acceptodds
Withdrawn from ICLR 2027

LLMs as Proposers: Generative Class-Level Counterfactual Explainer for GNNs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Counterfactual explanations for graph neural networks (GNNs) require finding the minimal perturbations on the input graphs for changing GNNs’ predictions. Nevertheless, the instance-level edits fail to provide a holistic view of the model’s global decision boundaries. To address this challenge, we introduce LAPCE (LLM-as-Proposer for Counterfactual Explanation), a post-hoc class-level framework organized into the stages of propose–verify–select, identifying shared, discriminative subgraphs that systematically govern decision boundaries across a whole class. In the proposal stage, an instruction-conditioned chemical large language model (LLM) generates candidate subgraphs for counterfactual explanation, circumventing the combinatorial search in the discrete space. In the verification stage, each candidate subgraph is matched to a molecule, whose residual is further evaluated: The candidate subgraph is valid only if the residual molecule is valid and the GNN’s predicted label changes. In the selection stage, the candidates are selected and ordered using their verification results to form a compact class-level explanation across semantic-cost thresholds and explanation budgets, while controlling redundancy. Empirical results show that LAPCE outperforms the strongest baseline across four molecular tasks.

Primary area:
interpretability and explainable AI
Keywords:
counterfactual explanations, graph neural networks, explainable AI, large language models

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_zVrEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LAPCE, a class-level counterfactual explanation method for molecular GNNs. A chemical LLM, fine-tuned with SFT and PPO, proposes molecular fragments for deletion. Each fragment is then tested on all molecules of the same class, and a fragment counts as valid for a molecule only if deleting it flips the GNN’s prediction. Finally, a selection step builds one ordered list of fragments that covers as many molecules as possible at low cost. On four molecular datasets, LAPCE achieves higher coverage and lower cost than five baselines.

Critical strengths

1.Novelty. Unlike prior LLM-based work that generates whole counterfactual molecules for single instances, LAPCE uses an LLM to propose reusable deletion fragments for class-level explanation.

2.Verified and reusable explanations. Each fragment is accepted only if deleting it actually flips the GNN’s prediction, and the selected fragments form one ordered list that serves any explanation budget without reselection.

3.Thorough experimental reporting. The paper includes bootstrap confidence intervals, results on five GNN backbones, step-by-step ablations of the proposer and selector, and a detailed breakdown of computational cost.

Critical weaknesses

1.Necessity of the LLM is unclear. The LLM requires 8 to 13 hours of SFT and PPO training per task and is partly fine-tuned on BRICS fragments, yet BRICS is the only non-LLM strategy compared for candidate proposal, so cheaper alternatives such as frequent or randomly sampled substructures should also be tested.

2.Only one LLM family is tested. All experiments use ChemLLM, so it is unclear whether the method works with other LLMs, and without comparing a general-purpose LLM, the claimed benefit of chemical knowledge is not verified.

3.Modest gains at high cost. The gains over RLHEX are modest, while LAPCE spends several hours per task on verification in addition to LLM training, and since no baseline runtime is reported, it is unclear whether such small gains justify the extra cost.

4.Evaluation relies on a single distance measure. All methods are evaluated only with WNode, which LAPCE directly optimizes, while baselines such as RLHEX are trained with other distances, so results under other measures such as graph edit distance should also be reported.

Questions for authors

(1) Can you compare with LLM-GCE, the most closely related LLM-based method, using the same pooling and selection procedure as RLHEX?

Reviewer_cQcMOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a LLM-based framework for class-level counterfactual explanation of GNNs. The method uses a chemical LLM to propose,verify and select. The expriments was conducted on chemical classification datasets demonstrate improvements in counterfactual coverage and intervention cost compared with existing baselines.

Critical strengths

None

Critical weaknesses
  1. The motivation is unclear. The authors identify three challenges in the third paragraph of the Introduction, but it is unclear why existing methods are insufficient to address these challenges or what fundamental limitations the proposed approach overcomes. Class-level counterfactual explanation for GNNs is not new. I didn't see a clear point what this paper try to address compared to the existing works. In addition, the overall presentation is difficult to follow, making it challenging to identify the central technical innovation.

  2. The scope of this work is limited. This is not a general approach for class-level counterfactual explanations. It is very specific to molecular graphs and relies on several domain-specific components, including a pretrained chemical LLM, SMILES-based fragement generation, and chemical-validity verification. The experiments are also constrained to the chemical datasets. This narrow experimental scope limits the generalizability and broader significance of the framework.

  3. Why we need to use LLM to solve this problem? The paper does not convincingly demonstrate why an LLM is necessary compared with simpler fragment-generation approaches. Instead, I feel like using LLM make this solution to be overly complicated, a class-level counterfactual explanation can be simply done by some graph generative model. Considering the computational overhead of LLM, I didn't see a clear benefit that why LLM is used here.

Questions for authors

See weakness above

Reviewer_h6A7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

LAPCE is to provide global counterfactual explanations, leveraging the LLM's knowledge to generate the hypothesis of counterfactual explanations, then verifying and selecting the final global explanations using the target model to explain.

Critical strengths
  1. The way to leverage the LLM for XAI is interesting. Plus, the adaptation part of the LLMs is generally reasonable.
  2. Regarding the evaluation metrics such as coverage and cost that the authors propose, LAPCE outperforms recent global counterfactual explanations.
Critical weaknesses
  1. The overall representation is quite poor, and it is hard to understand the critical information. Especially, the authors decide to move the basic explanations of notations and details of rewards in Eq (5) to Appendix, even though they are crucial to understanding the core novelty. While it is reasonable to move the details to the appendix due to page limitations, omitting even the notation significantly undermines the readability of the paper. Besides Eq (5), the terms overall in the paper are not quite intuitive to understand easily. For example, (a) Source-class Parents in Figure 1 could be replaced with a more conventional term, "input graphs to be explained," if that is its intended meaning. This is just one example of a broader issue, as several other terms throughout the paper are similarly unclear or unnecessarily complicated.
  2. Generally, the explanations and motivation are weak in the step of verifying and selecting the explanations.
  3. Even though the authors leverage LLMs, the interpretation of the found counterfactual explanations is missing in terms of qualitative analysis.
Questions for authors
  1. What is your motivation for choosing the WNode? What advantages does WNode offer over these alternatives in evaluating counterfactual explanations?
  2. How did you decide the thresholds based on Figures 3 and 4?
  3. The evaluation metrics cov and cost equations differ from the current global counterfactual explanations e.g., COMRECGC or GCFExplainer. Can the results in Table 2 be evaluated using the metrics adopted by these existing methods?

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

Related papers

Discussion (0)

Sign in to comment.