One-Step Causal Mechanisms: Counterfactuals from Two Network Passes per Node

Answering a counterfactual question ("what would this patient's blood pressure have been had the dose been different, given everything else we saw?") with a learned causal model takes three steps. First, recover the hidden background noise that produced the observation; this step is called abduction. Second, change the variable of interest. Third, push the recovered noise forward through the changed system. The strongest current learned models do the recovery and the push-forward with diffusion models, which are generators that turn random noise into data by many small denoising steps. Run deterministically, those steps can be reversed. The cost is heavy: each node of the causal graph needs a full numerical integration in each direction, often 50 to 100 network calls apiece. At the image scale, a counterfactual can take 3 to 5 minutes per image. Separately, the fast-generation literature has learned to produce samples in one network call by predicting the average displacement over a time interval rather than the instantaneous velocity. That work optimizes sample quality and says nothing about whether a one-step map remains an accurate, invertible stand-in for a structural equation.

This note proposes parent-conditioned bidirectional one-step flows as causal mechanisms. Each node gets one network pass for abduction and one for prediction. They are trained with an extra objective aimed at exactly the two errors that a known bound says control counterfactual error: round-trip reconstruction error, and leakage of parent information into the recovered noise. The scientific question is whether this objective lets one-step mechanisms keep the counterfactual accuracy of multi-step ones. The competing possibility is that the speed gain silently trades away counterfactual validity while observational fit looks fine.

Core Claim

Our claim is that fast causal generative models should be trained and judged against the counterfactual error budget, not against distribution-matching quality. Two results motivate this. The Diffusion-based Causal Model paper, Modeling Causal Mechanisms with Diffusion Models for Interventional and Counterfactual Queries, models each graph node with a deterministic diffusion process conditioned on its parents. It proves that for encoder-decoder mechanisms, counterfactual error is bounded when reconstruction is accurate and the encoded noise carries no information about the parents. Separately, Exogenous Isomorphism for Counterfactual Identifiability shows that models can match observational data well, as measured by $OBS_{WD}$, a Wasserstein distance to the observed distribution. They can do so while failing badly on $CTF_{RMSE}$, the root-mean-square error of counterfactual predictions, once the noise mapping departs from a component-wise bijection.

One-step generators are trained on objectives that are blind to both conditions. That includes MeanFlow (Mean Flows for One-step Generative Modeling), which reaches FID 3.43 on ImageNet 256×256 in a single step. We therefore expect naive one-step causal mechanisms to reproduce the "good observational fit, bad counterfactual" failure. We expect that explicitly penalizing round-trip error and parent leakage in the one-step encoder and decoder closes most of the gap to multi-step models at a small fraction of the latency.

What is new relative to the closest work is the combination of two things. BiFM: Bidirectional Flow Matching for Few-Step Image Editing and Generation already learns forward and backward average-velocity fields with a consistency term, down to 4 steps. It has no causal graph, no parent conditioning, no Pearl-style interventions, and no independence requirement on the inverted noise. DCM and the ODE-based Triangular Velocity Structural Model have the causal semantics but use multi-step integration. We put the two together, and we test whether the bound that justifies the multi-step model can serve as a training objective for the one-step one.

Research Program

For each node $X_i$ with parents $PA_i$, we train two average-velocity networks with the MeanFlow identity. One runs noise to data and defines a one-step decoder $D_i(u, pa)$. The other runs data to noise, BiFM-style, and defines a one-step encoder $E_i(x, pa)$. On top of the two MeanFlow regression losses $\mathcal{L}_{\text{MF}}$ we add

$$ \mathcal{L}_{\text{cf}} = \mathbb{E}\,\big\|D_i(E_i(x,pa),pa) - x\big\|^2 + \lambda\,\mathrm{HSIC}\big(E_i(x,pa),\, pa\big), $$

where the expectation is over observed $(x, pa)$ pairs for node $i$ and $\lambda$ weights the independence penalty. HSIC is the Hilbert–Schmidt independence criterion, a kernel measure of statistical dependence that is zero when the two arguments are independent. The two terms are exactly the two premises of the DCM bound.

Counterfactual inference then has two phases. Abduction uses factual values only, so it runs for all nodes in parallel in one pass. Prediction proceeds in topological order with one decoder pass per graph layer. A graph of depth $K$ therefore needs $K+1$ sequential network passes, against roughly $(K+1)\times$NFE for DCM, where NFE is the number of network evaluations per integration.

