{"id":"9b0972fc-3ab8-43ee-97f4-29cb5b71f1d3","arxiv_id":"2504.14530","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"A dissertation assembling the author's prior publications on causality for NLP, centered on two LLM causal reasoning benchmarks and their implications.","lead":"This PhD thesis compiles eight peer-reviewed studies on causal reasoning in large language models, spanning benchmarks, interpretability, and computational social science. A reader might open it to see the current evidence on whether LLMs can reason causally, and what causal methods offer NLP tasks.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corr2Cause's MEC construction may conflate graph isomorphism with Markov equivalence, making some validity labels inconsistent with formal causal discovery.","rationale":"The thesis's central claim is that current LLMs lack robust formal causal reasoning, and Corr2Cause is the primary evidence for the 'near random' causal discovery result. My concern is not a dispute about LLM capability; it is an internal correctness question about the benchmark's label generation. The paper defines validity via shared properties of all graphs in the MEC (§2.3.5), but the graph enumeration in §2.3.3 retains one representative per isomorphism class, and §2.3.4 groups graphs by d-separation sets over that reduced set. If MECs are computed over unlabeled representatives rather than over the labeled variable set in each question, the resulting validity labels can disagree with formal causal discovery. The chain/fork example makes this precise: for the same d-separation statement, the full MEC contains graphs where 'A directly causes B' is true and graphs where it is false, so the hypothesis is not entailed. If Corr2Cause labels such samples as valid, a model that correctly follows the formal rules would be marked wrong, and the near-random scores would not support the conclusion. The variable-refactorization robustness result is also weakened, because a fixed topological order in graph enumeration makes alphabetical order a confounded predictive feature. I do not think this invalidates the entire thesis: CLadder provides independent, carefully constructed evidence of weak causal reasoning, and Chapters 4 and 5 do not depend on Corr2Cause. But the Corr2Cause-based portion of the central claim should be treated as unverified until the label-generation code is audited. The reader's conditional verdict is therefore unchanged, though the conditionality should now explicitly include this internal audit.","tokens_in":52531,"tokens_out":21778,"duration_ms":203518,"concrete_test":"Download Corr2Cause v2.0 and run the label-generation script (or inspect the released data) for an N=3 sample whose premise states 'A is independent of C given B; A correlates with B; B correlates with C' and whose hypothesis is 'A directly causes B.' Independently compute the entailment label by enumerating all labeled DAGs with the same d-separation relations (A→B→C, C→B→A, B→A,B→C) or by running PC on the implied conditional-independence oracle and checking whether 'A→B' is in every DAG represented by the output CPDAG. If the dataset label is 1, the ground truth is wrong; if it is 0, replace this concern with a narrower one and re-examine the isomorphism/MEC code path.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Chapter 2's ground-truth labels depend on the Markov equivalence classes (MECs) computed in §2.3.4. The graph enumeration in §2.3.3 keeps only one DAG per isomorphism class (Table 2.1: 6 unique DAGs for N=3), and MECs are then computed by grouping these reduced graphs. But Markov equivalence is not graph isomorphism: for N=3, the chain A→B→C and the fork B→A,B→C have the same d-separation set (A⊥C|B) and belong to the same MEC, yet they are not isomorphic. If the enumeration and MEC clustering are done only on the 6 unlabeled representatives, the fork is never considered when labeling chain-shaped samples. Concretely, for the correlation statement 'A is independent of C given B; A correlates with B; B correlates with C', the hypothesis 'A directly causes B' is not entailed: the MEC contains A→B→C, C→B→A, and B→A,B→C, and in the last graph A does not cause B. A correct label is 0. If Corr2Cause v2.0 labels such samples as 1, the benchmark's 'validity' is not the formal causal discovery answer advertised in §2.3.5. Because the headline claim ('LLMs score near random') is measured against these labels, a systematic labeling bias would mean the scores reflect agreement with an incorrect ground truth rather than absence of causal reasoning. The paper's own ablation (variable refactorization, §2.4.5) is also affected: if the training graphs always use a fixed topological order A<B<C, alphabetical order is a legitimate in-distribution cue, so fine-tuned models' collapse under renaming does not cleanly separate 'memorization' from 'reasoning over an accidentally confounded dataset.