acceptodds
Under review as a conference paper at ICLR 2027

RECAP: Relative Evaluation with Consensus Augmented Policy Search for Causal Discovery

Abstract

Reinforcement learning can search for causal graphs by scoring sampled structures, but learning a graph policy from limited candidate graphs is often unstable. Rewards vary across inputs, while promising edges found in earlier updates may be lost as the policy changes. We propose RECAP, Relative Evaluation with Consensus Augmented Policy Search, which couples within-input relative evaluation with cross-update structural memory. RECAP uses group-relative advantages to distinguish candidate graphs for the same input and guide the current policy update. To carry structural evidence across updates, it aggregates edges from high-reward graphs into an evolving consensus teacher, whose selective, annealed guidance favors recurring edges without permanently constraining the search. A reference-policy KL term moderates policy changes, while the graph policy samples candidates from contextual variable representations and scores them using a BIC-based reward with an acyclicity penalty. Across four benchmarks spanning Bayesian networks, simulated fMRI, and protein signaling, RECAP achieves the lowest SHD and highest precision, with the best or tied-best F1 on three. Ablations confirm the contributions of both components, while equal-budget comparisons further demonstrate RECAP's F1-SHD advantage.

Primary area:
causal reasoning
Keywords:
Causal Discovery, Reinforcement Learning, Deep Learning

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_BCY2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RECAP, a reinforcement-learning framework for causal structure discovery from observational data. It combines group-relative policy updates with an evolving edge-consensus teacher that retains information from high-reward candidate graphs across iterations. Candidate graphs are evaluated using a BIC-based score with acyclicity penalties, while selective, annealed distillation guides subsequent search. Validation includes four main benchmarks, component ablations, a matched-budget configuration-search comparison, and supplementary experiments on 100- and 200-node graphs. The authors report improved SHD and precision on the main benchmarks, with ablations suggesting that both relative evaluation and consensus guidance contribute to structural recovery.

Critical strengths
  1. RECAP combines within-input relative evaluation with cross-update structural memory. The former distinguishes candidates within the current group, while the latter retains recurring edges from high-reward proposals.

  2. The reported ablations show that removing either relative evaluation or consensus guidance reduces F1 and increases SHD across all four main benchmarks.

Critical weaknesses
  1. The experiments in the main text evaluate the proposed method only on graphs with up to 11 nodes. Although the appendix includes experiments on 100- and 200-node graphs, the method does not achieve the best performance in either setting. I am therefore concerned about its performance on graphs with 20–100 nodes.

  2. Generating and scoring multiple candidate graphs increases computational overhead, leading to slow runtimes as the number of nodes grows.

Questions for authors

Could the authors compare the total runtime required by RECAP and the baseline methods, including main experiments and experiment on graphs with 100 and 200 nodes?

3 comments
AuthorsOpenReview ↗

Response to Reviewer BCY2 — Part 1/2: Recovery and Configuration Search

Thank you for recognizing the complementary roles of relative evaluation and structural memory. We address recovery at larger graph sizes below.

W1 Recovery on medium-sized and larger graphs.

We add a 50-node experiment with matched configuration-search budgets. Within the rebuttal period, we select ALIAS, DAGMA, SDCD and GraNDAG as representative baselines covering reinforcement-learning search and continuous optimization. RECAP and each baseline receive ten configuration trials, compared with twenty in the paper. Tuning uses independently generated development data under the same generative settings as the test data. Selected configurations are evaluated on five test datasets. Each method can therefore adapt its configuration to the data-generating setting before evaluation on the shared tests.

Each Erdős–Rényi DAG has 50 nodes and approximately 100 directed edges. Each dataset contains 2,000 independent observations from linear Gaussian structural equations with equal noise variances. Results are means ± sample standard deviations. F1 measures directed-edge recovery.

MethodTest SHD ↓Test F1 ↑
RECAP0.20 ± 0.451.00 ± 0.00
ALIAS1.60 ± 2.190.99 ± 0.01
DAGMA3.60 ± 1.670.98 ± 0.01
SDCD98.80 ± 27.420.49 ± 0.09
GraNDAG87.20 ± 15.170.29 ± 0.16

RECAP has the lowest mean SHD and highest mean F1, recovering the directed structure almost exactly. ALIAS and DAGMA also have high F1. RECAP further reduces the remaining structural errors. With F1 near its upper bound, SHD makes this difference clearer: the mean graph-edit distance is 0.20.

With configurations selected on separate development data, the five-test average assesses recovery on independent samples. RECAP's lower SHD distinguishes it even from the near-complete recovery of ALIAS and DAGMA.

Appendix G.2, Table 7 also reports competitive recovery on 100-node nonlinear graphs. RECAP has SHD of 93.8 ± 13.3, compared with the lowest cited baseline value of 105.0 ± 19.5. The baseline values come from Appendix D, Table 8 of Yuan et al. [1]. These results and the additional 50-node experiment extend the evidence beyond small-graph tasks.

The nonlinear 100-node result addresses a different recovery regime. Here, the advantage is lower graph-edit distance relative to the cited baselines. The 50-node study provides a direct comparison under the same tuning budget and test datasets. Taken together, the two experiments show where RECAP is competitive and extend the empirical evidence across graph sizes and structural equation models.

