acceptodds
Under review as a conference paper at ICLR 2027

GLORI: GRAPH-LOCAL ORDERING FROM RESIDUAL INSTABILITY ACROSS ENVIRONMENTS

Abstract

We study causal ordering from multi-environment data when an undirected graph skeleton is available but intervention targets are unknown. Our starting observation is that a stable causal mechanism can leave similar prediction errors across envi- ronments, whereas reverse prediction can change with the input distribution. We introduce GLORI (Graph-Local Ordering from Residual Instability). For each node, GLORI trains one predictor using all environments together, evaluates the same predictor separately in each environment, and measures changes in its environment- wise residual mean. The resulting node instability scores define a causal order. We establish the residual-instability contrast theoretically in a tractable one-root setting and then evaluate whether the same signal remains effective beyond this setting. Across 1,440 synthetic graph tasks spanning diverse graph structures and causal mechanisms, as well as real-world CausalBench data, GLORI consistently demonstrates strong causal-ordering performance while remaining computationally efficient relative to the compared methods.

Primary area:
causal reasoning
Keywords:
causal discovery, causal ordering, multi-environment learning, distribution shifts, invariant mechanisms

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zd8aOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper describes GLORI, a method for causal orientation at a relatively low cost. GLORI uses the neighbors of each causal node to fit a regression model, then calculates the mean signed residual for each environment, then generates scores based on these residual means, and uses the scores to sort the nodes in descending order, then orients the edges in the causal graph.

Critical strengths

1.The paper’s idea is very interesting. It determines the directions of edges in a causal graph by fitting regression models. The method has a low computational cost, has practical value, and is relatively novel.

2.The paper demonstrates the effectiveness of the method through sufficient experiments, and the experiments show that the method’s effectiveness is relatively evident.

3.The paper provides a rigorous theoretical proof of the correctness of the method.

Critical weaknesses
  1. The main problem with the paper lies in the coverage of the method. For the method in the paper, suppose there is a causal relationship X → Y. When using it, if different environments cause X to change while the mechanism generating Y does not change, then there is no particular problem. However, if Y is changed, that is, the environment causes Y to change, then when using X to predict Y, the error will vary with this change, which may cause Y to receive a higher score and be placed earlier, leading to an incorrect direction. The experiments in Table 9 of the paper also reflect this point. I think the paper could discuss this issue further. Of course, I think this is an issue related to the mechanism, and from the perspective of the paper, this flaw does not outweigh its merits.
Questions for authors

Please refer to W1

Reviewer_W3zFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces GLORI, a method for causal ordering from multi-environment data using a known undirected graph skeleton. It estimates causal directions by measuring variations in prediction residual means across environments and ordering variables accordingly. The authors provide theoretical guarantees for a restricted linear causal model and report promising results on synthetic benchmarks and real-world biological datasets.

Critical strengths

The paper addresses an interesting problem in causal discovery by exploiting distributional changes across multiple environments to infer causal ordering. The proposed GLORI method is conceptually simple, computationally efficient, and relatively straightforward to implement. The authors also provide a theoretical analysis under restricted assumptions and conduct experiments on several synthetic configurations and two real-world biological datasets. The reported computational efficiency compared with the selected baselines is a positive aspect of this work.

Critical weaknesses

The paper suffers from several significant limitations that substantially weaken its contribution.

  • The methodological novelty is very limited, as the proposed approach mainly combines established ideas from invariant causal prediction and residual-based causal discovery without introducing a fundamentally new theoretical or algorithmic framework.
  • The theoretical analysis is particularly restrictive, covering only a simple linear, one-root causal structure, with no meaningful guarantees for general DAGs involving multiple parents, colliders, or heterogeneous interventions. Moreover, the central assumption that residual instability provides a reliable causal ordering is insufficiently justified and may fail under realistic distribution shifts. The requirement for a known causal skeleton further limits the practical relevance of the approach and makes the experimental setting considerably easier than real-world causal discovery.
  • The empirical evaluation is also unconvincing, as it relies heavily on synthetic data with ground-truth skeletons and lacks systematic investigations of skeleton errors, latent confounding, and violations of the underlying assumptions. The baseline comparisons are incomplete, particularly given the exclusion of NoGAM from the full-data results due to memory limitations. Furthermore, the real-world improvements are relatively modest, especially on K562, and do not provide sufficiently strong evidence of reliable causal discovery. The performance degradation with alternative predictors also raises concerns about the robustness and generalizability of the proposed method.
Questions for authors

Significant work need to make the paper sound for ICLR.

Reviewer_XyKzOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed GLORI, a method that learns graph causal ordering from residual instability in multi-environments. The method fits a regression for each node using its neighbors and measures the instability of mean residuals across environments. The authors provide a theoretical recovery guarantee for a one-root linear model and evaluate the method on synthetic datasets with known causal graphs and real-world CausalBench data.

Critical strengths
  1. The proposed method is simple and computationally efficient. It requires only one regression per node and avoids expensive graph search procedures.
  2. The experimental results are promising. GLORI outperforms the compared baselines in overall edge-direction accuracy on synthetic datasets and achieves substantially lower computational costs. The authors also provide ablation studies and evaluations on real-world biological datasets.
Critical weaknesses
  1. Table 1 motivates GLORI through the instability of reverse conditional distributions across environments. However, changes in conditional distributions do not necessarily imply changes in mean residuals. Consider X_1=X_e + epsilon_1, X_2=X_e + epsilon_2, where X_e is normal. The variance of X_e changes across environments, while its mean remains zero. Hence the later metric defined on mean difference shown in (12) will fail.
  2. This method doesn’t compare GLORI with R^2 sortability baseline proposed in [1].
  3. The method requires a supplied skeleton and does not explicitly account for unmeasured confounders.

[1]. A Scale-Invariant Sorting Criterion to Find a Causal Order in Additive Noise Models

Questions for authors

In Table 9, GLORI achieves 71.8% accuracy when neither endpoint is directly, but only 46.2% when the target alone is shifted. Could the authors explain this substantial difference?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.