further evidence.
We thank the reviewer for the suggestion. We would like to clarify that the term “inaccurate priors” groups together two distinct types of prior knowledge in our study. The first consists of masked, coarse-grained reachability relations derived from the true causal structure. Although these relations do not directly specify the causal graph, the information they provide is correct. Describing them as “inaccurate priors” is therefore imprecise. Our experiments show that providing more of this knowledge improves performance. The second type consists of random errors deliberately introduced in our robustness tests. The results in the appendix D.5 show that the model can accept some incorrect prior knowledge under the error types and proportions tested.

Our primary focus remains on enabling the model to use coarse-grained knowledge, such as reachability relations and forbidden-edge constraints. For example, knowing the topology of a production line or the order of its operations can be sufficient to rule out an edge from a downstream stage to an upstream stage. This makes such constraints easier to use in practice. Domain experts can often judge these relations with greater confidence than they can identify direct causal edges. In practice, users can also select high-confidence statements to provide as input.

As an initial exploration of this problem, our work addresses the question “How can a model use coarse-grained knowledge?” through its architecture and training strategy. The reviewer raises equally important questions about systematic errors and varying levels of confidence in expert knowledge. In future work, we will clarify the conditions under which our current robustness findings apply and focus on modeling errors that better reflect those made by experts in practice and on accounting for differences in confidence across prior statements.

Q1:In real-world settings where the ground-truth causal graph is unavailable, how should the strength of the soft knowledge biases in the latent space be chosen?
In real-world applications, even when both the causal graph and its reachability relations are entirely unknown, experts can still readily provide forbidden-edge constraints based on their domain knowledge. This is precisely what motivates our work. For example, Table 1 shows how the topology of a production line or the order of its operations can be used to mark edges from downstream stages to upstream stages as forbidden while leaving all other edges unknown. In this setting, prior knowledge is encoded using only {0,−1}\{0, -1\}.

Even without any prior knowledge, the model still shows an advantage, as illustrated by the “0% knowledge” results in Figures 4 and 6, Table 1, and Figures 13--16.

Q2:Would it be useful to include comparisons with causal discovery methods that incorporate soft prior knowledge through loss regularization?
For PC, LiNGAM, and GES, prior knowledge must be incorporated into each method's testing, fitting, or search procedure, making it difficult to use a common additional loss term across these methods. For NOTEARS-MLP, expert knowledge encoded as {−1,0,1}\{-1, 0, 1\} . Such discretization is difficult to use directly as a target for optimizing the graph's continuous edge weights or as a supervisory signal for those weights. For pretrained models such as CSIvA,AVICI and BCNP, an auxiliary loss for reachability could be added, but it would provide reachability information already implicit in the graph prediction objective. More importantly, changing the training loss alone does not enable a model that lacks a knowledge input to use newly supplied priors at inference time.

We explicitly introduce prior knowledge at both the encoding and decoding stages, allowing it to interact with the observational data within the model and jointly inform structure prediction. Compared with imposing constraints solely through loss regularization, this design allows the model to use prior knowledge supplied at inference time.

Q3:When expert prior knowledge conflicts with the causal evidence from observational data, how does Kode balance the two sources of information?
We thank the reviewer for the question. Kode learns to use observational data and prior knowledge jointly by allowing them to interact during both encoding and decoding. The prior guides the prediction without imposing constraints that the model must satisfy. In the random noise setting, Kode performed at least as well as it did without prior knowledge. The additional robustness tests further show that Kode can tolerate some incorrect priors. Under the conditions tested, a small number of conflicts between prior knowledge and data did not compromise performance relative to the no-knowledge baseline.

The coarse-grained knowledge used in this work also makes prior knowledge easier to obtain. Experts can provide useful, high-confidence judgments without having to distinguish direct from indirect effects. In a production process, for example, experts can use the process topology or the order of operations to identify forbidden edges based on the judgment that “downstream variables cannot influence upstream variables.” Users can further select high-confidence judgments and leave uncertain relations unknown. This coarse-grained formulation reduces the likelihood of introducing priors that conflict with the data, while the model's robustness helps limit the impact of conflicts that do arise.

The reviewer raises an interesting question. A focus of our future work will be to quantify, explain, and adjust the relative influence of prior knowledge and observational evidence when they conflict.

We sincerely thank the reviewers for their time and consideration. We hope our responses have adequately addressed the main concerns raised in the reviews. We would be happy to provide further clarifications or address any additional questions. We look forward to further discussion and would greatly appreciate the reviewers taking our clarifications into account when revisiting their assessments.