Configuration search and the learning mechanisms

The Fixed Config. ablation in Section 4.3 and Appendix G.1, Table 6 retains relative evaluation and consensus guidance while removing outer search. On NetSim, it achieves F1 of 0.33 and SHD of 12.00, compared with 0.56 and 10.90 for RECAP. Configuration selection helps choose suitable capacity and training parameters.

Tuning alone does not explain the advantage. In Section 4.4, Table 3 and Appendix G.5.1, Table 10, eligible learned methods receive the same outer-search budget. ABCNAS-RL retains its original search procedure. RECAP still has the highest F1 and lowest SHD. The strongest competing values are 0.51 and 11.60, respectively.

The fixed-configuration and matched-budget comparisons address different explanations. The first measures how much choosing a suitable configuration helps RECAP. The second asks whether giving competing methods the same opportunity for configuration search removes its advantage. Together they show that tuning contributes to performance, while differences in the learning procedure remain important under matched search budgets.

Table 6 shows that removing either learning mechanism reduces recovery across all four benchmarks. Relative evaluation gives the largest F1 gain on NetSim. Consensus guidance gives the largest SHD improvement on Sachs. These results support their complementary roles in comparing current candidates and retaining information for later updates. The new teacher and interaction experiments are reported in our responses to Reviewers GbrL and 8VsV, each in Part 1/3.

Reference

[1] Wei Yuan, Zixuan Shao, and Shuhui Wang. Global Directional Priors with Local Statistical Validation for Scalable Causal Discovery. International Conference on Machine Learning, 2026.

AuthorsOpenReview ↗

Response to Reviewer BCY2 — Part 2/2: Runtime and Total NAS Cost

W2 & Q1 End-to-end runtime and total NAS cost.

We report complete runs at 20, 50 and 100 nodes. Each size uses a shared ER linear Gaussian dataset with 2,000 observations. All 16 methods run on the same Xeon Platinum 8558 and H100 server. Each timing includes training, graph selection and postprocessing at fixed settings.

Method20 nodes / s50 nodes / s100 nodes / s
PC3.898.7415.86
GES23.333959.9015591.01
ICA-LiNGAM4.406.6411.79
NOTEARS14.72195.891033.66
GraNDAG22.0938.0764.94
GOLEM339.57372.27379.11
SDCD136.60162.35159.63
COSMO182.27181.71187.80
DAGMA14.1620.0737.52
DiffAN36.96647.71— OOM
CaPS14.53434.418224.86
CORL204.38961.223684.99
ALIAS28.38216.20743.99
ABCNAS-RL193.612517.268524.59
RL-BIC526.232050.07— OverflowError
RECAP51.44205.41962.74

OOM denotes GPU out-of-memory for DiffAN at 100 nodes. RL-BIC encounters an integer OverflowError in the gCastle implementation.

RECAP is faster per run than CORL and ABCNAS-RL at all three sizes and uses less peak GPU memory at 100 nodes. It retains multi-candidate exploration and teacher guidance at lower per-run resource cost than these baselines. DAGMA and SDCD are faster at 50 and 100 nodes, reflecting the efficiency of dedicated graph optimization.

At 100 nodes, RECAP takes 962.74 seconds, versus 3684.99 for CORL and 8524.59 for ABCNAS-RL. This lowers the end-to-end cost of each configuration evaluation while retaining exploration and guidance.

Total search cost at 200 nodes

We also measure 200-node search on 2,000 observations from linear Gaussian structural equations. Given the rebuttal period, we use NOTEARS as a representative differentiable causal discovery baseline.

MethodSingle-run time / sNAS trials
RECAP3674.4120
NOTEARS6953.16Not applicable

RECAP includes NAS, so its complete search requires repeated training. Using the paper's 20 trials, estimated total cost is 3674.41 × 20 = 73488.20 seconds, or 20.41 hours. A complete NOTEARS run at fixed settings takes 1.93 hours. RECAP is faster for one run, but its cumulative NAS cost exceeds this fixed-configuration baseline. Reducing configuration-search cost is therefore important for larger graphs.

Single-run time measures the cost of obtaining a graph at fixed settings. The NAS estimate covers repeated training to select those settings. Equal trial counts align configuration evaluations, while total runtime depends on the cost of each architecture and optimizer.

Further scaling

RECAP focuses on improving graph search through relative evaluation and structural memory. SDCD [2] combines a stable acyclicity constraint with sparse-graph training and scales to thousands of variables. DAGMA [3] and DAG-NoCurl [4] improve efficiency through their acyclicity formulation and graph representation.

RECAP separates candidate generation and scoring from the two learning mechanisms. We plan to combine them with sparse candidate screening, more efficient acyclicity computation and order-based graph generation. These directions target lower search cost while retaining the benefits of candidate comparison and historical guidance.

Candidate screening targets the number of connections explored. Efficient acyclicity computation and DAG parameterization aim to reduce the work needed to enforce valid structure. Relative evaluation and the teacher can then use the resulting candidate evaluations to guide recovery.

References

[2] Achille Nazaret, Justin Hong, Elham Azizi, and David Blei. Stable Differentiable Causal Discovery. International Conference on Machine Learning, 2024.

