Neural causal discovery models learn to read a dataset and output a causal graph in one pass. They are trained on simulated data because real ground-truth graphs are rare. On simulations they look nearly perfect. On real single-cell and river-network data they often lose to crude baselines such as random-forest feature importance. Most attempts to close this synthetic-to-real gap change the model: bigger priors, more mechanism families, test-time adaptation.
This note asks a narrower question that is cheaper to answer: how much of the gap sits in the last step, where a matrix of continuous edge scores is turned into a discrete graph? That step is usually a hard rule. Keep everything above a threshold, then delete edges to break cycles or remove "redundant" direct links. Object detection hit the same wall in crowded scenes. There, hard non-maximum suppression (deleting every candidate box that overlaps a stronger one by more than a fixed amount) threw away real, adjacent objects. Soft-NMS fixed this by decaying overlapping scores instead of zeroing them Soft-NMS - Improving Object Detection with One Line of Code. We propose the causal analogue and use it as a measuring instrument.
When a strong path X→Y→Z has been accepted, the direct edge X→Z is not deleted. Its score is decayed by how much of X's measured effect on Z is already explained by Y. That "overlap" is computed from the data, and interventional data where available, rather than from the model's own scores, which are exactly what become unreliable off the simulator.
Core Claim
The claim has two parts. First, a measurable share of the real-world failure of amortized causal discovery comes from hard decoding, not only from mis-scored edges. Real regulatory and hydrological systems contain feed-forward loops, meaning a regulator that acts on a target both directly and through an intermediate. They also contain feedback. Both structures are rare in the acyclic simulators these models are trained on. Hard transitive pruning and forced cycle-breaking systematically delete true edges in these motifs.
Second, a decoder that decays scores using a data-derived overlap should raise recall at a fixed false-positive budget on real benchmarks. It should do so for neural scorers and for non-causal scorers alike, and its single decay parameter should transfer from synthetic tuning to real data. If raw-score ranking already matches it, the claim is false, and the experiments are designed so that this outcome is visible.
What is new relative to the closest work is where the suppression signal comes from and what the method is used for. Learning Non-maximum Suppression and SEA's learned aggregator both learn how to resolve conflicting candidates, so they inherit whatever the training distribution teaches. Our overlap is a fixed, data-computed quantity with one tuned parameter, and the method doubles as a diagnostic that separates scoring error from decoding error.
Research Program
The decoder works on any score matrix $s_{ik}$ for candidate edges $i \to k$. Like Soft-NMS, it processes edges in descending score order and maintains an accepted set. When an edge $j \to k$ is accepted, every not-yet-processed edge $i \to k$ for which $i \to j$ is also accepted gets the update
$$ s_{ik} \leftarrow s_{ik}\,\exp\!\left(-\frac{o_{ijk}^{2}}{\sigma}\right), $$
where $o_{ijk} \in [0,1]$ is the fraction of $i$'s effect on $k$ that is carried by $j$, and $\sigma$ is the single decay width. This is the Gaussian kernel of Soft-NMS with geometric box overlap replaced by mediated overlap.
We define the overlap two ways. The interventional overlap applies when $i$ was knocked out, as in CausalBench's CRISPR perturbations. It is the share of the knockout-induced shift in $k$ that is predicted by the knockout-induced shift in $j$, estimated by regression across cells in that knockout condition. The observational overlap is $o_{ijk} = \max(0,\, 1 - |\rho_{ik\cdot j}| / |\rho_{ik}|)$. Here $\rho_{ik}$ is the correlation of $i$ and $k$, and $\rho_{ik\cdot j}$ is the partial correlation given $j$. A true feed-forward loop leaves a residual direct effect, so its overlap stays below 1 and the edge survives with a reduced but ranked score. A purely mediated edge has overlap near 1 and decays toward zero.
The same logic replaces hard cycle-breaking. For reciprocal or cyclic candidates, the weaker direction is decayed by how well the stronger direction explains the shared variance, instead of being deleted. This matters because CausalBench's evaluation does not require acyclicity, while decoders such as the one in CauScale: Neural Causal Discovery at Scale enforce it with an ad-hoc cycle-breaking heuristic once cycles appear at 1,000 nodes.
To keep cost manageable at 622 genes, mediators are restricted to the top-$K$ accepted parents of each target. This gives roughly $O(d^2 K)$ overlap computations.
The analysis has three components. First, a decomposition of the gap: for each base scorer, compare real-benchmark performance under raw ranking, hard decoding, and soft decoding. The difference between hard and soft decoding bounds how much of the loss is decoding-induced. Second, a motif audit: on biological reference edges from STRING and CORUM, identify which true edges hard decoding removes, and test whether they sit disproportionately in feed-forward and feedback motifs. Third, a transfer test: tune $\sigma$ on synthetic priors only, then evaluate on real data without retuning.
The base scorers are SEA Sample, estimate, aggregate: A recipe for causal discovery foundation models, CausalTab CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery, and CauScale. None of their code appears among the file's resources, so we will use public checkpoints if they exist and otherwise retrain from the published recipes. That is a real reproducibility risk, and we will report it. We also include GRNBoost and Sortnregress on CausalBench, and VAR and cross-correlation on CausalRivers. Classical estimators come from causality-lab, and time-series baselines from PCMCI-Omega.
The motivating evidence comes from CausalBench CausalBench: A Large-scale Benchmark for Network Inference from Single-cell Perturbation Data. It reports that "most methods either identify a small number of high-confidence edges (high precision, low recall) or a vast number of edges with many false positives." It also reports that interventional methods do not beat observational ones there. A decoder that uses interventions only to judge redundancy is a direct test of whether the interventional signal is being wasted at decoding time.
