- Rating
- 2
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper makes a clear distinction between counterfactual text editing and associational text generation, and constructs a benchmark to test whether LLMs can perform causally valid counterfactual editing. It introduces CoBe, which contains 2,000 samples, each consisting of a short scenario and conversational prompts.The data covers seven domains and six types of causal graph structures. Experiments show that even frontier models often fail to produce causally valid counterfactual edits.
Critical strengths
- This paper makes a necessary distinction between counterfactual text editing and associational generation, and the task is properly grounded in Pearl’s causal hierarchy.
- Well conducted experiments, with a comparative study of nine LLMs and detailed analyses across evaluation criteria, graph types, and query phrasings.
- The dataset covers seven domains and six graph types. Unlike prior work that mostly only focus on simple linear timelines, it actually tests complexer structures like colliders, forks, and diamonds.
Critical weaknesses
- The paper is a bit unbalanced. The abstract and introduction take 3.5 pages because the intro includes too much background, like the Pearl Causal Hierarchy. This should really be moved to a separate section.
- The comparison between the LLM-Judge and human annotators doesn't seem fair. Humans get the original context, query, and LLM output, but not the evaluation criteria (App. D.4.1). Meanwhile, the LLM-Judge only gets the output and the criteria (App. D.2). Can the authors get human annotations for the three criteria checks in lines 292–316, compare them with the LLM-Judge results, and report Krippendorff’s for each one?
- Only one LLM-Judge is used. All responses are judged by Claude-Opus-4.7 (l.319), and it is also one of the evaluated models. This may introduce self-preference bias [1]. Could the authors include more frontier models as LLM judges?
- The authors state that “models frequently default to associational generation rather than preserving the causal structure” (l.28-39). However, the experiments only measure three checks, and none of them directly tests for associational generation. Could the authors add a direct test? For example, the same question could be asked in two forms: “what would have happened if...” and “what happened if...” If the model gives similar answers to both forms, this would support the claim.
[1] Panickssery et al. LLM Evaluators Recognize and Favor Their Own Generations. NeurIPS 2024.
Questions for authors
Please see weaknesses.
3 comments
Author Response [2/2]
(W4) The evaluation criteria don’t test associational bias. Could the authors add an associational prompt test?
Associational reasoning:
Actually, our eval criterion E1 does test failures due to associational reasoning. This criterion checks whether models update prior causes, background premises, or other facts which should remain invariant under counterfactual edit. E.g., in Fig. 3 when changing the treatment Gemini also edits the patient symptoms. Frontier models exhibit this error ~30% of the time (red bars in Fig. 6). Such non-causal updates are failures related to associational reasoning (App. A.2) because the edit affects merely-correlated events.
Associational prompting:
We follow your suggestion and directly compare GPT-5.4-Pro by asking the same question under counterfactual (“would have happened…had”) and associational (“happened…if”) query forms on the same 500 samples, holding the stories, judge, and evaluation criteria fixed (App. K). Model accuracy is similar: 62.1% (counterfactual) versus 61.1% (associational).
Beyond aggregate accuracy, responses under both query versions receive matching judge verdicts on 80.9% of paired examples overall and 88.2% for E1 eval criterion, which specifically checks for changes to invariant facts (associational reasoning failures). Failures on E1 criterion across each query form are also similar at 27.2% vs 28%, and 80% of counterfactual-prompting E1 failures recur under the associational-prompting form.
These results are consistent with associational bias toward updating facts in violation of causal rules.
(a) Accuracy under each query form
| Query form | Accuracy |
|---|---|
| Counterfactual: “would have happened…had” | 62.1% |
| Associational: “happened…if” | 61.1% |
(b) E1 failure rate under each query form
| Query form | E1 failure rate |
|---|---|
| Counterfactual: “would have happened…had” | 27.2% |
| Associational: “happened…if” | 28.0% |
(c) Match rates between query forms
| Evaluation check | Match rate |
|---|---|
| Overall judge verdict | 80.9% |
| E1: preservation of invariant facts | 88.2% |
Lastly, we highlight the discussion with reviewer M9Pf extensively discussing the novelty and rigor of our benchmark over prior works that go by “counterfactual editing” but are not rigorously grounded in Pearl causal semantics.