[3] Kevin Bello, Bryon Aragam, and Pradeep Ravikumar. DAGMA: Learning DAGs via M-matrices and a Log-Determinant Acyclicity Characterization. Advances in Neural Information Processing Systems, 2022.

[4] Yue Yu, Tian Gao, Naiyu Yin, and Qiang Ji. DAGs with No Curl: An Efficient DAG Structure Learning Approach. International Conference on Machine Learning, 2021.

Thank you for your suggestions on scalability. We look forward to your feedback on the new experiments and welcome further discussion.

AuthorsOpenReview ↗

Follow-up on the Scaling Experiments and Runtime Analysis

Dear Reviewer BCY2,

Thank you again for highlighting the importance of scalability and computational efficiency. In response, we have added further experiments on medium-sized and larger graphs, together with comprehensive runtime comparisons under consistent computational settings. We have also provided an analysis of the total NAS cost to offer a more complete picture of the computational requirements.

As the discussion period is nearing its end, we wanted to check whether our additional experiments and clarifications have addressed your concerns, or if there are any further questions you would like us to discuss. We would be more than happy to continue the discussion and provide any additional information that might be helpful.

Thank you again for your time and consideration.

Best regards,
The Authors

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

RECAP searches for a graph of cause-and-effect relationships. It learns from two sources: how candidate graphs compare with one another on the current data, and which connections have appeared repeatedly in good graphs during earlier searches. The paper tests whether these two ideas improve graph recovery on four benchmarks.

Critical strengths
  1. The two sources of feedback address a reasonable search problem: a method should learn from current candidates without forgetting promising connections found earlier. In the authors' tests, removing either source makes the recovered graphs worse.
  2. The paper compares several kinds of graph error across four benchmarks. RECAP reports strong results for the number of incorrect or missing connections, especially on NetSim.
Critical weaknesses
  1. The paper does not clearly show what is new beyond combining familiar ideas. Comparing current candidates by relative score and averaging connections from earlier high-scoring candidates are both established strategies. Section 2.2 does not explain why existing causal-graph search methods need this particular combination. The mathematical appendix mostly describes properties of the separate ingredients; one result assumes an ideal optimization that differs from the actual training process. A comparison with a simpler method that remembers edges from earlier good graphs would make the contribution clearer.
  2. Key details of the comparison are missing. The paper scores graphs using a regression-based formula, but does not explain how it applies that formula to the discrete Asia and Cancer data (Appendix Eq. 35). It also does not explain how it scores graphs when the data cannot distinguish some edge directions, or when a baseline leaves edges undirected. Table 5 lists settings that could be tried, but not the settings ultimately chosen. The number of tuning trials, graph-pruning rules, and the meaning of the claimed equal search budget are also unclear. These choices can materially change the reported graph scores.
Questions for authors
  1. How much better is RECAP than a simpler method that averages edges from previous high-scoring graphs, with the same graph model and evaluation budget?
  2. How is the regression-based score in Eq. 35 applied to the discrete Asia and Cancer data? How are undirected or observationally indistinguishable edges scored?
  3. What settings were selected for each method, how many settings were tried, and what exactly was held equal in the budget comparison?
  4. Table 2 reports a mean structural Hamming distance of 10.91 for NetSim across ten runs. Since this distance counts discrete graph edits, was each run also averaged over multiple graphs or other units?
4 comments
AuthorsOpenReview ↗

Response to Reviewer GbrL — Part 1/3: Teacher Design and Mechanism Evidence

Thank you for recognizing the motivation, ablations and recovery results across four benchmarks.

W1 & Q1 Why combine relative evaluation with a consensus teacher?

Relative evaluation guides the current update. The teacher retains useful structural information for later updates. A graph reward evaluates a complete structure, while the policy must update interdependent edge decisions. Comparing candidates sampled from the same input focuses updates on their structural differences. The teacher aggregates edge probabilities across input groups, weighted by their contribution to selected high-scoring candidates. These soft targets preserve promising connections as candidate groups change and guide new edge combinations.

EMA discounts older targets as the sampling distribution changes. The mask focuses guidance on supported connections. Annealing gradually reduces its strength so that later rewards can revise early preferences.

Temporal memory and teacher controls

We compare RECAP with the proposed history average and related controls using the same parameters. Random DAGs have 15 nodes and exactly 15 edges. Each dataset has 2,000 observations from nonlinear MLP structural equations with additive Gaussian noise. Results are means ± sample standard deviations over ten runs per condition.

SettingF1 ↑SHD ↓
No teacher0.49 ± 0.2710.80 ± 4.94
Probability EMA0.51 ± 0.2910.90 ± 5.57
Current consensus0.46 ± 0.2411.40 ± 3.86
History average0.45 ± 0.2512.20 ± 5.14
All-candidate EMA0.52 ± 0.2511.20 ± 5.29
No selection mask0.53 ± 0.2310.20 ± 4.42
No annealing0.49 ± 0.2611.10 ± 5.20
RECAP0.58 ± 0.259.40 ± 4.55

RECAP has the highest mean F1 and lowest mean SHD. Current consensus discards previous targets. History averaging weights early and recent targets equally. EMA retains history while adapting to the changing policy. Probability EMA averages predictions without reward selection. All-candidate EMA retains adjacency aggregation but removes reward selection. These controls distinguish temporal averaging from the information selected for guidance.

