acceptodds
Withdrawn from ICLR 2027

EICR: Executable Identifiability Counterfactual Reasoning for Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Counterfactual reasoning requires more than executing an alternative action. A model must first infer which hidden explanations remain compatible with the factual observation, and then determine whether those explanations agree after intervention. We identify a failure mode in which a model selects one factually consistent explanation, obtains a locally valid counterfactual value, and reports it as a uniquely supported answer even though another equally valid explanation gives a different result. We call this phenomenon counterfactual identifiability collapse. We introduce EICR, an executable completion-set reasoning framework that represents a counterfactual answer by all observation-consistent structural completions, their branch-wise rollouts, the resulting answer set, and an instance-level identification label. We construct EIBench as a unified benchmark with an oracle-driven construction protocol over a common symbolic task space rendered as executable code, structural causal models (SCMs/DAGs), and natural language. It distinguishes point-identified, set-identified, interval-partially-identified, and unresolved cases. Under an explicit and complete candidate mechanism family, we establish soundness and completeness properties for the oracle certificate, define a training objective for learning task-dependent certificates, specify representation-aware dataset splits, and introduce metrics for false certainty, false uncertainty, set coverage, interval coverage, bound error, and certificate validity. Experiments across five model backbones and three representations show that completion-set supervision reduces false certainty on executable tasks and improves recovery of the supported answer set, while natural-language parsing remains more challenging than structured representations. These results support preserving the evidence-compatible completion set rather than committing prematurely to a single abductive explanation.

Primary area:
causal reasoning
Keywords:
counterfactual reasoning, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_svGPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces EICR, a framework for counterfactual reasoning that keeps all compatible causal explanations. An oracle generates structured certificates containing factual checks, counterfactual outcomes, answer sets or bounds, and identification labels, which are used for supervised fine-tuning. The authors also introduce EIBench, a synthetic benchmark with code, structural causal model, and natural-language representations. Experiments report improvements in answer recovery and reductions in false certainty over several baselines.

Critical strengths
  1. The distinction between hidden-state ambiguity and target ambiguity is useful: several compatible explanations can still imply the same counterfactual answer.
  2. The Answer-Set SFT baseline is relevant because it helps separate supervision of the final answer and identification label from supervision of the complete supporting trace.
Critical weaknesses
  1. The conceptual and methodological contribution of this paper appears to be very limited. Once a complete candidate mechanism family and admissible completion space are supplied, retaining the explanations consistent with the observation and comparing their intervention outcomes is a direct solution to the stated problem. The training objective is standard supervised fine-tuning on oracle traces, and the theoretical results largely restate the consequences of exhaustive enumeration under exact execution. The coupon example illustrates familiar underdetermination, but does not establish why the proposed certificate framework is needed or what substantial new capability it provides beyond existing executable counterfactual reasoning.
  2. The benchmark provides insufficient evidence for the broader value of the approach. A synthetic benchmark is appropriate for controlled evaluation, but here the tasks, oracle, certificate format, and supervision are all built around the proposed procedure. Improvements on this benchmark only proves performance within that construction. The parallel code, SCM, and natural-language versions share the same underlying tasks, so they do not provide independent evidence of broader task generalization. Evaluation on an independently constructed task set or clearly specified structural extrapolation would make the contribution more convincing.
  3. The oracle and data-generation protocol are not sufficiently justified or specified. The authors introduce sampling mechanisms and executing completions, but leave the mechanism grammar, variable and domain sizes, completion-space distributions, template counts, and split sizes unclear. For bounded tasks, the paper refers to symbolic interval procedures without specifying the function restrictions or solver guarantees needed to establish sharp bounds. The soundness and completeness propositions assume a complete candidate space and correct execution, and they do not validate how these assumptions are ensured in the benchmark. The construction of unresolved cases also needs clarification, including whether incompleteness is explicitly signaled in the input.
  4. The certificate-validity results appear inconsistent with the stated definition. Table 4 reports 84.7 certificate validity but only 79.6 identification accuracy for the full model. A valid certificate is described as including the correct identification label, so certificate validity should not exceed identification accuracy under the same evaluation population and averaging.
  5. The training comparisons and ablations are difficult to assess. Concrete optimization settings, training steps, decoding budgets, self-consistency sample counts, and exact checkpoint identifiers are missing. It is also unclear how removing certificate components changes the training targets and evaluation: scoring an intentionally omitted field against the full certificate could make part of the ablation degradation automatic.
Questions for authors

See weakness.