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a PhD thesis that compiles peer-reviewed papers on causality for natural language processing. It argues that large language models (LLMs) lack robust formal causal reasoning, based on two new benchmarks: Corr2Cause (Chapter 2), which tests causal discovery from correlational statements, and CLadder (Chapter 3), which tests causal effect reasoning across the three rungs of Pearl's ladder. It then presents studies on mechanistic interpretability of LLMs (Chapter 4), a causal framework for quantifying robustness of mathematical reasoning (Chapter 5), the implications of causal and anticausal learning directions in NLP tasks (Chapters 6-7), and two applications to text-based computational social science (Chapters 8-9). The thesis concludes that current LLMs perform near random on formal causal inference and that their apparent causal ability is partly memorization rather than reasoning.","tokens_in":52849,"tokens_out":24007,"duration_ms":201134,"significance":"If the central claims hold, the thesis provides valuable benchmark resources and a comprehensive framework for evaluating causal reasoning in LLMs. The manuscript's strengths include open-sourced data and code for each chapter, formal grounding of dataset generation in causal inference theory, and explicit robustness checks such as paraphrase and variable-refactorization perturbations (Chapter 2), anti-commonsensical and nonsensical verbalizations (Chapter 3), and adversarial interventions on input factors (Chapter 5). The constituent chapters have been peer-reviewed at major NLP/ML venues, lending credibility to the individual results. The thesis also offers a useful synthesis of causality-for-NLP research directions and applications.","major_comments":[{"comment":"The construction of Corr2Cause conflates graph isomorphism with Markov equivalence, which corrupts the ground-truth labels of the benchmark. In §2.3.3, DAGs are generated by enumerating upper-triangular adjacency matrices (edges only from smaller to larger alphabetical index) and then reduced to one representative per isomorphism class before §2.3.4 computes MECs from the d-separation sets of these representatives. However, Markov equivalence is not preserved by isomorphism-based deduplication: two DAGs in the same MEC need not be isomorphic, and isomorphic DAGs can induce different d-separation statements on a fixed label set. Concretely, for N=3 the true MEC of the chain A→B→C contains the fork B→A,B→C and the reverse chain C→B→A. The fork is isomorphic to the source A→B,A→C and the reverse chain to the chain, so both are deduplicated away. As a result, the correlation set {A independent of C given B; A correlates with B; B correlates with C} is mapped to an 'MEC' containing only the chain, and the hypothesis 'A directly causes B' is labeled valid. In the true MEC over labeled variables, this hypothesis is not entailed, because it is false in the fork and in the reverse chain. Thus the labels in Corr2Cause are not the formal causal-discovery answer advertised in §2.3.5, and the headline near-random results in Table 2.4 may reflect disagreement with an incorrect labeling scheme rather than absence of causal reasoning. The statistics in Table 2.1 (6 DAGs → 5 MECs for N=3, whereas the 25 labeled DAGs on three variables form 9 MECs) signal that the deduplication has collapsed distinct MECs. This is a load-bearing issue for Chapter 2 and for the thesis's central claim; it requires recomputing MECs over the full set of labeled DAGs (or at least before isomorphism deduplication) and regenerating the benchmark and experiments.","section":"§2.3.3-§2.3.5, Table 2.1"}],"minor_comments":[{"comment":"The human sanity check reports 82% accuracy on 50 questions from one expert evaluator; this is a low ceiling and raises the possibility that some questions are ambiguous or that the verbalization does not uniquely encode the symbolic query. Please report inter-annotator agreement and a larger human evaluation to strengthen the benchmark's validity.","section":"§3.3.4"},{"comment":"Many LLM evaluations are reported as point estimates without confidence intervals, and some GPT-3 metrics in Chapter 5 are approximate due to API constraints (Appendix A.4.3, asterisks in figures). Please report variance or significance tests for the key comparisons, such as the CausalCoT improvement over vanilla GPT-4 in Table 3.2, so readers can assess whether differences are meaningful.","section":"Tables 2.4, 3.2, 5.3"},{"comment":"The categorization of NLP tasks into causal, anticausal, and mixed is central to the meta-analysis in Chapter 6, but the assignment relies on the authors' judgment. Please provide more explicit decision rules or inter-annotator agreement for the categorization, and discuss how sensitive the meta-analysis results are to borderline cases (e.g., question answering, machine translation).","section":"§6.2, Table 6.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a PhD thesis composed of previously peer-reviewed papers. The MEC/isomorphism flaw in Chapter 2 is a genuine load-bearing error, but it is fixable by regenerating the dataset with correct MEC computation over labeled DAGs. If the author can do that and re-run the experiments, the thesis's central claim would be substantially strengthened. The editor should also consider whether a thesis compilation is within the journal's scope; my recommendation assumes the venue is willing to consider such manuscripts."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this before you read it: it's a PhD thesis, not a new research paper. It reprints eight peer-reviewed papers (ICLR, NeurIPS, ACL, EMNLP) with introductory framing. No new experiments, but the underlying work is real and influential. Corr2Cause and CLadder are now standard benchmarks for whether LLMs can do formal causal reasoning, and the causal/anticausal learning frame for NLP is genuinely useful.\n\nThe strongest part is the negative result: seventeen off-the-shelf LLMs score near random on Corr2Cause (best F1 33.38), and CLadder shows GPT-4 at 62% with fine-tuned gains collapsing under paraphrasing or variable renaming. The robustness checks (anti-commonsense and nonsensical verbalizations) are well done and rule out simple memorization. Code and data are public, which is a real credit.