RECAP recovers 7.9 true edges, versus 6.3 without a teacher, with 3.0 false positives in both cases. All-candidate EMA recovers 7.0 true edges with 4.0 false positives. RECAP thus recovers more true connections without expanding false predictions. Together with the mask and annealing controls, these results support selecting historical information and regulating its use during search.

Without the mask, guidance includes less-supported connections. Without annealing, its loss weight stays fixed. RECAP outperforms both controls, supporting selective guidance whose strength decreases during training.

Interaction with relative evaluation

All four conditions below use the same ten 100-node Erdős–Rényi DAG datasets with an expected 200 edges. Each contains 2,000 observations from linear Gaussian structural equations. Results are means ± sample standard deviations.

EvaluationSettingF1 ↑SHD ↓
PooledNone0.26 ± 0.24254.00 ± 79.70
PooledConsensus0.38 ± 0.21264.40 ± 72.40
Within-group relativeNone0.63 ± 0.31160.20 ± 140.20
Within-group relativeConsensus (RECAP)0.81 ± 0.1785.20 ± 101.60

Adding the teacher under relative evaluation reduces SHD by 46.8% and improves F1. The average predicted edge count remains similar, but true positives increase and false positives decrease. Under pooled evaluation, F1 improves while SHD increases. The benefit therefore depends on how candidate feedback is used. The results support complementary roles for current within-group comparisons and historical structural guidance.

The shared teacher-quality analysis is reported in our response to Reviewer 8VsV, Part 1/3. On the same ten 100-node datasets, RECAP improves both F1 and SHD over probability EMA on nine datasets. Its guidance mask is smaller but covers more true edges and contains a higher proportion of true edges.

Appendix B.5 explains the role of annealing under a fixed base objective and idealized optimization assumptions. Under these assumptions, limiting solutions remain optimal or stationary for the base objective as the guidance coefficient approaches zero. The experiments above assess recovery under the actual training procedure.

AuthorsOpenReview ↗

Response to Reviewer GbrL — Part 2/3: Discrete Scoring and Graph Evaluation

W2 & Q2 Discrete-variable scoring and evaluation of unoriented edges.

Regression surrogate and categorical BIC

Eq. 35 is a regression-based surrogate for numerically encoded discrete variables on Asia and Cancer. Each candidate defines parent sets for predicting variables. The score combines residual sum of squares with a complexity penalty. It measures predictive fit under the encoding rather than categorical likelihood. We will clarify this in Appendix E.3.1.

The coding and categorical BIC conditions use the same experimental settings as RECAP for each dataset. All conditions use the same ten samples of 2,000 observations. Results are means ± sample standard deviations over ten runs.

DatasetConditionF1 ↑SHD ↓
AsiaRECAP0.61 ± 0.035.10 ± 0.74
AsiaSwapped coding + regression score0.08 ± 0.1012.30 ± 5.27
AsiaOriginal coding + categorical BIC0.42 ± 0.138.80 ± 2.15
CancerRECAP0.57 ± 0.003.00 ± 0.00
CancerSwapped coding + regression score0.55 ± 0.032.20 ± 0.42
CancerOriginal coding + categorical BIC0.51 ± 0.322.30 ± 1.25

Categorical BIC lets RECAP use a discrete likelihood while retaining both learning mechanisms. On Asia, it recovers 4.0 true edges on average with recall of 0.50, matching RECAP's mean recall. The F1 difference mainly reflects precision. This demonstrates compatibility with discrete scoring, while its effect on accuracy depends on the dataset and configuration.

On Cancer, categorical BIC predicts 3.0 edges on average, including 1.8 true edges. Recoded regression predicts 3.3 edges, including 2.0 true edges. Sparser predictions reduce both true and false positives. Categorical BIC has slightly lower F1 and similar mean SHD, with greater variation across runs. RECAP has the highest mean F1 on both datasets.

Coding invariance of the score does not imply invariance of the search. For a fixed graph, swapping binary labels preserves the residual sum of squares of regression with an intercept and the complexity penalty in exact arithmetic. The policy nevertheless receives different numerical inputs, which can change sampling and updates. The Asia and Cancer results illustrate this distinction. RECAP's modular interface supports categorical BIC for discrete conditional distributions and separate improvements to categorical input representations.

Relabeling lowers F1 from 0.61 to 0.08 on Asia and from 0.57 to 0.55 on Cancer. With observations held fixed, this tests sensitivity to numerical inputs. Categorical BIC instead changes how parent sets are scored.

Skeleton and CPDAG evaluation

PC and GES can output CPDAGs with unoriented edges, while the other evaluated methods output directed graphs. Appendix E.4 evaluates directed recovery against a reference graph. Observationally equivalent DAGs can therefore receive different scores. The original adjacency-based evaluation counts an unoriented connection represented by symmetric entries as two directed predictions.

The additional evaluation separates connection recovery from orientation states. Skeleton metrics count each connection once. CPDAG pair-state differences compare the directed, undirected or absent state of each unordered node pair.

