PaperlayerView exact revision on arXiv

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

arXiv:2605.09169v2cs.LGcs.AI

TL;DR

Whether next-step forecasters reveal causal structure from their learned weights remains an open question. This paper tests that claim with a reusable falsification benchmark and finds that the method-level claim does not survive its controls, leaving only narrow characterization findings.

  • Problem

    It remains an open question whether next-step forecasters’ learned weights expose causal or dependency structure as a byproduct, potentially complementing dedicated causal-discovery methods.

    Source: §1 · p. 1

  • Method

    The paper tests prediction-as-causal-discovery with a reusable five-stage falsification benchmark applied to next-step bottleneck readouts and matched control arms.

    Source: §1 · p. 2, §2 · p. 2

  • Results

    The method-level claim fails: linear bottlenecks match or beat SSMs, Lasso beats them on CauseMe-style benchmarks, and size-matched do(X=c) controls collapse the intervention advantage, whose residual reproduces in classical Granger.

    Source: §1 · pp. 1–2

  • Takeaways & Limitations

    The supported conclusion is a narrow characterization rather than a general causal-discovery method, while the reusable falsification benchmark is the paper’s lasting artifact.

    Source: §6 · p. 5

    The intervention experiments compared the bottleneck only with prediction-fit baselines, not dedicated causal baselines.

    Source: §5 · p. 5

Abstract

from arXiv · show

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 and concludes that the proposed bottleneck is not a competitive causal-discovery method, while preserving a reusable falsification benchmark and limited characterization findings.

1 Introduction

  • The motivating hypothesis was that next-step forecasters might expose causal or dependency structure as a byproduct of prediction, complementing dedicated causal-discovery methods.The paper frames its contribution as primarily negative: testing that hope and identifying controls behind the failures.Source: §1 · p. 1
  • Initial experiments suggested that Mamba’s S = |WoutWin| recovered synthetic VAR Granger graphs and that interventions amplified its advantage at p < 10^-5 across twelve (K, T) configurations.The paper reports that these initial claims proved less robust than they first appeared.Source: §1 · p. 1
  • The paper releases its five-stage protocol as a reusable falsification benchmark spanning standardized synthetic generators, three intervention semantics, and edge-provenance cards for three real datasets.It also includes fixed seeds, a Dockerfile, and a single-command reproduction target.Source: §1 · p. 2
  • Tuned Lasso beats the bottleneck on synthetic CauseMe-style benchmarks, while PCMCI and Granger lead a tight Lorenz-96 cluster in which the bottleneck trails; a linear bottleneck matches or beats the SSM.Lorenz-96 is the only real benchmark with unambiguous ground truth, and the linear comparison uses matched capacity and seeds.Source: §1 · p. 2
  • With size-matched controls, the intervention advantage collapses under proper do(X = c) interventions, while its residual reappears larger in classical Granger.The residual therefore does not establish a bottleneck-specific advantage.Source: §1 · p. 2
  • The findings that survive—mild-nonlinearity, sample-efficiency, and target-corruption robustness—are characterization claims, not evidence for a superior method.These are the paper’s narrow surviving findings.Source: §1 · p. 2

2 Method: The Extraction We Are Falsifying

The method tests whether a model trained only on next-step MSE yields directed edge scores from its input and output projections, without a causal-specific loss. For lagged-input bottlenecks, it extends the extraction across lags and evaluates against ground-truth adjacency.

2 Method: The Extraction We Are Falsifying

  • A next-step-MSE model uses S = |WoutWin| as directed edge strengths (j→i), then zeroes the diagonal, normalizes, and thresholds the score matrix.For lagged-input bottlenecks, S(τ) = |WoutW(τ)in| is evaluated by flat-lag AUROC against ground-truth adjacency.Source: §2 · p. 2

3 Five Falsifications

The bottleneck does not show a general advantage in causal-structure recovery: a linear bottleneck matches or beats the SSM, tuned Lasso wins on synthetic graph recovery, and classical methods lead on clean Lorenz-96. The apparent intervention advantage largely reflects sample size, and its controlled residual persists under random forcing but also appears in classical Granger.

F1. Architecture does not matter

  • Across six synthetic datasets, the matched-capacity linear bottleneck matches or beats the SSM; on Lorenz, linear scores AUROC 0.94 ± 0.09 versus SSM 0.52 ± 0.24.The authors characterize the apparent emergence as low-rank lag-1 regression by the linear bottleneck, with the SSM achieving the same estimator at higher parameter cost.Source: §3 · p. 2

F2. Tuned Lasso beats the bottleneck on graph recovery

  • On the 48-cell stress grid, the bottleneck wins 12% of graph-recovery runs and 0% of prediction-MSE runs; CauseMe-style Lasso reaches AUROC 0.98 at K=20 versus 0.73 for the bottleneck.These results qualify the framing of causal discovery as a byproduct of strong forecasting: on this grid the bottleneck is not a competitive forecaster either.Source: §3 · p. 2, Table 1 · p. 3

