Prediction Bottlenecks Don't Discover Causal Structure (But Here's What They Actually Do)
Summary
This paper challenges the claim that prediction bottlenecks in models like Mamba recover causal structure, demonstrating through a new benchmark that gains are largely due to confounds and robustness artifacts rather than true causal discovery.
View Cached Full Text
Cached at: 05/13/26, 04:13 AM
Paper page - Prediction Bottlenecks Don’t Discover Causal Structure (But Here’s What They Actually Do)
Source: https://huggingface.co/papers/2605.09169 This paper falsifies the claim that next-step prediction bottlenecks—especially Mamba/SSM weight projections—recover causal structure, showing instead that their apparent gains are mostly low-rank regression, sample-size confounds, intervention-semantics artifacts, and target-corruption robustness, with the main durable contribution being a reusable falsification benchmark.
➡️ 𝐊𝐞𝐲 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬 𝐨𝐟 𝐭𝐡𝐞𝐢𝐫 𝐏𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧-𝐚𝐬-𝐂𝐚𝐮𝐬𝐚𝐥-𝐃𝐢𝐬𝐜𝐨𝐯𝐞𝐫𝐲 𝐅𝐚𝐥𝐬𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧 𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤:
🧪 𝑹𝒆𝒖𝒔𝒂𝒃𝒍𝒆 𝑭𝒊𝒗𝒆-𝑺𝒕𝒂𝒈𝒆 𝑭𝒂𝒍𝒔𝒊𝒇𝒊𝒄𝒂𝒕𝒊𝒐𝒏 𝑩𝒆𝒏𝒄𝒉𝒎𝒂𝒓𝒌: Introduces a control-heavy benchmark spanning VAR, Lorenz-96, CauseMe-style generators, real datasets with edge-provenance cards, matched-capacity architectures, size-matched observational controls, and multiple intervention semantics to stress-test claims that prediction models implicitly recover causal graphs.
🧩 𝑾𝒆𝒊𝒈𝒉𝒕-𝑷𝒓𝒐𝒋𝒆𝒄𝒕𝒊𝒐𝒏 𝑪𝒂𝒖𝒔𝒂𝒍𝒊𝒕𝒚 𝑫𝒐𝒆𝒔 𝑵𝒐𝒕 𝑺𝒖𝒓𝒗𝒊𝒗𝒆 𝑪𝒐𝒏𝒕𝒓𝒐𝒍𝒔: Tests the extraction rule (S = |W_{out}W_{in}|) for bottleneck predictors and shows that linear bottlenecks match or beat Mamba SSMs, tuned Lasso dominates on synthetic graph recovery, and classical PCMCI/Granger-style methods outperform the bottleneck on clean Lorenz-96 ground truth.
🧠 𝑰𝒏𝒕𝒆𝒓𝒗𝒆𝒏𝒕𝒊𝒐𝒏 𝑮𝒂𝒊𝒏𝒔 𝑨𝒓𝒆 𝑪𝒐𝒏𝒇𝒐𝒖𝒏𝒅𝒔, 𝑵𝒐𝒕 𝑪𝒂𝒖𝒔𝒂𝒍 𝑬𝒙𝒕𝒓𝒂𝒄𝒕𝒊𝒐𝒏: Demonstrates that the reported interventional advantage mostly comes from extra sample size and a non-standard per-step random-forcing intervention; under proper (do(X_i=c)) interventions the effect nearly vanishes, while the residual appears even more strongly in classical bivariate Granger, indicating method-agnostic target-corruption robustness rather than learned causal discovery.
Similar Articles
How Does Bayesian Causal Discovery Fail? Characterising Structural Consequences in Linear Gaussian Networks under Latent Confounding
This paper analyzes how Bayesian causal discovery fails under latent confounding in linear Gaussian networks, deriving a correlation threshold that causes spurious edges to be favored by the score function and characterizing two distinct posterior failure regimes.
Interpretable Causal Discovery via Causal-Effect Constraints
This paper introduces a Bayesian approach to conditional causal discovery, where the posterior over causal graphs and parameters is conditioned on user-specified causal-effect constraints (e.g., a large causal effect). They adapt rare-event estimation techniques to handle events with small posterior mass and validate the method on synthetic data and the Sachs protein dataset.
When Prediction Error Is Not Enough: Evaluating Nuisance-Function Prediction for Causal Estimation
The paper evaluates how prediction error of nuisance functions relates to causal estimator performance, finding that prediction error is not a consistent measure of causal bias or confidence interval coverage across methods like XGBoost and Double Machine Learning.
From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation
This paper presents a decision-centric causal optimization framework for large-scale targeting and recommendation, combining a causal Transformer, Bayesian bandit layer, and dual-based linear programming. It reports a statistically significant +7.20% lift in LinkedIn Feed marketing traffic via online A/B testing.
Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models
This paper analyzes latent reasoning models (LRMs) and demonstrates that observable patterns in latent states are not causal explanations of reasoning; it advocates for matched controls and causal tests in interpretability research.