NodesMethodSkeleton F1 ↑Skeleton SHD ↓CPDAG pair differences ↓
20PC0.72 ± 0.0820.60 ± 6.3525.60 ± 6.95
20GES0.72 ± 0.1432.60 ± 23.8447.20 ± 32.20
20RECAP0.98 ± 0.031.60 ± 2.192.00 ± 2.74
50PC0.69 ± 0.0861.80 ± 20.3978.20 ± 20.90
50GES0.83 ± 0.0542.60 ± 18.8156.40 ± 23.68
50RECAP0.93 ± 0.0515.20 ± 11.7318.40 ± 12.86
100PC0.67 ± 0.07131.00 ± 27.47162.40 ± 27.47
100GES0.72 ± 0.14161.00 ± 125.59186.60 ± 133.64
100RECAP0.93 ± 0.0429.80 ± 16.1536.20 ± 19.15

Each size uses five shared ER linear Gaussian datasets with an expected edge count twice the number of nodes and 2,000 observations per dataset. Methods use fixed configurations. Results are means ± sample standard deviations over five runs.

RECAP has higher mean skeleton F1, lower mean skeleton SHD and fewer mean CPDAG pair-state differences at all three sizes. Its recovery advantage remains when reversible directions within an equivalence class are allowed.

AuthorsOpenReview ↗

Response to Reviewer GbrL — Part 3/3: NAS Protocol and Statistical Corrections

W2 & Q3 NAS budgets, search ranges and output rules.

Each method receives 20 configuration trials. Original output thresholds and postprocessing are retained and excluded from NAS. They are not adjusted using test-graph F1 or SHD. RECAP's search space is in Appendix E.3.2 and Table 5. ABCNAS-RL retains its original search procedure.

MethodParameters and suggested rangesOriginal output rule
NOTEARSlambda1: LogU[10⁻², 1]; max_iter: {50, 100, 200}; h_tol: LogU[10⁻¹⁰, 10⁻⁶]0.30
GraNDAGhidden_dim: {32, 64, 128, 256}; hidden_num: Int[2, 9]; batch_size: {32, 64, 128}; lr: LogU[10⁻⁴, 3×10⁻³]; iterations: Int[2000, 8000]Remove edges from weakest to strongest until acyclic
RL-BIC / ABCNAS-RLhidden_dim / num_heads: pairs listed below*; num_stacks: Int[2, 9]; decoder_hidden_dim / hidden_dim: {0.25, 0.50, 1.00}; batch_size: {32, 64, 128}; lr1_start: LogU[10⁻⁴, 3×10⁻³]; nb_epoch: Int[2000, 8000]0.30
GOLEMlearning_rate: LogU[10⁻⁴, 3×10⁻³]; num_iter: Int[2000, 8000]; lambda_1: LogU[2.00×10⁻³, 0.20]; lambda_2: LogU[0.50, 50]0.30
CORLembed_dim: {32, 64, 128, 256}; encoder_heads: {8, 16}; encoder_blocks: Int[2, 9]; batch_size: {32, 64, 128}; actor_lr: LogU[10⁻⁴, 3×10⁻³]; critic_lr: LogU[10⁻⁴, 3×10⁻³]; iteration: Int[2000, 8000]0.30
DiffANlearning_rate: LogU[10⁻⁴, 3×10⁻³]; batch_size: {32, 64, 128}; epochs: Int[2000, 8000]1.00×10⁻³, CAM pruning significance cutoff
DAGMAlr: LogU[10⁻⁴, 3×10⁻³]; lambda1: LogU[3.00×10⁻³, 0.30]; T: {3, 5, 7}; mu_init: LogU[0.10, 10]; mu_factor: LogU[0.01, 0.50]; warm_iter: Int[2000, 8000]; max_iter: Int[2000, 8000]0.30
CaPSeta_G: LogU[10⁻⁴, 10⁻²]; eta_H: LogU[10⁻⁴, 10⁻²]1.00×10⁻³, CAM pruning significance cutoff
COSMOlearning_rate: LogU[10⁻⁴, 3×10⁻³]; batch_size: {32, 64, 128}; max_epochs: Int[2000, 8000]; l1_adj_reg: LogU[10⁻⁴, 10⁻³]; l2_adj_reg: LogU[10⁻³, 5×10⁻³]; priority_reg: LogU[10⁻³, 3×10⁻³]; temperature: LogU[5×10⁻⁴, 10⁻³]; shift: LogU[5×10⁻³, 2×10⁻²]0.30
SDCDlearning_rate, both stages: LogU[10⁻⁴, 3×10⁻³]; batch_size, both stages: {32, 64, 128}; n_epochs, each stage: Int[2000, 8000]; dim_hidden: {32, 64, 128, 256}; num_layers: Int[2, 9]0.10
ALIASrl_params.learning_rate: LogU[10⁻⁴, 3×10⁻³]; steps_per_env: Int[2000, 8000]; n_envs: {32, 64, 128}; rl_params.batch_size: {32, 64, 128}; rl_params.ent_coef: LogU[1.00×10⁻³, 0.90]; rl_params.max_grad_norm: LogU[0.10, 2.00]; rl_params.normalize_advantage: {true, false}0.30
RECAPSee Appendix E.3.2 and Table 50.30

LogU is log-uniform sampling. Int is an integer range. Braces indicate discrete choices. *Hidden-dimension/attention-head pairs: 32/8, 32/16, 64/8, 64/16, 64/32, 128/8, 128/16, 128/32, 256/16, 256/32, 256/64.