Reviewer_F915OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Kode, a knowledge-informed pretrained model for causal discovery that takes observational data together with optional prior knowledge. The prior knowledge is encoded as a reachability matrix with three states: plausible (1), impossible (0), and unknown (-1). The model has a dual-source encoder and a graph decoder that follows BCNP, with a permutation decoder and a lower triangle decoder. Training uses 300,000 synthetic graphs (20 to 40 variables, linear and neural network mechanisms), a curriculum learning strategy that gradually samples sparser priors, and a similarity constraint added to the graph likelihood. Experiments on synthetic data (10 to 40 variables, in-distribution and out-of-distribution), Sachs, and two CIPCaD-Bench datasets (Ultra-processed Food and Tennessee Eastman Process) report gains over classical and pretrained baselines. They also report robustness to a random prior and steady improvement as the retention ratio of the prior increases.

Critical strengths

● A practical problem setting. Experts often know which relations are plausible or impossible without knowing exact edges. Encoding this as a three-state reachability matrix and training one model to use any amount of it fills a real gap between methods that need hard edge constraints and purely data-driven pretrained models. ● A broad evaluation across prior strengths. Each setting is tested with a random prior (r), 0%, 30%, 60%, and 100% retention, and the full ground-truth graph (GT), across graph sizes, densities, and mechanisms. Performance at r stays close to 0% and rises steadily as more correct prior is given. This supports the central claim that the model uses the prior as soft guidance without over-relying on it.

Critical weaknesses

● The simulated prior is much stronger than the "weak" or "coarse" knowledge the paper motivates. Every prior is the true reachability matrix with random masking, and 10% of training priors are ground-truth graphs used as anchors. At high retention, even the Neg Only prior tells the model which variables cannot be ancestors of which, which reveals much of the causal order. Real expert knowledge contains systematic errors, such as wrong directions or missed relations, not random masking of a correct matrix. The only test of incorrect knowledge is the random matrix r, and the paper does not say how r is generated. Without tests that mix correct and incorrect entries at controlled rates, the robustness and practicality claims are not well supported. ● The baseline comparisons are not on equal footing. (a) The in-distribution test data follow Kode's own training distribution (heteroscedastic linear and neural network mechanisms with latent inputs), while CSIvA, AVICI, and BCNP use the BCNP training corpus with a fixed 20 variables. The paper does not explain how these baselines are run at 10, 30, and 40 variables (Fig. 5, Figs. 12 to 15). (b) In the main results, no baseline receives the prior knowledge, and the paper does not compare against methods that use soft or ancestral prior knowledge. (c) Appendix D.4 shows that PC and GES with edge constraints close the gap as more knowledge is given and outperform Kode with the full ground-truth graph. (d) On the real-world data without priors, the gains are small (Tennessee Eastman SHD 35.47 vs. 35.58 for BCNP), and no variance is reported over the 100 trials. The claims of being "first" and "consistently" outperforming baselines need softening or stronger evidence.

Questions for authors

● How is the random prior r generated (density and share of each entry type)? How does Kode perform when only part of the prior is incorrect, for example, 10%, 30%, or 50% of entries flipped? ● How were CSIvA, AVICI, and BCNP run on graphs with 10, 30, and 40 variables, given the 20-variable training corpus described in Appendix A.2.2? Were they retrained on your training dataset? ● Is the test data standardized? With weights sampled from U(-10, 10), varsortability (Reisach et al., 2021) may favor some methods. Results on standardized data would address this. ● Please report standard deviations or confidence intervals over the 100 trials, and the Sachs setup used (observational samples only, and sample size). ● In Fig. 9, the model trained without ground-truth anchors performs worse in the GT setting than at 100% retention. What does this say about how the model interprets the prior?

5 comments

We address the weaknesses identified by the reviewers and respond to their questions below.
1.The simulated prior is much stronger than the ”weak" or "coarse" knowledge the paper motivates.
We thank the reviewer for their comments. We would like to clarify that the prior knowledge used in this paper describes reachability rather than direct causal adjacency between variables. It is therefore not equivalent to a known causal graph and cannot directly serve as the output of causal structure learning.