Reviewer_U7SfOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies counterfactual queries whose abduction step is underdetermined: several completions of a declared mechanism family reproduce the factual observation but give different target values under the intervention. Reporting one such branch value as the unique answer is called counterfactual identifiability collapse (Eq. 5). EICR defines the answer as the set of target values over all observation-consistent completions (Eq. 3), labels it point-identified, set-identified, interval-partially-identified or unresolved, and has an executable oracle build a certificate (completions, factual checks, branch rollouts, result, label, completeness status) that an LLM is trained to emit with SFT (Eq. 6). EIBench has 8,000 synthetic items per representation in five families, rendered as Code, SCM/DAG and natural language from one canonical record and split by canonical task and template before rendering. Props. 1-6 and Cor. 1 state soundness, completeness and related set-theoretic properties, mostly of the oracle and its certificates under a complete declared candidate space. On five 2.5B-14B backbones, EICR-SFT beats CoT, CoT-SC, an executable-counterfactual baseline and Answer-Set SFT on PointAcc, FCR, SetCov and Tightness in every cell of Table 3, with NL hardest. Table 2 tracks EICR-SFT's FCR on hidden-divergent items per backbone, and the appendix adds an ablation, seed stability and a per-family breakdown for Qwen2.5-14B (Tables 4-6).

Critical strengths

A precise, scoreable notion of when a counterfactual answer is warranted. Eq. 3 (l.183-184) and Eq. 5 (l.216-222) separate a branch value that is correct under one valid explanation from an answer the evidence supports uniquely, and the four labels turn this into an instance-level output designed to penalize both false certainty and blanket hedging: the hidden-irrelevant family has multiple valid completions but an invariant target (l.200-201, l.267-268), FCR is paired with FUR (l.373-374), and App. F states that declaring every task unresolved "is not solving the identification problem" (l.791-792). The scope is stated honestly, with guarantees only "under an explicit and complete candidate mechanism family" (l.024-025), Prop. 6 on the dependence on the declared space, and a caution against reading synthetic results as deployment evidence (l.496-499). The idea is classical non-identifiability applied per instance and is sharpest for the enumerable families, so it can be counted as a modest but useful lens for evaluating LLM counterfactual reasoning.

The benchmark construction and the first control are set up the right way. As described, the oracle certificate is computed from one canonical record before any rendering, and canonical task IDs and generator templates are split before rendering, so no task or complete template crosses partitions (l.318-321, l.354-358; App. C.2, l.732-735); this closes the main leakage route for multi-format benchmarks and makes the Code/SCM/NL comparison content-matched. Answer-Set SFT, supervised on the label and result payload but not on provenance (l.368-371), is the necessary first control, though not a length-matched one (Q9), and EICR-SFT beats it in all 60 backbone-representation-metric cells of Table 3: FCR is 3.2-10.3 points lower (23-32% relative, mean 27.7%), SetCov is 4.8-11.8 points higher, and Answer-Set SFT's Tightness exceeds its SetCov by 4.9-8.8 points against 1.4-2.3 for EICR-SFT, which is the pattern expected if provenance supervision recovers branches that answer-only training drops.

Critical weaknesses

The certificate-validity and per-family results contradict the paper's own metric definitions. Sec. 3.2 says a certificate is valid only when its soundness and completeness conditions hold, including that "result sets or interval bounds must match the oracle" (l.207-210). App. B calls a certificate that omits any valid completion incomplete (l.711-715), and D.2 gives incomplete certificates zero validity (l.755-757). So on an enumerable item a valid certificate has R^=R\hat R = R and Coverage =1=1 (Eq. 7), and mean CertValid cannot exceed mean SetCov on the same items. App. B also requires "a correct unresolved label" for unresolved validity (l.716-717). Table 6 (p.17) breaks both conditions: CertValid exceeds SetCov in all three families its caption treats as enumerable-result families (91.4 vs 88.7, 84.1 vs 78.5, 81.9 vs 75.9), and unresolved CertValid is 75.7 against IdentAcc 71.5. The aggregates also fail checks that need no assumption about the test composition. CertValid minus IdentAcc is 2.8/1.8/3.7/3.8/4.2 across the five families of Table 6 but 84.7−79.6=5.184.7-79.6=5.1 for the same model in Table 4, which no common family weighting can produce even with rounding. Table 4's SetCov (79.4) together with Table 6's bounded and unresolved CertValid (80.6, 75.7) caps aggregate CertValid at 80.6 under any common weighting (79.2 with Table 1's proportions), not the 84.7 reported in Tables 4-5. Table 3 has no internal contradiction of this kind, but Table 4's full row equals its Qwen2.5-14B EICR-SFT averages exactly (83.9/14.4/79.4/80.9) and Table 5 repeats those cells next to the CertValid values, so the main table appears to come from the same experiments and scorer. The PDF has no code or data link (the Reproducibility Statement mentions only an unspecified "accompanying implementation", l.508-509), no raw outputs and no benchmark item beyond the coupon toy, and one backbone ("Qwen2.5-2.5B", l.363) matches no Qwen2.5 size I know of (0.5B/1.5B/3B/7B/14B/32B/72B), so I cannot tell whether the definitions, the scoring code or the tables are wrong. This matters beyond the appendix: Tables 4-6 are the only evidence on certificate validity, on the bounded and unresolved families taken separately, and on which certificate components matter, which is what separates EICR from answer-set prediction, and the same problems lower my trust in Table 3. Resolving it needs corrected tables that satisfy the stated definitions, per-family test counts and the aggregation formula, and per-seed predictions with the scoring code. Table 5 is one more reason to want per-seed values: 11 of its 12 SDs follow 0.5 + 0.1 per column + 0.2 per row (Code 0.5/0.6/0.7/0.9, SCM 0.7/0.8/0.9/1.0, NL 0.9/1.0/1.1/1.2; only Code CertValid is off, by 0.1). A three-seed SD has a relative standard error of about 50%, so a grid this exact is unlikely unless seed effects are almost perfectly shared across representations and across metrics scored on different items (PointAcc on point items, FCR on non-point items, l.372-373)-- per-seed values would settle it.

