- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2(+1)
Read review
Paper summary
The paper investigates whether, and under what conditions, the predictive parents identified by a transformer's gradient-based readout correspond to the structural causal parents of a target variable. The paper uses synthetic time-series data to evaluate the claim. The experiments show that the transformer can learn predictive dependencies corresponding to structural causal parents when the identification assumptions hold. The results on real-world benchmarks are provided in the appendix and provide inconclusive evidence for the main claim of the paper.
Critical strengths
The paper demonstrates that Transformers can achieve strong performance in identifying the causal parents of a target variable under certain assumptions. The proposed method outperforms several baselines on simulated datasets and achieves competitive performance on real-world-derived datasets. Although the Transformer architecture may not be the main factor contributing to these improvements, the paper provides an interesting investigation into the relationship between the predictive capabilities of ML models and their ability to identify causal parents from learned conditional distributions.
Critical weaknesses
Table 1 (in appendix) shows Transformer vs MLP comparison in causal discovery performance. It appears that the causal learning mechanism is not specific to transformers as MLP can use the same gradient readout to achieve better causal parent identification on AQI dataset. This may indicate transformer architecture is not the main factor for the performance improvement. Other strong conditional density estimators may achieve comparable performance using the same gradient readout.
In general, I feel the title and central framing of the paper may be misleading. Although the experiments demonstrate that transformer-based models can identify causal parents under certain assumptions, the theoretical result is not specific to transformers or the proposed gradient readout. Under the identifying assumptions, causal parents can be recovered through conditional dependence testing implemented using various nonlinear ML methods with or without gradient-based attribution.
Questions for authors
see weaknesses
5 comments
W1: Transformer vs. MLP and Other Conditional Density Estimators Using the Same Gradient Readout ("MLP can use the same gradient readout to achieve better causal parent identification on AQI")
To compare conditional density estimators without sparsity penalties (), we trained four architectures sharing the same mixture head, NLL objective, and gradient readout across synthetic (), released and expanded CausalTime (–, expanded via the same CausalTime data generation process), and physical Causal Chambers wind tunnel ( real sensor steps [5]) benchmarks:
Table 1: Cross-Architecture Comparison Under Identical Mixture Head & Gradient Readout (All Pure NLL, )
| Benchmark (All Pure NLL, ) | Metric | DOT (Transformer) | CNN | GRU | MLP |
|---|---|---|---|---|---|
| Synthetic Scale / Shape () | AP | 1.000 / 1.000 | 1.000 / 1.000 | 1.000 / 1.000 | 1.000 / 1.000 |
| Synthetic Mixed () | AP | 0.998 | 0.922 | 0.906 | 0.855 |
| CausalTime @ (AQI / Traf / Med) | AUROC | 0.689 / 0.643 / 0.643 | 0.866 / 0.673 / 0.626 | 0.819 / 0.663 / 0.582 | 0.840 / 0.696 / 0.506 |
| Expanded CausalTime @ (AQI / Traf / Med) | AUROC | 0.977 / 0.883 / 0.813 | 0.974 / 0.775 / 0.632 | 0.966 / 0.759 / 0.615 | 0.911 / 0.751 / 0.555 |
| Expanded CausalTime @ (AQI / Traf / Med) | AUROC | 0.991 / 0.909 / 0.851 | 0.985 / 0.807 / 0.656 | 0.971 / 0.790 / 0.630 | 0.939 / 0.763 / 0.592 |
| Expanded CausalTime @ (AQI / Traf / Med) | AUROC | 0.997 / 0.944 / 0.871 | 0.993 / 0.852 / 0.700 | 0.985 / 0.795 / 0.656 | 0.953 / 0.795 / 0.623 |
| Physical Wind Tunnel ( real obs.) | AP / AUROC | 0.526 / 0.762 | 0.367 / 0.608 | 0.357 / 0.630 | 0.451 / 0.771 |
Note: On the Physical Wind Tunnel (chance ), classical baselines PCMCI, VAR-LiNGAM, and DYNOTEARS reach 0.229, 0.157, and 0.123 AP (0.748, 0.636, and 0.512 AUROC).
We see that at steps, simpler models have lower estimation variance on AQI and Traffic, so CNN and MLP score higher there. From steps onward, the ranking reverses on AQI and Traffic as well, and DOT leads all four architectures across all three CausalTime domains, multi-mechanism Mixed, and Wind Tunnel AP.