For example, suppose the true causal structure is A -> B -> C. Even if we do not know the direct connections, we can provide the model with the coarse-grained information that “A may influence C” to aid learning. This information does not determine whether there is a direct causal edge from A to C. Similarly, based on the topology of a production line or the order of its operations, we may infer that “C cannot influence A” and translate this knowledge into a forbidden-edge constraint on C -> A without knowing the complete causal structure.

Thus, the “weak” or “coarse-grained” nature of this prior knowledge lies in the fact that it neither specifies direct causal connections nor determines the complete causal graph. To the best of our knowledge, this paper is the first to systematically use such indirect, coarse-grained relationships to aid causal structure learning. The value of this approach lies in translating domain knowledge that is relatively easy to obtain in practice into an effective learning signal.

2.Every prior is the true reachability matrix with random masking, and 10% of training priors are ground-truth graphs used as anchors. At high retention, even the Neg Only prior tells the model which variables cannot be ancestors of which, which reveals much of the causal order. 
We thank the reviewer for pointing this out. Although 10% of the pretraining samples use the ground-truth adjacency matrix as an anchor, these priors are also randomly masked, so the model receives only part of each matrix. They differ from the coarse-grained priors only in their source: they are derived from the adjacency matrix rather than the reachability matrix.

We agree that negative priors with a high retention rate can reveal substantial information about the causal ordering. However, ground-truth graphs used as anchors and priors with high retention rates account for only a small fraction of the knowledge configurations used during pretraining. Neither is used at inference time. Our experimental results confirm that the model can benefit from weaker, sparser prior knowledge at inference time. Even with no prior knowledge, the model remains competitive.

3.Real expert knowledge contains systematic errors, such as wrong directions or missed relations, not random masking of a correct matrix. 
We share the reviewer's concern that random masking cannot fully simulate systematic errors in real-world expert knowledge. As the earlier example of A -> B -> C illustrates, coarse-grained judgments such as “A may influence C” and “C cannot influence A” are often easier to obtain in practice and can be made with greater confidence than judgments about direct causal connections. In practice, users can also manually select other high-confidence knowledge to provide as input. Such coarse-grained knowledge makes the method easier to use in real-world settings and reduces the likelihood of systematic errors.

The random noise experiments and the robustness tests reported in Appendix D.5 (line 1051) of the revised manuscript further show that, for some of the error rates tested, model performance does not drop sharply and remains above the baseline without prior knowledge. The reviewer's concern about incorrect reachability information raises an interesting question. In future work, we will investigate how to avoid or mitigate the effects of incorrect prior knowledge.

4.The only test of incorrect knowledge is the random matrix r and the paper does not say how r is generated. Without tests that mix correct and incorrect entries at controlled rates, the robustness and practicality claims are not well supported.
We thank the reviewer for pointing this out. In the “Random noise” experiment, each entry of the matrix rr is assigned a value of 00, 11, or −1-1 with equal probability, consistent with the encoding of prior knowledge during training. The results show that, under this random noise setting, the model maintains performance comparable to that obtained without prior knowledge. We have added this detail about how the noise is generated at line 411 of the manuscript.

We agree that this experiment alone is insufficient to fully assess robustness to incorrect prior knowledge. We have therefore added tests in Appendix D.5 of the revised manuscript that mix correct and incorrect entries in controlled proportions to assess the impact of incorrect prior knowledge more directly.
5.The baseline comparisons are not on equal footing. (a) The in-distribution test data follow Kode's own training distribution (heteroscedastic linear and neural network mechanisms with latent inputs), while CSIvA, AVICI, and BCNP use the BCNP training corpus with a fixed 20 variables. The paper does not explain how these baselines are run at 10, 30, and 40 variables (Fig. 5, Figs. 12 to 15). 
This ambiguity arose from an imprecise description in the appendix. To ensure a fair comparison, we generated the pretraining data for each method using the same mechanism we used for our model. We adapted the inputs to each method's requirements to accommodate different numbers of variables. CSIvA uses a seq2seq input format and requires no additional processing. AVICI also supports varying numbers of variables, but its input format requires intervention indicators. BCNP accommodates different numbers of variables through padding, as documented in its open-source code and discussed on OpenReview. We have clarified the training data and model configurations at line 753 of the revised manuscript to resolve this ambiguity.
6. In the main results, no baseline receives the prior knowledge, and the paper does not compare against methods that use soft or ancestral prior knowledge. 
As stated in the paper, to the best of our knowledge, existing work on causal discovery has yet to explore the use of “coarse-grained reachability relations” as defined in this paper as prior knowledge. We therefore cannot make a fair comparison with methods that use soft or ancestral prior knowledge. Using such indirect causal relations as prior knowledge is one of the main contributions of this work.