The named failure mode is never measured as defined, and the interval and unresolved labels are evaluated only thinly. Collapse (Eq. 5, l.216-222) is an in-set value y^∈R\hat y \in R with ∣R∣>1|R|>1 presented as unique, which the paper calls "distinct from ordinary hallucination and numerical execution error" (l.079-080). But FCR only counts point labels on non-point items (l.373), the label is scored separately from the payload (App. D.1, l.750-751), and Table 3 pools bounded and unresolved items into FCR (l.397-399), so no reported number separates in-set point answers (y^∈R\hat y \in R) from other false-certain outputs. "Experiment 1: Simple Test of Single-Point Collapse" evaluates only EICR-SFT, and the paper concedes that it "isolates backbone sensitivity ... rather than comparing different supervision targets" (l.413-414); no compared method is built to follow the single-completion procedure of Eq. 4, and the closest evidence, the otherwise undescribed "w/o completion enumeration" ablation that App. E reads as "enumeration prevents collapse" (l.779-780), uses the same pooled FCR on one backbone. The reported numbers do not show in-set collapse in the baselines either. An in-set singleton on a set-identified item scores Tightness 1 and Coverage at most 1/2 (Eq. 7), yet CoT, the most false-certain method in all 15 cells (FCR 32.1-50.1), has Tightness only 1.2-2.2 points above SetCov, about the same as EICR-SFT (1.4-2.3); only Answer-Set SFT (4.9-8.8) shows that signature. CoT's false-certain answers may instead lie outside RR, be offset by over-broad sets on other items, or sit mostly in bounded and unresolved items, and the reported numbers cannot separate these cases. For bounded and unresolved items, Table 6 gives IdentAcc 76.9 and 71.5, FCR 12.6 and 18.9, and CertValid 80.6 and 75.7 for Qwen2.5-14B EICR-SFT (App. B ties validity there to matching endpoints and a supported reason, l.715-717). But interval coverage, bound error (Eq. 8), exact set match and reason validity, promised at l.394-396, listed among the "secondary analyses" at l.425-428 and invoked in Table 6's caption (l.866-868), never appear (I searched "IntervalCoverage", "interval coverage", "bound error", "exact set", "reason validity"), and no baseline or other backbone has any IdentAcc, FUR or CertValid. So, beyond the pooled FCR, the paper never shows that EICR-SFT beats Answer-Set SFT or Executable CF at producing intervals or recognizing unresolved cases, and the Conclusion's "EICR therefore reports a point, a set, an interval, or an unresolved result" (l.484-485) describes the oracle rather than the trained model, whose four-way IdentAcc is 79.6% on one backbone (Table 4). Since the paper describes its closest prior work as already evaluating "multiple valid counterfactual answers" (l.115-117), these labels carry much of the added value, and they are asserted more than demonstrated. Most of this can be fixed from existing outputs: split false-certain outputs into y^∈R\hat y\in R and y^∉R\hat y\notin R for every method, add the baselines and a point-answer SFT model to Table 2, and report the promised interval and reason metrics plus FUR and IdentAcc for all methods.