\n\nThe most serious soft spot is a potential flaw in Corr2Cause's ground-truth labels. The construction enumerates DAGs up to isomorphism and then computes Markov equivalence classes (MECs) on those representatives. But Markov equivalence is not isomorphism: for N=3, the chain A→B→C and the fork B→A,B→C have the same d-separation set and belong to the same MEC, yet they are not isomorphic. If the enumeration keeps only one orientation, the fork is never considered when labeling chain-shaped samples. Then a hypothesis like 'A directly causes B' can be labeled valid when the true MEC includes a graph where A does not cause B. That would be a systematic labeling bias. I haven't checked the released code, but the Table 2.1 numbers (6 DAGs, 5 MECs for N=3) are consistent with this. If it's real, the near-random scores are still bad, but they're measured against an incorrect ground truth, and the fine-tuning collapse findings could be affected too. This deserves a careful check before building on Corr2Cause's exact validity labels.\n\nOther soft spots are minor: several LLM evaluations report point estimates without confidence intervals; Chapter 6's meta-analysis is selective and its task categorization is subjective; GPT-3 numbers are approximate due to API constraints. The thesis is transparent about these.\n\nWho is this for? Anyone using Corr2Cause or CLadder as benchmarks, or working on causality in NLP. It's a convenient entry point, and the limitations sections are honest.\n\nMy recommendation: this deserves serious referee time, not a desk reject. But if it goes to review, the MEC issue should be resolved before the benchmark is fully trusted.","headline":"A thesis compilation of important causality-in-NLP benchmarks; the Corr2Cause MEC construction may have a subtle labeling flaw worth checking.","tokens_in":53378,"tokens_out":19356,"would_cite":true,"duration_ms":159140,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T50","62D20","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The thesis argues that current large language models lack robust formal causal reasoning: they perform near chance on correlation-to-causation inference, stay below 62% on multi-step causal questions, and their fine-tuned gains collapse…","keywords":["causal reasoning","large language models","causal discovery","causal effect estimation","benchmark datasets","chain-of-thought prompting","causal and anticausal learning","mechanistic interpretability"],"falsifier":"A single transformer whose fine-tuned Corr2Cause F1 stays above roughly 90 after full variable refactorization, and whose CLadder accuracy on anti-commonsense and nonsensical story variants rises well above chance, would refute the paper's central diagnosis, because the current results collapse specifically on those perturbations.","tokens_in":52300,"feed_emoji":"🧠","tokens_out":9521,"duration_ms":84838,"temperature":0.7,"pith_summary":"This thesis sets out to determine whether large language models can perform formal causal reasoning rather than recite causal facts memorized from training text. Using newly constructed, symbolically grounded benchmarks, it reports that current models perform close to chance when asked to infer causation from correlation and stay below 62% accuracy on multi-step causal questions, with fine-tuned gains collapsing under paraphrasing or variable renaming. It then uses causal graphs and interventions as an interpretive lens to show where these failures originate inside the models and how the causal direction of data collection shapes task difficulty. A sympathetic reader would take the central claim to be that LLM causal competence is largely memorization-like, and that formal causal machinery is needed to measure, locate, and eventually repair it.","feed_headline":"LLMs score near random at inferring cause from correlation","feed_subtitle":"Two benchmarks show top models under 62 percent on causal questions; fine-tuned gains vanish when variables are renamed.","key_machinery":"The load-bearing machinery is the directed acyclic graph treated as a causal model, paired with d-separation and Markov equivalence to generate correlation statements and validity labels, and with the ladder of causation and do-calculus to generate query-answer pairs for interventions and counterfactuals. The benchmarks convert these graph-theoretic conditions into natural-language verbalizations, and CausalCoT forces a language model through six formal steps: extract the graph, classify the query, formalize it, gather data, derive the estimand, and calculate. In the interpretability chapters the same graph idea is inverted: interventions on inputs, on attention entries, and on textual surface forms quantify direct versus total causal effects, separating desired sensitivity to the ground-truth result from brittleness to spurious factors.","core_discovery":"The paper's central claim is that today's LLMs do not yet possess robust formal causal reasoning. On Corr2Cause, a benchmark built from all legal directed graphs of two to six variables, a causal hypothesis is labeled valid only when it holds across the whole Markov equivalence class of the stated correlations; seventeen off-the-shelf LLMs score near random, with the best F1 at 33.38 for BART MNLI and GPT-4 at 29.08. On CLadder, which generates 10K questions spanning the three rungs of the ladder of causation, GPT-4 reaches 62.03% accuracy, and a purpose-built chain-of-thought prompt, CausalCoT, raises that to 70.40%. Fine-tuning on Corr2Cause can reach 94.74% F1 in-distribution, but this apparent mastery drops by up to 62.30 F1 points under paraphrasing or variable