Research map / Large language models / Language, safety and interpretability
Natural Counterfactuals With Necessary Backtracking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes natural counterfactuals that allow minimal necessary backtracking to causal ancestors, keeping counterfactual values within the observed data distribution, and solves this via a feasible intervention optimization over learned non-parametric SCMs.
Problem
Pearl's non-backtracking counterfactuals often require interventions that are too far from the observed data distribution to be feasible, leading to unrealistic counterfactual outcomes and unreliable predictions.
Approach
Replace the direct do-intervention with change(A=a*), realized by interventions on causal ancestors of A, chosen by minimizing L1 perception distance (necessary backtracking) subject to a naturalness constraint that noise CDF values lie in (epsilon, 1-epsilon), optimized via a Lagrangian loss over noise variables of a learned SCM.
Key results
- On Toy 1, natural counterfactuals reduce MAE on n3 from 0.297 to 0.114 for do(n2) (about 61.6% reduction, as stated in the text).
- On Strong-3DIdent, the method reduces MAE for d from 0.100 to 0.058 and for alpha from 0.495 to 0.316 relative to non-backtracking counterfactuals.
- On Morpho-MNIST with V-SCM, epsilon=1e-2 lowers t-MAE for do(t) from 0.336 to 0.139 relative to non-backtracking.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use feasible intervention optimization with necessary backtracking instead of direct nonbacktracking intervention | Counterfactual generation with normalizing flows · Toy 1, change(n2) → n3 | MAE: 0.297 -> 0.114 | Helped |
| Use feasible intervention optimization with necessary backtracking instead of direct nonbacktracking intervention | Counterfactual generation with normalizing flows · Toy 3, change(n2) → n3 | MAE: 0.488 -> 0.127 | Helped |
| Use feasible intervention optimization with necessary backtracking instead of direct nonbacktracking intervention | Counterfactual generation with normalizing flows · Toy 4, change(n1) → n3 | MAE: 0.446 -> 0.443 | No clear effect |
| Use feasible intervention optimization with necessary backtracking instead of direct nonbacktracking intervention | Counterfactual generation with H-SCM · Weak-3DIdent, β | MAE: 0.077 -> 0.036 | Helped |
| Use feasible intervention optimization with necessary backtracking instead of direct nonbacktracking intervention | Counterfactual generation with H-SCM · Strong-3DIdent | MAE | Helped |
| Increase naturalness threshold ε from 10⁻⁴ to 10⁻² (ablation) | Natural counterfactuals with V-SCM · MorphoMNIST, change(t), thickness t | MAE: 0.314 -> 0.139 | Helped |
| Increase naturalness threshold ε from 10⁻⁴ to 10⁻² (ablation) | Natural counterfactuals with V-SCM · MorphoMNIST, change(t), intensity i | MAE: 4.506 -> 4.367 | Helped |
| Increase naturalness threshold ε from 10⁻⁴ to 10⁻² (ablation) | Natural counterfactuals with V-SCM · MorphoMNIST, change(i), thickness t | MAE: 0.171 -> 0.145 | Helped |
| Increase naturalness threshold ε from 10⁻⁴ to 10⁻² (ablation) | Natural counterfactuals with V-SCM · MorphoMNIST, change(i), intensity i | MAE: 4.424 -> 3.959 | Helped |
| Increase naturalness threshold ε from 10⁻⁴ to 10⁻² (ablation) | Natural counterfactuals with H-SCM · MorphoMNIST, change(t), thickness t | MAE: 0.260 -> 0.093 | Helped |
| Increase naturalness threshold ε from 10⁻⁴ to 10⁻² (ablation) | Natural counterfactuals with H-SCM · MorphoMNIST, change(t), intensity i | MAE: 2.495 -> 2.338 | Helped |
| Increase naturalness threshold ε from 10⁻⁴ to 10⁻² (ablation) | Natural counterfactuals with H-SCM · MorphoMNIST, change(i), thickness t | MAE: 0.105 -> 0.083 | Helped |
Surprising findings
- Non-backtracking counterfactuals sometimes produce out-of-distribution points; the authors report that some changes have no feasible intervention at all, so these are excluded from evaluation.
Limitations
- Feasible interventions may not exist for some desired changes, so the constraints can be unsatisfiable.
- Evaluation relies on learned SCMs (normalizing flows, VAEs) whose imperfections affect results.
- Naturalness is defined via the data distribution and the assumed monotone, additive-noise style structure; identifiability requires smooth strictly monotonic mechanisms.
Experimental setup
Datasets: Toy 1-4 (synthetic, 10,000 train / 10,000 test each) Morpho-MNIST 3DIdentBOX Weak-3DIdent 3DIdentBOX Strong-3DIdent
Benchmarks: Toy 1-4 Morpho-MNIST Weak-3DIdent Strong-3DIdent
Metrics: mean absolute error counterfactual effectiveness (MAE of predicted parents)
Baselines: non-backtracking counterfactuals (Pearl do-intervention) D-SCM H-SCM
Builds on: Pearl structural counterfactuals Morpho-MNIST counterfactual framework H-SCM D-SCM
Closest papers by meaning
- A New Approach to Backtracking Counterfactual Explanations: A Unified Causal Framework for Efficient Model Interpretability (ICML 2025 · similarity 0.88)
- Counterfactual reasoning: an analysis of in-context emergence (NeurIPS 2025 · similarity 0.86)
- Counterfactual Graphical Models: Constraints and Inference (ICML 2025 · similarity 0.86)
- Counterfactual Realizability (ICLR 2025 · similarity 0.86)
- LeapFactual: Reliable Visual Counterfactual Explanation Using Conditional Flow Matching (NeurIPS 2025 · similarity 0.86)
- Generating Counterfactual Explanations Under Temporal Constraints (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.