F3. Real data: not competitive on the clean benchmark

  • On clean Lorenz-96, PCMCI (0.986), DYNOTEARS (0.983), bivariate Granger (0.979), tuned linear models (0.974), and VAR-LiNGAM (0.968) beat the SSM bottleneck (0.722 ± 0.031).The lagged bottleneck scores 0.916; this is the real benchmark with unambiguous ground truth.Source: Table 2 · p. 3, §3 · p. 3
  • Removing the definitional ENSO↔SOI edges drops climate Granger AUROC from 0.819 to 0.605 and promotes the SSM bottleneck to first, illustrating ranking sensitivity to ground-truth choices.The corrected climate ground truth has only three positive edges over K=6; the authors caution that its ranking is a sensitivity analysis, not a causal-discovery claim.Source: §3 · pp. 3–4, Table 3 · p. 4

F4. Interventions: a confound story

  • Size-matching reduces the bottleneck’s intervention-specific gain to +0.03 to +0.05 AUROC; constant clamps yield only 3/12 significant configurations, while soft interventions yield 6/12.The controlled gain survives only under the original per-step random-forcing scheme, which the paper attributes to robustness to corrupted target rows rather than causal-content extraction.Source: §3 · p. 4

F5. Method-agnosticity (the strongest single result)

  • The size-matched intervention gain is method-agnostic: bivariate Granger matches the bottleneck’s significant gain and exceeds it at K=20 and K=30.The bottleneck gains +0.026 to +0.054 AUROC (p < 10^-3, n=10); Granger gains +0.040 at K=20 and +0.095 at K=30. Lasso uniquely fails under target corruption.Source: §3 · p. 4

4 What Survives

After five falsifications, three narrow characterization results remain: a specific mild-nonlinearity configuration, gains from more observational data, and robustness to target corruption. These results do not support proposing the bottleneck as a causal-discovery method.

Mild-nonlinearity configuration

  • At (K, T) = (20, 300) and nonlinear=0.3, the bottleneck beats tuned Lasso/RRR on 87% of nine cells, with mean AUROC improvement +0.121.At nonlinear=0.6 and 1.0 both methods fail equally; at zero nonlinearity Lasso is preferred, and K or T has not been swept.Source: §4 · pp. 4–5

Sample efficiency

  • Under size-matched control arms, additional observational data improve bottleneck AUROC over Lasso by ∼0.07 at K ∈{20, 30}.The passage characterizes this gain as modest but useful where collecting more observations is cheap.Source: §4 · p. 5

Target-corruption robustness

  • The bottleneck’s shared Wout absorbs per-step random forcing better than Lasso’s per-equation fits, explaining a residual intervention effect as reliability rather than causal discovery.This is presented as a target-corruption robustness result, not evidence of causal discovery.Source: §4 · p. 5

What survives

  • These remaining findings are characterization claims; the bottleneck is not proposed as a causal-discovery method.Source: §4 · p. 5

5 Lessons

The experiments show why causal-discovery claims need structural controls, pre-registered intervention semantics, and classical causal baselines. Benchmark rankings also depend on which positive edges count as ground truth, motivating edge-provenance cards and sensitivity audits.

5 Lessons

  • Both 5-seed Exp 19 and 20-seed Exp 21 (p < 10^-5 across 12 configs) found the same effect, but neither detected the sample-size confound.More seeds tighten a noisy measurement; they cannot diagnose a missing structural control, so size-matched controls are needed.Source: §5 · p. 5
  • The gap between random forcing and proper do(X_i = c) interventions changed the result from significant to null, so intervention semantics should be pre-registered.The distinction was not obvious from the code.Source: §5 · p. 5
  • Intervention experiments need at least one classical causal baseline, because Granger and PCMCI may benefit from the same intervention scheme as neural methods.The experiments compared the bottleneck only with prediction-fit baselines: OLS, Ridge, Lasso, and RRR.Source: §5 · p. 5
  • Removing two definitionally coupled edges flipped Granger from first to last in the climate comparison, motivating edge-provenance cards and leave-one-group-out sensitivity audits.The released benchmark labels positive edges as causal, definitional, proxy, or soft.Source: §5 · p. 5

6 Conclusion

After five stages of controls, the paper finds no free causal-discovery method hidden in prediction bottlenecks. What remains is a narrow set of characterization findings and a reusable falsification benchmark scaffold.

  • The headline effects fail under size-matched controls, depend on a non-standard intervention scheme, or reproduce as larger effects in classical causal methods.Source: §6 · p. 5
  • The surviving contributions are narrow characterization findings and a reusable falsification benchmark scaffold.Source: §6 · p. 5
  • The reproduction package includes experiments, data loaders, edge-provenance cards, optional modern baselines, fixed seeds, pinned dependencies, a Dockerfile, and a single-command reproduction target.Source: §6 · p. 5

Related papers

Loading 2605.09169v2…