W2 (Part 2): Conditional Dependence Testing Without Gradient-Based Attribution (CMIknn and GPDC)
We also compared DOT against non-gradient nonlinear conditional independence tests (Tigramite PCMCI with CMIknn and GPDC) across all five nonlinear mechanisms at . CMIknn and GPDC are more sample-efficient at , whereas DOT overtakes CMIknn on 3 of 5 mechanisms by and all 5 mechanisms by while running significantly faster ( min vs. 3.5–9.5 hours per run), avoiding the scalability bottleneck of nonparametric CI tests at larger :
Table 2: Nonparametric PCMCI (CMIknn, GPDC) vs. DOT Across Sample Sizes and Runtime
| Mechanism | : CMIknn / GPDC / DOT | : CMIknn / DOT | : CMIknn / DOT | : DOT |
|---|---|---|---|---|
| Interaction | 0.347 / 0.314 / 0.132 | 0.602 / 0.473 | 0.710 / 0.790 | 1.000 |
| Scale | 0.704 / 0.954 / 0.344 | 0.912 / 0.972 | 0.966 / 1.000 | 1.000 |
| Shape | 0.603 / 0.560 / 0.131 | 0.767 / 0.896 | 0.863 / 0.980 | 1.000 |
| Transport | 0.331 / 0.574 / 0.125 | 0.469 / 0.802 | 0.652 / 0.982 | 1.000 |
| Mixed | 0.511 / 0.740 / 0.116 | 0.732 / 0.703 | 0.824 / 0.909 | 0.998 |
| Runtime / run | 12–45m / 3.3–4.5h / 0.2–0.3m | 1.2–3.4h / 0.5–0.8m | 3.5–9.5h / 0.8–1.3m | 8–10m |
We hope these clarifications and experiments thoroughly address your concerns!
References
- [1] Rohekar, Raanan Y., Yaniv Gurwicz, and Shami Nisimov. "Causal interpretation of self-attention in pre-trained transformers." Advances in Neural Information Processing Systems 36 (2023): 31450-31465.
- [2] Nichani, Eshaan, Alex Damian, and Jason D. Lee. "How transformers learn causal structure with gradient descent." arXiv preprint arXiv:2402.14735 (2024).
- [3] Vafa, Keyon, et al. "Evaluating the world model implicit in a generative model." Advances in Neural Information Processing Systems 37 (2024): 26941-26975.
- [4] Ansari, Abdul Fatir, et al. "Chronos-2: From univariate to universal forecasting." arXiv preprint arXiv:2510.15821 (2025).
- [5] Gamella, Juan L., Jonas Peters, and Peter Bühlmann. "Causal chambers as a real-world physical testbed for AI methodology." Nature Machine Intelligence 7.1 (2025): 107-118.
Dear Reviewer uxfh,
We want to kindly follow up regarding our rebuttal. We truly appreciate the time and effort you have already dedicated to reviewing our work and providing your feedback.
We have worked to address your concerns as follows:
Title and Central Framing (W2): We agree with you that our identifiability results are architecture-agnostic, since they characterize what any next-step predictor can recover from the observed history. We centered the paper on Transformers because whether next-step Transformers learn causal structure is still unsettled in the literature, and our results clarify when they do (the log-density gradient reaches 0.998 AP on the same checkpoint where attention rollout drops to 0.321 AP) and when they cannot. We also showed that the same readout extends to the pretrained Chronos-2 (a time-series foundation model). On reflection, since you and another reviewer independently regard the title as a misleading claim, we are happy to revise our framing in the revision and change to an architecture-agnostic title in the final version, since ICLR does not allow title changes during the discussion period. We are currently considering candidates such as "Autoregressive Models Are Limited Causal Learners" and "When Predictive Parents Are Structural Parents: Causal Structure from Next-Step Likelihood" (suggested by Reviewer qpPt).
Transformer vs. MLP and Other Density Estimators (W1): We trained Transformer, CNN, GRU, and MLP predictors with the same mixture head, the same pure likelihood objective, and the same gradient readout. We found that MLP and CNN lead on AQI and Traffic only at the released sample size, where simpler models have lower estimation variance. Once we expand CausalTime with its official data-generating process, the ranking reverses, and the Transformer becomes the best of the four architectures on all three domains at every sample size, as well as on the multi-mechanism Mixed system and in AP on the real wind tunnel data.
Conditional Dependence Testing Without Gradients (W2): We compared DOT against nonparametric PCMCI with CMIknn and GPDC on all five nonlinear mechanisms. These tests do recover the parents and are more sample-efficient at small sample sizes, whereas DOT overtakes CMIknn on all five mechanisms as the sample size grows, while taking about a minute instead of 3.5–9.5 hours per run.
You can find the full tables and details in our responses above. We would greatly appreciate it if you could let us know whether our responses have addressed your concerns. Thank you!
Best,
The Authors
We thank the reviewer for recognizing our investigation linking learned predictive conditional distributions to structural causal parent identification. Below we directly address both critical weaknesses (W2 on title, central framing, and non-gradient conditional dependence testing; and W1 on Transformer vs. MLP and other conditional density estimators).
W2 (Part 1): Title and Central Framing ("The theoretical result is not specific to transformers or the proposed gradient readout")
We agree that Theorem 1 and Proposition 2 are architecture-agnostic, as they characterize the conditional distribution p∗(Xi,t∣Xt−L:t−1) targeted by next-step prediction. We center our framing and title on Autoregressive Transformers for two reasons:
qpPt) further supports this link, reaching 0.764 zero-shot and 0.940 fine-tuned AP on Linear, and 0.814 fine-tuned AP on variance-only Scale.Following ICLR policy, we cannot modify the title during the rebuttal period. If you still view the Transformer-centered framing as a main weakness after considering these additional experiment results, we will consider changing our title and update our framing to be architecture-agnostic (e.g., "Autoregressive Models Are Limited Causal Learners") in the final version.