Please let us know if this addresses your questions. We are happy to receive any follow up clarifications/suggestions.
Thank you!
Dear reviewer XYva,
We wanted to enquire if you had a chance to read our response. In particular, our revision incorporates your feedback: (1) shortened Introduction; clearly marked optional background, (2) human-agreement study on individual criteria (95-97% agreement), (3) experiments using different judge model, showing no model-family-bias. We also (4) ran the associational prompt experiment you requested, and documented evidence for associational bias.
Please let us know if there are any remaining concerns, and we would be glad to address these.
Thank you for your time and consideration!
Sincerely, Authors of submission 50487
Author Response [1/2]
We thank reviewer XYva for their very helpful feedback, and have incorporated several of their suggestions into our latest draft (changes in blue).
We especially appreciate the acknowledgment of the formal causal grounding of our benchmark, the diversity of domains and graphs compared to prior work, and the rigor of the prompt ablation experiments and error analysis.
(W1) Introduction is long
This is useful feedback, and we have now reduced Sec. 1 to 3 pages and shifted some content to later.
We still retain some background about the Pearl Causal Hierarchy because a general audience may lack your causal expertise. They may not understand clearly why our evaluation is novel vs. earlier benchmarks (e.g. TimeTravel, PASTA discussed in the other threads) which use the word “counterfactual” but are actually not properly grounded in causal rules for editing.
We want to strongly clarify the difference between counterfactual and associational text edits. We have added a note at Line 048 that readers familiar with PCH can skip this part. Please let us know if that helps balance out the paper.
(W2) Comparison between the Judge and human experts doesn't seem fair. Can the authors compare human annotations for the three criteria checks with the Judge?
We clarify an important point here.
In our human agreement study (App. D.4), the LLM-judge receives model output and the self-contained eval criteria which contain the facts which should hold, or be changed in the output, plus the eval prompt in App. D.2. This setting is how we evaluate models in our benchmark. The human experts receive the original story, intervention, and model output, plus instructions of how to perform the counterfactual edit based on formal causal rules (App D.4.1). This is our unbiased ground truth for what counts as a correct evaluation.
Indeed, we deliberately don’t show the judge-facing eval criteria to the human annotators because we want to compare our actual eval pipeline to ground truth correctness according to the causal rules (92.1% agreement). The judge-facing instructions are internal to the judge pipeline. We don’t want to bias the annotators based on what the judge sees.
To address your concern, we added App. J with newly conducted independent expert annotations on the judge-facing criteria checks themselves.
Per-criteria agreement of the judge with experts is 95-97%. The Krippendorff’s α for E2, E0 is misleading because of class imbalance in labels, with such skewed labels, α attributes nearly all agreement to chance and penalizes each disagreement on a rare failure heavily, exhibiting wide intervals, as reported in App. J. Thus, we additionally report Gwet’s AC1 which is high across all criteria.
(W3) Only one LLM judge is used. This may introduce self-preference bias.
We tested our eval pipeline also using GPT-5.6-Sol as a judge (n=500) and discussed results in App. I.
We see no evidence of model-family-bias. Claude-judge widens the GPT–Claude score advantage, compared to GPT-judge. GPT-judge reduces the GPT–Claude score advantage, compared to Claude-judge. The scores are also roughly similar. The judges differ mainly in calibration: GPT-5.6-Sol is slightly stricter on the “unchanged facts” criterion and slightly more lenient on the “downstream facts changed” criterion, for both models alike.
(Minor correction: our submission had a typo at Line 367. Our evals used Opus-5 as a judge, not Opus-4.7; we corrected this now)