Questions for authors
  1. Do Tables 3-6 share models, seeds and test sets, how are SetCov, IdentAcc and CertValid computed and aggregated, and how does "w/o completion enumeration" get CertValid 36.9 if D.2 zeroes incomplete certificates?
  2. Which checkpoints ("Qwen2.5-2.5B" matches no Qwen2.5 size I know of) and tuning were used, are CoT/CoT-SC fine-tuned, how do the baselines emit labels and sets, and what are per-method parse-failure rates?
  3. How long are EICR-SFT versus Answer-Set SFT targets ("matched budgets", l.429), and does a length-matched trace without completion lists, or a single-completion trace, keep EICR-SFT's gain?
  4. For each method, what share of false-certain outputs on hidden-divergent items has y^∈R\hat{y}\in R versus y^∉R\hat{y}\notin R (CoT's Tightness exceeds SetCov by only 1.2-2.2 points), and what are predicted set sizes?
  5. How are bounded gold intervals computed and incompleteness signaled in unresolved inputs, and is any cross-representation or parsing result behind the transfer (l.095-096) and NL-parsing (l.030-031) claims?
  6. Reference fixes: "Pearl et al. (2003)", Econometric Theory 19, is L. G. Neuberg's review of Pearl's book (Econometric Theory 19(4):675-685); cite Pearl, Causality, Cambridge University Press, 2000 (2nd ed. 2009). Jin et al. (2024): remove David Ha (not an author), add Spencer Poff. "Youssef et al. (2024)" → Nguyen, Youssef, Seifert and Schlötterer (Findings of EMNLP 2024). Hüyük et al. (2025): "Javier Hernandez" → Javier González. Wei et al. (2022): add Brian Ichter, drop the "et al." after the last author. Kaushik et al. (2019) → ICLR 2020; Wang et al. (2022) → ICLR 2023; Chu et al. (2025) and Vashishtha et al. (2024) → ICML 2025. Trailing "et al." → Schölkopf in Jin et al. (2023), Daya Guo in Shao et al. (2024), Xu Yang in Wu et al. (2026).
Reviewer_HYiAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to build training data and benchmarks that teach LLMs to return all counterfactual outcomes consistent with the factual evidence, rather than selecting one plausible outcome and presenting it as certain (called "counterfactual identifiability collapse", Eq. 5). In structural causal model terms, this counterfactual inference process amounts to identifying the compatible exogenous states (in some tasks also parameters or causal mechanisms, from a candidate set given in the input), applying the intervention while preserving each state, and collecting the resulting outcomes. The authors propose EICR, a training framework built around certificates, which are structured, verifiable reasoning traces of this process generated by an executable oracle.

The paper also introduces EIBench, with training and evaluation tasks expressed in code, in structural equations with graphs, and in natural language. EIBench contains not only the questions and answers but also the certificates, which can be used directly for SFT to teach the model this reasoning trace. Experiments are conducted on five backbones, comparing EICR-SFT with prompting baselines (CoT, CoT-SC) and an SFT baselines on the proposed benchmark.

Critical strengths

S1. The paper connects traditional counterfactual reasoning with LLM reasoning in a focused but useful way. It is an important direction to make the community aware of why correct counterfactual inference matters: a model should not present one locally valid branch as the unique answer when the evidence supports several.

S2. The benchmark covers tasks in three important representations: code, SCMs (structural equations with graphs), and natural language.

S3. The results comparing Answer-Set SFT and EICR-SFT clearly suggest the importance of the reasoning trace. Both are trained on the same tasks and differ only in whether the certificate is supervised, and EICR-SFT is better on every backbone.

Critical weaknesses

Presentation

W1. The notation is confusing and easy to misread for readers from the causal inference literature. The paper uses MM for the "task-level candidate mechanism family" and McM_c for the "instantiated mechanism" (Sec. 3.1, ll. 171–174). In the traditional literature (e.g., Pearl, 2009), the causal mechanisms are usually denoted by FF, and MM denotes the SCM itself, M=⟨U,V,F,P(U)⟩M = \langle U, V, F, P(U) \rangle, not its mechanisms. Moreover, a completion cc may fix hidden variable values, parameters, or causal relations (ll. 172–174), so McM_c is really a fully specified SCM together with an exogenous realization, not a mechanism. I suggest adding a section in the appendix that connects the formulation in Sec. 3.1 with the traditional causal inference literature, i.e., stating how completions, Agree\mathrm{Agree}, Exec\mathrm{Exec}, and R(O,I,Y)R(O, I, Y) correspond to abduction, action, and prediction in an SCM (Pearl, 2009).