Training epochs or updates span 2,000–8,000. NOTEARS uses outer-solver iterations. DAGMA separates stages and updates. CaPS tunes kernel parameters with default pruning and edge addition. Output thresholds select final edges. The teacher threshold in Table 5 selects guidance connections.

Equal trials align configuration counts, not runtime. We estimate NAS cost as measured single-run time × trial count. See Reviewer BCY2, Part 2/2, for timings.

NAS selects configurations. Relative evaluation and consensus guidance govern learning. Fixed Config. in Section 4.3 and Appendix G.1, Table 6 removes outer search but retains both mechanisms. Adding NAS changes NetSim F1/SHD from 0.33/12.00 to 0.56/10.90. Under matched budgets in Section 4.4 and Appendix G.5.1, RECAP has the highest F1 and lowest SHD. The best competing values are 0.51 and 11.60.

Table 6 shows that removing either mechanism reduces recovery across all four benchmarks. Relative evaluation gives the largest F1 gain on NetSim, from 0.29 to 0.56. Consensus guidance gives the largest SHD improvement on Sachs, from 14.67 to 11.00. Both mechanisms contribute beyond configuration selection.


Q4 Correction of summary statistics.

One run was counted twice in summary statistics. Tables 2 and 3 use the same ten RECAP runs. Corrected graph distance is 10.90 ± 0.99 and recall is 0.64 ± 0.09. The other five metrics are unchanged after rounding. This also corrects Tables 6 and 10 and related text.

Thank you for identifying the error. We welcome further discussion of the method and evaluation.

AuthorsOpenReview ↗

Follow-up on Methodological Clarifications and Additional Experiments

Dear Reviewer GbrL,

Thank you again for your detailed feedback and insightful suggestions. We have made substantial efforts to address the concerns raised in your review, including additional comparisons of the consensus mechanism, clarifications on discrete-variable scoring and graph evaluation, and further details on the configuration search protocol. We have also corrected the reported summary statistics.

With the discussion period coming to a close, we would like to ask if you have any remaining questions or concerns regarding our responses and the additional results. We sincerely welcome any further discussion and would be glad to provide additional clarification if needed.

Thank you for your time and valuable feedback.

Best regards,
The Authors

Reviewer_8VsVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RECAP, a reinforcement-learning-based framework for causal discovery that combines group-relative policy updates with a cross-iteration consensus teacher. The method uses relative rewards to guide graph search and aggregates high-reward structural information to improve subsequent updates. Experiments across several benchmarks demonstrate competitive performance, particularly in SHD and precision.

Critical strengths
  • The motivation for combining relative evaluation and structural memory is intuitive and clearly presented.
  • The paper includes comparisons with a broad range of baselines, along with ablation studies and matched-budget experiments.
  • RECAP achieves promising results in structural recovery, especially in SHD and precision.
Critical weaknesses
  1. The method combines existing optimization techniques, but it remains unclear whether the consensus teacher provides meaningful structural memory beyond standard self-distillation. This weakens the methodological contribution.

  2. The performance advantages are less consistent on larger graphs, and the computational cost is not sufficiently evaluated against competing methods. This raises concerns about practical applicability.

Questions for authors
  1. How does the consensus teacher differ fundamentally from standard EMA-based self-distillation, and what evidence demonstrates that it preserves useful structural information?
  2. Which consensus aggregation strategy is used in the experiments: sampled adjacency matrices or policy edge probabilities?
  3. How are reward differences across inputs handled when selecting high-reward graphs for the consensus teacher?
  4. Can the authors provide stronger large-scale comparisons and end-to-end runtime measurements under comparable computational budgets?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 8VsV — Part 1/3: Teacher Evidence and Structural Memory

Thank you for your careful review and for recognizing the motivation, baseline comparisons and ablations.

W1 & Q1 What does the consensus teacher add beyond ordinary EMA?

The teacher uses candidate quality to select information for later search. It aggregates edge probabilities from groups that produce high-scoring candidates. EMA retains these preferences across updates. The mask focuses distillation on supported connections, and annealing allows new rewards to revise earlier preferences.

Aggregation and selection controls

All conditions use the same ten paired 100-node Erdős–Rényi DAG datasets and training parameters. Each graph has an expected 200 edges and 2,000 observations from linear Gaussian structural equations. Results are means ± sample standard deviations.

SettingDirected F1 ↑SHD ↓
Reward-selected adjacency EMA0.73 ± 0.26122.80 ± 139.24
Probability EMA0.75 ± 0.24106.40 ± 108.67
Group-relative Top-M0.72 ± 0.33117.20 ± 146.63
Equal-count random selection0.76 ± 0.16108.60 ± 89.35
RECAP0.81 ± 0.1785.20 ± 101.58

RECAP has the highest mean F1 and lowest mean SHD. It improves both metrics over probability EMA on nine of ten paired datasets. Uniform probability EMA weights every input group equally. RECAP gives greater weight to groups that contribute more high-scoring candidates. The advantage over equal-count random selection further supports using candidate quality rather than retaining a fixed number of candidates without regard to reward.