Baselines
The naive reference point is raw-score ranking: sort all candidate edges by the scorer's output and cut at a fixed edge budget, with no pruning of any kind. This is the most dangerous baseline. Soft decay only reorders edges, and if the scorer's ranking is already right, reordering cannot help. Every comparison reports curves over edge budgets, so no method benefits from a lucky threshold.
The second baseline is each model's own published hard decoder. This means global thresholding for CausalTab, CauScale's threshold plus cycle-breaking, and SEA's output thresholding. It represents current practice and is the thing we claim is harmful.
The third baseline is hard transitive reduction in the style of Computationally and statistically efficient learning of causal Bayes nets using path queries. An edge is deleted whenever an accepted path explains it, using the same overlap measure with a hard cutoff. This is the key ablation. It holds the data-derived overlap fixed and changes only hard versus soft, which isolates the decay rule.
A second ablation replaces the data-derived overlap with a score-derived one, the product of the path's model scores. This isolates the claim that overlap must come from data rather than from miscalibrated model outputs.
The strongest existing methods on each benchmark are also included. On CausalBench these are GRNBoost and the challenge winners Guanlab and BetterBoost The CausalBench challenge: A machine learning contest for gene network inference from single-cell perturbation data. No reported numerical result for these was found on Lacuna, since the file's abstract snippet cuts off, so we will rerun them with the benchmark's own evaluation. On CausalRivers the strongest established methods are VAR and CDMI, which CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series reports "often hitting AUROC scores between 0.80 and 0.85 on 'Random' and 'Close' graph structures." On Sachs, no reported result was found on Lacuna.
All decoders receive identical score matrices and the same tuning budget. That budget is a sweep over one parameter ($\sigma$, or the hard cutoff), tuned on synthetic data only in the main table and on a held-out real split in a secondary table.
Beating the baselines means the following. On CausalBench, soft decoding must lower the false omission rate at matched mean Wasserstein distance, and raise STRING/CORUM precision at matched edge count, relative to both raw ranking and hard decoding, on both K562 and RPE1. On CausalRivers, it must exceed the VAR/CDMI AUROC range when applied on top of those same scorers.
Why not just…?
Why not just tune the hard threshold on real data? Threshold tuning moves the operating point along a fixed ranking. It cannot change which edges rank above which. Our comparisons report full precision-recall and Wasserstein-versus-false-omission curves, so any benefit from threshold choice is already absorbed into the baseline. If soft decoding's curve lies on top of raw ranking's curve, the idea fails, and we will report that.
Why not just use the inverse covariance or partial correlations directly as the edge score? This is a real threat. SEA reports that "inverse covariance alone was found to be as informative as roughly 40 local subset estimates," so a partial-correlation scorer may already remove mediated edges. We include it both as a standalone scorer and as a scorer passed through our decoder. If the standalone version matches neural scorer plus soft decoding, the conclusion is that the neural scorer adds nothing on real data. That is itself a useful finding about the gap, but it would not support the decoder.
Why not just fine-tune the model on real data? CausalRivers shows that fine-tuning on one real region and testing on another "increased performance across almost all graph types." Where such labeled regions exist, fine-tuning is the stronger remedy. It is not available for most single-cell settings, where reference networks are partial and database-derived. The decoder needs no labels and composes with fine-tuning, so on CausalRivers we test fine-tuned VAR and CDMI scores with and without soft decoding to see whether the gains add.
Why not just learn the suppression, as learned NMS or SEA's aggregator do? A learned suppressor trained on simulated graphs learns simulated motif statistics, and those statistics are what fail to transfer. Our single-parameter rule is the controlled alternative. The synthetic-tuned-versus-real-tuned comparison of $\sigma$ directly measures whether the rule transfers where learned components do not.
Proposed Experiments
The main experiment runs every scorer × decoder combination on CausalBench K562 and RPE1, using the benchmark's Wasserstein/false-omission and STRING/CORUM evaluations. It is repeated on RiversEastGermany and RiversBavaria, with RiversElbeFlood as a shift test, and on Sachs. Synthetic sanity checks use the CausalTab mechanism suite, the SEA pretraining library, and SERGIO-GRN. On clean acyclic simulations, soft and hard decoding should roughly tie, since high reported synthetic performance (CausalTab F1 0.952, CauScale 99.6% mAP in-distribution) leaves little room. A large synthetic gain would suggest the decoder is fitting something other than real motifs.
The decisive checks are three. First, the soft-over-raw gain on CausalBench must be present with interventional overlap and larger than with observational overlap. Otherwise the "wasted interventions" explanation is wrong. Second, the motif audit must show that the true edges recovered by soft decoding are enriched in feed-forward and feedback motifs relative to all reference edges. Third, $\sigma$ tuned on synthetic data must land within the flat region of the real-data tuning curve.
If all three hold, we will have shown that part of the synthetic-to-real gap is a decoding artifact that can be fixed without retraining. If the first fails while raw ranking stays competitive, the result locates the gap firmly in the scorers, which narrows where future work should look.
References
- Soft-NMS - Improving Object Detection with One Line of Code
- Learning Non-maximum Suppression
- CauScale: Neural Causal Discovery at Scale
- Sample, estimate, aggregate: A recipe for causal discovery foundation models
- CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery
- CausalBench: A Large-scale Benchmark for Network Inference from Single-cell Perturbation Data
- Computationally and statistically efficient learning of causal Bayes nets using path queries
- The CausalBench challenge: A machine learning contest for gene network inference from single-cell perturbation data
- CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series