Existing methods rely on precise prior knowledge of direct causal edges. We provide the corresponding comparison in Appendix D.4, where Kode also outperforms the baselines. In this experiment, Kode continues to outperform the baselines even when no prior knowledge is provided (0% knowledge). This result shows that, although Kode can incorporate precise prior knowledge, its advantage does not arise solely from the additional knowledge provided. The model architecture and training strategy also contribute to its performance.

7.Appendix D.4 shows that PC and GES with edge constraints close the gap as more knowledge is given and outperform Kode with the full ground-truth graph. 
The advantage of PC, NOTEARS, and GES under fully correct direct-edge constraints is expected. These constraints narrow the search or optimization space, and edges specified as mandatory no longer need to be learned from data. When the complete ground-truth graph is supplied as a set of hard constraints, the prior knowledge effectively determines the output, so better performance in this setting is unsurprising. These constraints must accurately specify direct causal relationships, since incorrect constraints can introduce errors into the learned graph.

Kode, by contrast, uses both data and prior knowledge to predict the graph. Additional correct knowledge can improve its performance, but even supplying the complete ground-truth graph does not guarantee exact recovery, since Kode does not simply return the prior as its prediction. We acknowledge that hard-constraint methods have an advantage in this extreme setting. Kode's main contribution is to knowledge-informed causal discovery: it can use incomplete, coarse-grained prior knowledge to improve the causal structure learned from data.

8.On the real-world data without priors, the gains are small (Tennessee Eastman SHD 35.47 vs. 35.58 for BCNP), and no variance is reported over the 100 trials. The claims of being "first" and "consistently" outperforming baselines need softening or stronger evidence.
We thank the reviewer for the suggestion. We agree that, without prior knowledge, the improvement over BCNP on the Tennessee Eastman (TE) dataset is modest. To give a fuller account of the results, we now report standard deviations across 100 trials at line 448 of the revised manuscript. On TE, our model shows little advantage in SHD or in variability across trials. Its advantage in F1 is more pronounced, providing evidence of better performance on that metric.

Our use of “first” refers specifically to the use of coarse-grained reachability relations, as defined in this paper, to aid causal discovery. We will clarify this scope in the manuscript. Our original use of “consistently outperforming” did not adequately distinguish between results obtained with and without prior knowledge. We will describe the gains for each experimental setting and use more measured wording where the improvements are modest.

Q1:How is the random prior r generated (density and share of each entry type)? How does Kode perform when only part of the prior is incorrect, for example, 10%, 30%, or 50% of entries flipped? 
In the “Random noise” experiment, each entry of the matrix rr is assigned one of three values (00, 11, or −1-1) with equal probability, consistent with how prior knowledge is encoded during training. Under this random noise setting, the model maintains performance comparable to that obtained without prior knowledge. We have added this detail about noise generation at line 411 of the manuscript.

We agree that this experiment alone is insufficient to fully assess robustness to incorrect prior knowledge. As noted in our earlier response, we have added tests in Appendix D.5 of the revised manuscript that mix correct and incorrect entries in controlled proportions to assess the impact of incorrect prior knowledge more directly.

Q2:How were CSIvA, AVICI, and BCNP run on graphs with 10, 30, and 40 variables, given the 20-variable training corpus described in Appendix A.2.2? Were they retrained on your training dataset? 
We thank the reviewer for the question. The ambiguity arose from an imprecise description in the appendix. In our experiments, we used the same mechanism to generate pretraining data for all methods and adapted the inputs to each method's requirements. CSIvA and AVICI natively support varying numbers of variables, while BCNP uses padding to accommodate different numbers of variables. We have clarified the training data configuration in the revised manuscript to dispel the impression that all baselines were trained exclusively on datasets with exactly 20 variables. Kode used the same training data as the three baselines, with prior knowledge supplied as an additional input.

Q3:Is the test data standardized? With weights sampled from U(-10, 10), varsortability (Reisach et al., 2021) may favor some methods. Results on standardized data would address this. 
We thank the reviewer for raising this point. We standardized all data to prevent differences in variable scales and marginal variances from providing additional clues about causal order. Our primary reason for sampling weights from U(−10,10)\mathcal{U}(-10, 10) was to increase the diversity of weights in the causal mechanisms. We have emphasized the standardization procedure in the appendix of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.