We also build a diagnostic. For the scalar nodes of triangular monotonic benchmarks, we measure how often the one-step decoder violates monotonicity in the noise: the fraction of noise pairs whose order flips after decoding. The exogenous-isomorphism theory makes monotonicity necessary for identifiability. A one-step map $u \mapsto u - \bar{u}(u;pa)$ has no built-in guarantee of it, so we need to know whether and when it breaks.

Baselines

The naive reference is DCM run as published, with its standard multi-step DDIM integration. It fixes the accuracy ceiling we are trying to keep. The file reports DCM at 0.199% relative error for average treatment effect on IHDP. We also run it with a matched budget of 1, 2 and 4 steps using the same trained network. This is the "do nothing but take bigger steps" option.

The strongest existing competitor for the one-step regime is an analytically invertible conditional normalizing flow per node, as in the bijective generation mechanisms of Counterfactual Identifiability of Bijective Causal Models. Such a flow already gives one-pass abduction with an exact inverse, and it reported 1% MAPE on counterfactual ellipse generation, where non-bijective baselines exceeded 6000%. At the image scale, the strongest competitor is Diffusion Counterfactual Generation with Semantic Abduction. It reports 3–5 minutes per image on an RTX 4090 and more than 95% marker removal on EMBED mammograms. The file gives no numeric intervention accuracy for it on Morpho-MNIST.

On the TM-SCM-SYM and ER-DIAG-50/ER-TRIL-50 suites, TVSM is the reference. No reported $CTF_{RMSE}$ number was found on Lacuna for TVSM, so we rerun it ourselves.

All methods share the same graphs, data splits, a per-node parameter budget matched within 10%, and an equal hyperparameter search budget in GPU-hours. Beating the flow baseline means lower $CTF_{RMSE}$ at equal or lower wall-clock time on vector- or image-valued nodes. Beating Semantic Abduction means matching its intervention accuracy and reconstruction $L_1$ on Morpho-MNIST at under one second per image.

Two ablations isolate the proposal's ingredient. One drops $\mathcal{L}_{\text{cf}}$ entirely, which gives a causal BiFM with only its native consistency term. The other drops only the HSIC term.

Why not just…?

Why not just use a conditional normalizing flow, which is already exactly invertible in one pass? This is the real threat. On low-dimensional tabular nodes we expect flows to tie or win, and we will report that honestly. Our bet is restricted to nodes whose mechanisms are high-dimensional or strongly non-additive. There, the required invertible layer designs limit expressiveness. DCM reports that its node-wise diffusion trains 7x faster than CAREFL and 9x faster than VACA, and it reports markedly lower error under non-additive noise. If the flow matches us on Morpho-MNIST images, the proposal's case at scale fails.

Why not just run DCM with a 4-step solver? BiFM reports that inversion trajectories degrade far more under coarse discretization than generation trajectories. The exogenous-isomorphism analysis predicts that abduction errors compound down the graph. Our NFE sweep measures this directly: $CTF_{RMSE}$ against depth for step-reduced DCM versus our model.

Why not just distill DCM progressively (Progressive Distillation for Fast Sampling of Diffusion Models)? Distillation targets the generation direction and needs a converged teacher plus several rounds. Distilling the inverse requires a second student with no independence guarantee. We include a distilled encoder-decoder pair as a baseline. If it matches our model once $\mathcal{L}_{\text{cf}}$ is added to it, then the contribution is the objective rather than the MeanFlow parameterization, and we will say so.

Why not a conditional VAE? Its posterior is not a unique abduction. The ellipse results above show how badly non-bijective mechanisms fail, so we include one only as a sanity row.

Proposed Experiments and Decisive Checks

The experiments span TM-SCM-SYM (barbell, stair, fork and backdoor graphs), ER-DIAG-50 and ER-TRIL-50, IHDP and the Sachs network from the DCM evaluation, and Morpho-MNIST with thickness and intensity as parents of an image node. Metrics are $CTF_{RMSE}$, $OBS_{WD}$, ATE relative error, intervention accuracy and reconstruction $L_1$ for images, the monotonicity-violation rate, and sequential network passes and wall-clock time per query.

The first decisive check is the accuracy–latency curve. Our 1- and 2-step model should sit within a small margin of multi-step DCM on $CTF_{RMSE}$ while step-reduced DCM degrades with depth.

The second check tests the central claim directly. Across checkpoints and ablations, round-trip error plus HSIC should predict $CTF_{RMSE}$ substantially better than $OBS_{WD}$ does. If it does not, the bound is not a usable training signal, and the proposal is wrong in its core.

The third check isolates the ingredient. Removing $\mathcal{L}_{\text{cf}}$ should leave $OBS_{WD}$ nearly unchanged while $CTF_{RMSE}$ and monotonicity violations rise.