Adjacency EMA stores binary indicators from sampled graphs. Probability aggregation preserves the strength of an edge preference even when an edge is absent from a sample. Its better mean recovery supports retaining these continuous preferences. Group-relative Top-M ranks candidates within each group. Global Top-M preserves absolute quality differences across groups and gives better mean recovery in this comparison.

Information retained for distillation

The teacher supplies soft targets for later sampling and policy updates. Candidate scoring, selection and pruning determine the final DAG. At the fixed teacher threshold of 0.2, we measure mask size, the proportion of true edges within the mask and true-edge coverage. Edge AUPRC measures global edge-ranking quality. Results below are means ± sample standard deviations at the end of the ten paired runs.

SettingTrue edges within mask / % ↑True-edge coverage / % ↑Mask size / % of possible edgesEdge AUPRC ↑
Probability EMA2.80 ± 0.9485.30 ± 29.3959.90 ± 13.290.040 ± 0.013
RECAP3.12 ± 0.8789.34 ± 26.9056.19 ± 7.600.042 ± 0.011

RECAP covers more true edges within a smaller mask, with a higher proportion of true edges inside it. Reward selection thus makes guidance more selective. Subsequent updates receive more focused soft targets while graph rewards continue to guide exploration and revision.

The two mask measures describe different properties. True-edge coverage asks how much of the reference structure receives guidance. The proportion of true edges within the mask asks how concentrated that guidance is on correct connections. Coverage rises from 85.30% to 89.34%, while mask size falls from 59.90% to 56.19% of possible edges. Both measures improve together. The pattern is consistent with retaining more useful connections rather than simply making distillation more extensive.

Edge AUPRC and final DAG recovery evaluate different parts of the search. AUPRC ranks all possible edges. Distillation guides selected connections, and rewards evaluate complete graphs. The mask statistics characterize the retained information. The paired recovery gains show that search benefits from it.

Joint F1 and SHD gains on nine datasets show that better edge recovery accompanies fewer structural errors. The mask analysis characterizes the information retained for subsequent updates.

The shared 15-node nonlinear controls appear in our response to Reviewer GbrL, Part 1/3. RECAP has higher mean F1 and lower mean SHD than current consensus, history averaging and the other teacher controls. It recovers 7.9 true edges with 3.0 false positives, compared with 6.3 and 3.0 without a teacher. All-candidate EMA recovers 7.0 true edges with 4.0 false positives. These results support selecting information for memory as well as retaining it over time.

AuthorsOpenReview ↗

Response to Reviewer 8VsV — Part 2/3: Probability Aggregation and Global Selection

Q2 Why aggregate probabilities associated with selected graphs?

The main implementation aggregates policy edge probabilities associated with selected high-reward graphs. Candidates from the same input group share a probability matrix. Their Bernoulli samples and rewards can differ. A group receives more weight when more of its candidates are selected. Graph rewards therefore determine how strongly each group's edge predictions contribute to the teacher.

Selection determines the weight of a group's prediction. Selecting several candidates from one group contributes the same probability matrix several times. The number selected therefore controls that group's influence. The different sampled graphs provide evidence about the quality of the distribution that generated them. Aggregating the associated probabilities retains that distribution's edge preferences, rather than reducing the target to the edges present in one selected draw.

Uniform probability EMA weights every input group equally, including groups whose candidates score poorly. RECAP instead favors predictions from groups that produce more high-scoring graphs. The paired improvements in Part 1/3 support this use of candidate quality.

Probability targets retain more information than a single sampled adjacency. An adjacency records whether an edge appears in one draw. A probability preserves the strength of that preference even when the edge is absent. The higher mean F1 and lower mean SHD than adjacency EMA support retaining these continuous preferences. Reward selection controls each group's contribution, EMA carries it across updates, and masked distillation guides new edge combinations.

The resulting target combines preferences associated with different candidate structures and earlier updates. Its probability matrix need not itself be a DAG. The masked loss transfers these preferences to the policy, which generates new candidate combinations. Complete graphs are then scored and selected. This separation allows structural information from earlier candidates to remain useful even when those exact graphs are no longer sampled.

EMA provides continuity while the policy changes. Current consensus would replace the target with information from the latest candidates alone. A history average would give old and recent targets equal weight. EMA instead gradually reduces the influence of older predictions. Masking and annealing regulate where this memory is used and how strongly it affects later updates. These are the distinct choices examined by the controls in our response to GbrL, Part 1/3.


Q3 Why use relative evaluation for updates but Global Top-M for the teacher?

The two operations use rewards for different purposes. Input blocks condition candidate generation. Within each update, all candidates are scored on the same complete dataset with the same scoring function and penalty coefficients. Their parent sets and cycle structures yield different fitting errors and penalties. Raw rewards therefore compare candidate quality under a common objective.

Centering rewards within each input group lets the policy learn from differences among that group's candidates, even if the group scores poorly overall. The teacher instead retains predictions associated with high-scoring candidates across groups. Group-wise standardization can favor a candidate from a weaker group over one with a better absolute score. Global Top-M preserves the quality differences needed for this selection.

