Source-linked AI summary
Prediction Bottlenecks Don't Discover Causal Structure (But Here's What They Actually Do)
Ankit Hemant Lade, Sai Krishna Jasti, Indar Kumar, Aman Chadha
TL;DR
The paper asks whether a forecaster trained only for next-step prediction can reveal causal structure as a byproduct. It packages a five-stage falsification protocol covering architectures, synthetic and real benchmarks, intervention semantics, and matched controls, and finds that the method-level claim does not survive, leaving narrow characterization findings and a reusable benchmark.
Problem
The paper examines whether next-step prediction can expose causal or dependency structure through learned weights, potentially making every pretrained forecaster a causal-discovery engine.
Method
The paper tests the prediction-bottleneck claim with standardized generators, intervention semantics, real-data edge-provenance cards, matched controls, and the readout S = |WoutWin|.
Results
The method-level claim does not survive: a linear bottleneck matches or beats the SSM, tuned Lasso wins on synthetic CauseMe-style benchmarks, and causal baselines lead on Lorenz-96.
Takeaways & Limitations
The lasting contribution is a reusable falsification benchmark, while the surviving bottleneck findings are narrow characterization claims rather than a causal-discovery method.
Takeaways & Limitations
The mild-nonlinearity advantage is reported at one (K, T) configuration and is not established as a general regime.
Abstract
from arXiv · showhide
A Mamba state-space model trained only for next-step prediction appears to recover Granger-causal structure through a simple readout $S = |W_{out} W_{in}|$, with early experiments suggesting the phenomenon generalized across architectures and benefited from interventional data at $p < 10^{-5}$. We package the protocol used to test that claim -- standardized synthetic generators (VAR/Lorenz/CauseMe-style), three intervention semantics ($do(X=c)$, soft-noise, random-forcing), edge-provenance cards on three real datasets, and size-matched control arms -- as a reusable falsification benchmark, and walk the claim through it in five stages. The method-level claim does not survive: (i) a plain linear bottleneck does as well or better; (ii) tuned Lasso beats the bottleneck on synthetic CauseMe-style benchmarks, and on Lorenz-96 (the only real benchmark with unambiguous ground truth) classical PCMCI and Granger lead a tight cluster in which the bottleneck trails; (iii) the headline intervention advantage is roughly 60% a sample-size confound, and the residual disappears under standard $do(X=c)$ interventions, surviving only under a non-standard random-forcing scheme; (iv) even that residual reproduces, with a larger effect, in classical bivariate Granger -- the effect is method-agnostic. What survives is a narrow characterization result; the benchmark is the lasting artifact, and each stage above is one of its control arms.
1 Introduction
The paper tests whether next-step prediction can expose causal structure as a byproduct and finds that the initial claims are not robust. It instead contributes a reusable falsification benchmark and retains only narrow characterization findings.
- Benchmark contribution: The benchmark packages five control arms for testing prediction-as-causal-discovery claims across synthetic generators, intervention semantics, real datasets, and matched controls.It includes VAR, Lorenz-96, CauseMe-style generators, do(X = c), soft-noise, and random-forcing interventions, edge-provenance cards, fixed seeds, and reproducible tooling.
- Falsification results: A matched-capacity linear bottleneck matches or beats the Mamba state-space model, falsifying architecture specificity.
- Falsification results: Tuned Lasso beats the bottleneck on synthetic CauseMe-style benchmarks, while PCMCI and Granger lead it on Lorenz-96.Lorenz-96 is identified as the only real benchmark with unambiguous ground truth.
- Falsification results: Under size-matched controls and proper do(X = c) interventions, the headline intervention advantage collapses, while a residual reproduces under classical Granger.
- What survives: The surviving findings concern mild nonlinearity, sample efficiency, and target-corruption robustness as characterization rather than method claims.
2 Method: The Extraction We Are Falsifying
The tested method trains a next-step forecaster without a causal-specific loss, then interprets a projection-weight product as a directed dependency-strength matrix. The extracted matrix is normalized and thresholded for graph evaluation.