renaming. The thesis also argues, from the independent causal mechanisms principle, that prediction tasks aligned with the causal direction of data collection behave differently from anticausal ones, and it applies causal-effect reasoning to locate where factual and counterfactual mechanisms compete inside transformer layers.","pith_inferences":["The paper does not draw this conclusion, but its perturbation results imply that causal benchmark reports should include a variable-renaming audit as standard practice; any accuracy that does not survive renaming is template matching, not causal reasoning.","A practical corollary the paper leaves implicit: if current LLMs are causal parrots on formal questions, applications that surface causal claims from text should either stay at the level of correlation or route the query through an explicitly represented causal model before asserting an effect.","The causal-versus-anticausal taxonomy could be turned into a data-collection rule: record the order in which X and Y were generated, and choose which side gets annotation effort based on whether semi-supervised learning or domain adaptation is the target use.","The competition-of-mechanisms method suggests a testable extension to safety: apply the same attention modification to prompts that redefine a harmful concept in-context and check whether a few localized entries can suppress memorized associations."],"forward_implications":["Corr2Cause and CLadder provide out-of-distribution tests of formal causal skills: because their questions are generated symbolically, high scores cannot be explained by memorized verbatim examples.","Fine-tuned causal performance should be reported alongside paraphrased and variable-renamed test sets; under those perturbations the thesis observes F1 drops of up to 62 points, so i.i.d. gains alone overstate learning.","Structured causal prompting (CausalCoT) raises GPT-4's CLadder accuracy from 62.03% to 70.40%, with the largest gains on associational questions and diminishing returns toward counterfactual reasoning.","The causal direction of data collection predicts how much semi-supervised learning and domain adaptation help; practitioners can use the causal-versus-anticausal categorization to anticipate when unlabeled data will be effective.","Editing only two or three attention-map entries in specific heads can shift whether a model follows a counterfactual redefinition or recalls a memorized fact, indicating that mechanism-level control is possible without full retraining."],"supporting_citations":[{"why":"Supplies the formal framework of structural causal models and do-calculus that grounds both benchmark generations and the causal analyses.","marker":"(Pearl, 2009b)"},{"why":"Provides the PC algorithm and d-separation machinery on which Corr2Cause's correlation-to-causation labels are built.","marker":"(Spirtes et al., 2001)"},{"why":"Defines the ladder of causation and the causal inference engine that CLadder instantiates in natural language.","marker":"(Pearl and Mackenzie, 2018)"},{"why":"States the independent causal mechanisms principle and its formal implications used in the causal-versus-anticausal analysis.","marker":"(Peters et al., 2017)"},{"why":"Introduces the causal-versus-anticausal learning distinction that the meta-analysis of semi-supervised learning and domain adaptation studies tests.","marker":"(Schölkopf et al., 2012)"},{"why":"Provides chain-of-thought prompting, which CausalCoT extends with causal-inference-specific steps.","marker":"(Wei et al., 2022b)"},{"why":"Supplies causal tracing and the CounterFact data used to construct the factual-versus-counterfactual competition test set.","marker":"(Meng et al., 2022)"},{"why":"Locates factual knowledge recall in late-layer attention and MLP components, the baseline mechanism model that the competition-of-mechanisms study refines.","marker":"(Geva et al., 2023)"}],"fun_headline_variants":["LLMs near random on causal inference benchmark","GPT-4 scores 62% on causal reasoning, then drops","Fine-tuned causal skill erased by variable renaming","Causal reasoning in LLMs: fragile and superficial","New benchmark exposes LLMs' causal blind spot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the benchmark instruments measure formal causal reasoning and not template artifacts: the paper itself flags that Corr2Cause assumes a closed system with no hidden confounders and faithfulness, that CLadder assumes its verbalizations faithfully encode the symbolic causal query, and that the independent-mechanisms principle may not transfer to text data.","fun_headline_variants_meta":{"raw":{"variants":["LLMs near random on causal inference benchmark","GPT-4 scores 62% on causal reasoning, then drops","Fine-tuned causal skill erased by variable renaming","Causal reasoning in LLMs: fragile and superficial","New benchmark exposes LLMs' causal blind spot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1174,"prompt_tokens":905,"completion_tokens":269,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":193}},"tokens_in":521,"tokens_out":269,"duration_ms":3138,"temperature":1.0,"reasoning_tokens":193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:45:54.753510+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A single transformer whose fine-tuned Corr2Cause F1 stays above roughly 90 after full variable refactorization, and whose CLadder accuracy on anti-commonsense and nonsensical story variants rises well above chance, would refute the paper's central diagnosis, because the current results collapse specifically on those perturbations.","supporting_citations":[],"review_version":1}