A high within-group rank means that a candidate improves on its group's alternatives. It need not have the best absolute score across the batch. Keeping both comparisons lets the policy learn within groups while the teacher retains information from the highest-scoring candidates across groups.

The selection controls in Part 1/3 support using a common reward scale for teacher construction. Random selection ignores quality. A fixed quota per group weights each probability matrix equally and is equivalent to uniform probability EMA. Group-relative selection prioritizes within-group rankings. Global Top-M has the highest mean F1 and lowest mean SHD among these alternatives.

Relative evaluation improves the policy within each group. Global selection determines which predictions should guide later search. Their combination uses current comparisons and historical information for complementary purposes.

AuthorsOpenReview ↗

Response to Reviewer 8VsV — Part 3/3: Larger Graphs and Computational Cost

W2 & Q4 Recovery on larger graphs and computational cost.

Appendix G.2, Table 7 shows competitive recovery beyond small graphs. The 100-node nonlinear results use 1,000 observations. Baseline values, including GRaSP, come from Appendix D, Table 8 of Yuan et al. [1]. RECAP has lower mean SHD than the best cited baseline, indicating better graph-edit accuracy. For the paired 200-node comparison with 2,000 observations, we run the official OCMB-SciNO implementation. RECAP improves both precision and recall, so its F1 gain reflects better structural coverage and more reliable predicted connections.

The 100-node SHD is 93.8 ± 13.3 for RECAP, compared with 105.0 ± 19.5 for the best cited baseline. This comparison uses published baseline results under the same graph setting, with different graph realizations. The 200-node experiment instead compares methods on shared generated graphs. The two studies provide evidence at different scales and under different comparison protocols.

Our response to Reviewer BCY2, Part 1/2, reports the additional 50-node recovery experiment with matched configuration-search budgets. RECAP achieves the highest mean F1 and lowest mean SHD against ALIAS, DAGMA, SDCD and GraNDAG. This extends the evidence to medium-sized graphs under a shared evaluation protocol.

In the 50-node experiment, configurations are selected on independently generated development data and evaluated on five test datasets. Each method receives ten trials. RECAP reduces mean SHD to 0.20, compared with 1.60 for ALIAS and 3.60 for DAGMA. All three have high F1, so the graph-edit counts reveal the remaining differences in recovery. The result complements the larger nonlinear comparisons with a test of directed recovery on medium-sized linear graphs.

The complete 16-method runtime table is in our response to Reviewer BCY2, Part 2/2. At 20, 50 and 100 nodes, methods use a shared ER linear Gaussian dataset at each size with 2,000 observations. All run on the same Xeon Platinum 8558 and H100 server. Timings include training, graph selection and postprocessing at fixed settings.

RECAP takes 51.44, 205.41 and 962.74 seconds at these three sizes. It is faster than CORL and ABCNAS-RL at all sizes and uses less peak GPU memory at 100 nodes. It retains candidate exploration and teacher guidance at lower per-run cost than these policy-search baselines. DAGMA and SDCD are faster at 50 and 100 nodes.

Complete NAS requires repeated training. We estimate total cost as single-run time × trial count. The 200-node timings and total-cost calculation appear in the response to BCY2, Part 2/2.

At 200 nodes, one RECAP run takes 3674.41 seconds, compared with 6953.16 seconds for NOTEARS. Multiplying the RECAP measurement by 20 trials gives an estimated 20.41 hours for NAS. The fixed-configuration NOTEARS run takes 1.93 hours. These measurements separate the efficiency of one search run from the cumulative cost of selecting a configuration. They also identify repeated training as an important target for reducing total computation.

Further scaling can combine these learning mechanisms with efficient graph optimization. SDCD uses a stable acyclicity constraint and sparse-graph training to scale to thousands of variables. DAGMA uses a log-determinant acyclicity formulation. DAG-NoCurl learns in DAG space through a graph-potential representation. References for these methods are listed in our response to BCY2, Part 2/2.

RECAP separates candidate generation and scoring from relative evaluation and teacher guidance. We plan to explore sparse candidate screening and more efficient DAG parameterizations to reduce search costs while retaining both learning mechanisms.

Sparse screening can reduce candidate connections. Efficient acyclicity computation can lower evaluation costs. Order-based generation can produce DAGs directly. Relative evaluation and the teacher can then guide search using these cheaper candidate evaluations.

Reference

[1] Wei Yuan, Zixuan Shao, and Shuhui Wang. Global Directional Priors with Local Statistical Validation for Scalable Causal Discovery. International Conference on Machine Learning, 2026.

Thank you for the thoughtful questions. We would be glad to discuss the teacher mechanism and these additional results further.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Additional Results

Dear Reviewers and Area Chair,

Thank you for your time and consideration. We understand that this is a busy period for reviewers, and we would greatly appreciate even a brief response to our rebuttal. We have posted detailed responses to all three reviews, including additional experiments on the consensus mechanism, graph recovery, and computational cost, as well as clarifications on scoring and evaluation.

As the discussion period draws to a close, we would appreciate the reviewers’ feedback on whether these responses address their concerns and which questions, if any, remain unresolved. We would be happy to provide further clarification while there is still time for discussion.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.