{
 "artifacts": null,
 "category": "training",
 "description": "When an intervention variable is sampled from a distribution rather than fixed to a point, train a predictor on the full joint distribution instead of replacing intervention with conditioning on its realized value. This prevents selection bias caused by conditioning on mediators or downstream observations that reveal information about the random intervention.",
 "formulas_latex": [
  "$$P(\\mathcal{Y}\\mid\\mathrm{do}(\\mathcal{X}=x),\\mathcal{M}=m)=P(\\mathcal{Y}\\mid\\mathrm{do}(\\mathcal{X}=x,\\mathcal{M}=m))$$",
  "$$\\mathcal{X}_{\\mathrm{do}}\\sim\\mathcal{N}(0,1)$$",
  "$$p_{\\phi,\\theta}(y\\mid c)=\\int q_{\\theta}(x_{\\mathrm{do}})\\int p_{\\phi}(m\\mid x_{\\mathrm{do}},c)\\,p_{\\phi}(y\\mid x_{\\mathrm{do}},m,c)\\,dm\\,dx_{\\mathrm{do}}$$",
  "$$\\mathcal{L}_{\\mathrm{joint}}(\\phi,\\theta)=-\\sum_{k=1}^{B}\\left[\\log q_{\\theta}(x_{\\mathrm{do}}^{(k)}\\mid c^{(k)})+\\log p_{\\phi}(m^{(k)}\\mid x_{\\mathrm{do}}^{(k)},c^{(k)})+\\log p_{\\phi}(y^{(k)}\\mid x_{\\mathrm{do}}^{(k)},m^{(k)},c^{(k)})\\right]$$"
 ],
 "id": 3146,
 "implementation": "Integrate this into a world model, offline-RL transition model, or causal representation module whenever actions, perturbations, or latent interventions are stochastic. Add three heads sharing an encoder of context \\(c\\): an intervention-density head \\(q_{\\theta}(x_{\\mathrm{do}}\\mid c)\\), a mediator head \\(p_{\\phi}(m\\mid x_{\\mathrm{do}},c)\\), and an outcome head \\(p_{\\phi}(y\\mid x_{\\mathrm{do}},m,c)\\). For observed intervention samples, train the joint negative log-likelihood. For prediction under a desired intervention policy, draw \\(S\\) samples \\(x_{s}\\sim q_{\\theta}\\), then draw \\(m_{s}\\sim p_{\\phi}(m\\mid x_{s},c)\\), and average outcome distributions or samples: \\(\\hat p(y\\mid c)=S^{-1}\\sum_{s}p_{\\phi}(y\\mid x_{s},m_{s},c)\\). Pseudocode: encode \\(c\\); sample or observe \\(x_{\\mathrm{do}}\\); compute mediator and outcome log likelihoods; optimize their sum jointly; at evaluation, Monte Carlo marginalize over the intervention and mediator rather than conditioning on a mediator value alone. The paper contributes the singular-versus-nonsingular distinction; estimate the intervention density empirically with a normalizing-flow or Gaussian head and test calibration using held-out intervention samples. The first cheap experiment should use a synthetic SCM with \\(X_{\\mathrm{do}}\\sim\\mathcal{N}(0,1)\\), \\(M=X_{\\mathrm{do}}+U\\), and \\(Y=M+X_{\\mathrm{do}}+V\\), where \\(U,V\\) are independent Gaussian noises. Compare the proposed joint model against a baseline regressor trained on \\((M,Y)\\) that treats intervention as fixed or ignores it. Evaluate expected outcome under shifted intervention variance and under mediator selection \\(M\\approx m\\). Success is lower interventional negative log likelihood, calibrated uncertainty, and reduced bias in estimated \\(E[Y]\\) under intervention-distribution shift.",
 "math_summary": "The paper explicitly distinguishes a singular intervention, which takes one fixed value, from a stochastic intervention such as \\(\\mathcal{X}_{\\mathrm{do}}\\sim\\mathcal{N}(0,1)\\). It states that the replacement \\(P(\\mathcal{Y}\\mid\\mathrm{do}(\\mathcal{X}=x),\\mathcal{M}=m)=P(\\mathcal{Y}\\mid\\mathrm{do}(\\mathcal{X}=x,\\mathcal{M}=m))\\) is valid only when \\(P(\\mathcal{X}_{\\mathrm{do}})\\) is singular; for a random intervention, conditioning on \\(M=m\\) generally selects a nonrepresentative subset of intervention values. The required target is therefore the joint or appropriately marginalized law \\(P_{\\theta}(Y,M)\\), where \\(\\theta\\) parameterizes the intervention mechanism. For a neural model, let \\(q_{\\theta}(x_{\\mathrm{do}})\\) be the intervention distribution, \\(p_{\\phi}(m\\mid x_{\\mathrm{do}},c)\\) the mediator model, and \\(p_{\\phi}(y\\mid x_{\\mathrm{do}},m,c)\\) the outcome model, with context \\(c\\). The stochastic-intervention prediction is the integral \\(p_{\\phi,\\theta}(y\\mid c)=\\int q_{\\theta}(x_{\\mathrm{do}})\\int p_{\\phi}(m\\mid x_{\\mathrm{do}},c)p_{\\phi}(y\\mid x_{\\mathrm{do}},m,c)\\,dm\\,dx_{\\mathrm{do}}\\), not a predictor trained only on \\(p(y\\mid m,c)\\).",
 "math_tags": [
  "probability",
  "statistics",
  "measure-theory",
  "causal-inference"
 ],
 "ml_areas": [
  "world-model",
  "rl",
  "loss",
  "training",
  "generalization"
 ],
 "paper": {
  "arxiv_id": "2609.03697",
  "arxiv_url": "https://arxiv.org/abs/2609.03697",
  "summary_what_math_gives_to_ml": "The paper supplies a formal way to represent repeated, partially observed structures through families of graph embeddings whose mechanisms are shared and whose embeddings satisfy anchor, rigidity, and freeness constraints. This is directly transferable to neural networks as a principled alternative to treating non-IID temporal or multi-environment data as unrelated examples: aligned occurrences can share a causal transition module while equivariance controls how representations transform across environments. A second transferable asset is the warning that stochastic interventions cannot generally be handled by ordinary conditioning or do-calculus substitutions; the intervention distribution must remain part of the modeled joint law. The most promising experiments are an equivariant shared-mechanism world model and a joint intervention-conditioned predictor evaluated under distribution shift.",
  "title": "Symmetries and Causality: Causal Effect Identification Beyond IID Data",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 4,
  "novelty": 6,
  "usefulness": 7
 },
 "solves": [
  "accuracy",
  "stability",
  "generalization"
 ],
 "title": "Joint Modeling for Stochastic Interventions",
 "url": "https://synthcore.org/idea/3146/joint-modeling-for-stochastic-interventions",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)",
  "verification_axes": {
   "benchmark_mechanism": {
    "confirmed": null,
    "tested": false
   },
   "practical_benchmark": {
    "beats_baseline": null,
    "tested": false
   },
   "toy_mechanism_gate": {
    "confirmed": null,
    "tested": false
   }
  }
 }
}