- Extraction pipeline: The model uses input and output projections trained with next-step MSE and no causal-specific loss.
- Extraction pipeline: The readout S = |WoutWin| assigns S_i,j to the strength of j → i, with diagonal entries zeroed before normalization and thresholding.
- Lagged variant: For explicit lagged inputs, the lagged bottleneck extends the readout to S(τ) = |WoutW(τ)in| and evaluates flat-lag AUROC against ground-truth adjacency.
3 Five Falsifications
The five-stage benchmark tests architecture, sparsity, real-world ground truth, intervention semantics, and method-agnosticity. Across these controls, simple prediction bottlenecks are not consistently competitive, and dataset or intervention design can dominate the apparent result.
- Stress-regime matrix: The stress matrix spans five model classes, six generators, varying K and T, three real datasets, three intervention semantics, and four causal or predictive baselines.
- Sparsity and dimensionality: On the focused stress grid, the bottleneck wins only 12% of graph-recovery runs and 0% of prediction-MSE runs against tuned baselines.
- Sparsity and dimensionality: At K=20 on CauseMe-style processes, Lasso reaches AUROC 0.98 versus 0.73 for the bottleneck.
- Real-world ground truth: On Lorenz-96, classical and modern causal baselines lead by a wide margin, whereas climate rankings are unstable because the ground truth is weak.
- Real-world ground truth: Removing two definitional ENSO↔SOI edges changes the climate ranking, moving Granger from first to last and the SSM bottleneck to first.
- Intervention semantics: The intervention comparison originally mixed intervention content with additional samples, so size-matched controls are required to separate the two effects.
4 What Survives
After five falsifications, the paper retains three narrow characterization findings rather than a causal-discovery method: mild-nonlinearity performance, sample efficiency, and target-corruption robustness.
- Mild-nonlinearity configuration: At mild nonlinearity, the bottleneck beats the best tuned baseline on 87% of nine cells, with mean AUROC improvement +0.121.This holds at (K, T) = (20, 300) with nonlinear=0.3 over 10 seeds.
- Mild-nonlinearity configuration: At stronger nonlinearity, both methods fail equally, while Lasso is preferred at zero nonlinearity.The paper labels the result a configuration rather than a general regime because K and T were not swept.
- Sample efficiency: Additional observational data improve bottleneck AUROC more than Lasso’s by ∼0.07 at K ∈{20, 30} under size-matched controls.The gain is described as modest but useful when collecting more observations is cheap.
- Target-corruption robustness: Shared Wout makes the bottleneck more robust than per-equation Lasso fits to per-step random forcing.The paper treats this as a reliability result that explains the residual intervention effect, not as causal discovery.
- These surviving findings are characterization claims, and the paper does not propose the bottleneck as a causal-discovery method.
5 Lessons
The paper’s lessons emphasize structural controls over seed counts, preregistered intervention semantics, classical causal baselines, and audited ground truth.
- Size-matched controls diagnose sample-size confounds that additional seeds cannot reveal.More seeds tighten noisy measurements but do not replace missing structural controls.
- Intervention semantics should be preregistered because random forcing can produce significance where do(X_i = c) yields a null.
- Intervention studies should include classical causal baselines because Granger and PCMCI may benefit through the same mechanism.
- Soft observational benchmarks require ground-truth audits because definitional couplings can distort rankings when positive edges are scarce.
6 Conclusion
After five stages of controls, the proposed causal-discovery method does not survive, while narrow characterization findings and a reusable falsification benchmark remain.
- The headline effects fail under size-matched controls, depend on a non-standard intervention scheme, or reproduce more strongly in classical causal methods.
- The surviving contribution is a narrow set of characterization findings plus a reusable falsification benchmark scaffold.The released materials include experiments, data loaders, edge-provenance cards, baselines, fixed seeds, dependencies, and reproduction tooling.