W2. I also encourage the authors to illustrate this connection with at least one clear example, e.g., the coupon task (Fig. 1) as it is actually rendered in Code, SCM and NL, together with its full serialized certificate τ\tau.

Experiments

W3. Implementation details are missing for both the baselines and the proposed method. There is no learning rate, number of epochs, full fine-tuning vs. LoRA choice, decoding settings, or prompts for CoT and CoT-SC, and App. C.1 says all methods share "the same prompts, sample counts, context budget, and optimizer settings", which leaves unclear whether CoT and CoT-SC are fine-tuned. The model names also do not match released checkpoints: "Qwen2.5-2.5B" seems cannot be found in Qwen2.5 family.

W4. The evaluation is only on EIBench. Since EICR fine-tunes the model on a narrow synthetic task, it should also be evaluated on general benchmarks (e.g., MMLU) to test whether the fine-tuning causes forgetting of general abilities.

Dataset

W5. The queries and graphs are limited and not reported clearly. The target query appears always to be of the form YxY_x given the factual observation OO (Eq. 3), e.g., P(Yx∣x′,y′)P(Y_x \mid x', y') in the coupon example, where YY is the outcome, XX is the intervention, and the facts contain YY. The paper does not report the graphs, the number of variables, or which variables are observed when constructing datasets. Counterfactual reasoning can be very different when the query and the causal diagram vary (Shpitser & Pearl, 2008; Correa, Lee & Bareinboim, 2021). Since Eq. 3 allows a single intervention do(I)do(I) and a single target, nested counterfactuals such as the natural direct effect, or joint queries across worlds such as the probability of necessity and sufficiency, cannot be expressed. If the paper wants to claim general counterfactual reasoning, the benchmark should cover more diverse graphs and queries (e.g., ETT, NDE), not only queries like P(Yx∣x′,y′)P(Y_x \mid x', y'). Otherwise, the paper should state this limitation.

Questions for authors

Q1. What prompts are used for CoT and CoT-SC? Are there results on in-context learning, e.g., giving the model Pearl's abduction-action-prediction steps and a few worked examples? If in-context learning can already solve the benchmark, SFT may not be needed.

Q2. Is the admissible completion space C(M)C(M) given to the model as part of the input? Sec. 1 (ll. 81–82) and Sec. 3.1 (ll. 177–178) suggest that the task supplies it explicitly. If so, why not let the LLM propose the candidate explanations itself, and how do the models perform in that setting?

Q3. Why does the formulation not involve the distribution over UU? Are there cases in LLM applications where not only the support of the counterfactual but also its distribution matters (e.g., in the coupon example, the probability that the purchase would not have happened without the coupon)?

Q4. I could not find the results for IntervalCoverage and BoundError (Eq. 8) on the bounded tasks on the unresolved tasks. Could the authors point me to them, or these erros are only defined but not evaluated?

Q5. Out of curiosity (and not affecting my evaluation, given possible resource limitations): is there any intuition on how closed-source models perform on EIBench?

I am willing to raise my score if the authors address the weaknesses and questions above.

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

The paper proposes Executable Identifiability Counterfactual Reasoning (EICR). This framework certifies when counterfactual reasoning is point-identified, set-identified, interval-partially identified, or unresolved to support certainty. It creates EIBench to measure performance on matched code, known SCMs, and natural language.

Critical strengths
  1. Distinguishing counterfactual outcomes/predictions from the support for the reasoning is useful seems like a reasonable research direction.
Critical weaknesses
  1. The novelty of the paper is weak and relies heavily on existing work from Vashishtha et al's (2026) work on 'Executable counterfactuals: Improving llms’ causal reasoning through code.' This paper cites that previous work but taking a closer look at Vashishtha et al. (2026) shows they already evaluates and enumarates the full answer set for a counterfactual reasoning task and discusses hidden variable causes and identifiability issues. The EICR seems to create an implementation layer for explicit certificates but this doesn't seem to be particularly novel and not well distinguished from the predeccessor.
  2. The theoretical contributions of the paper are also weak and most results follow immediately from the construction. Eg. Corollary 1 just describes a singleton set. Proposition 1 and 2 follow immediately from definitions while 4 and 5 seem to say the same things. Overall, it appears that several concepts are dressed up and take away from clarity that simpler and common names would provide. Eg. Recall and precision are named coverage and tightness.
  3. Some benefits are overclaimed or not demonstrated. Most significant of which is 'This [EICR] makes it possible to distinguish representation transfer from causal-structure transfer.' This is never shown in the paper nor does it follow from what is shown. The paper also says 'Identification accuracy, FUR, set exact match, interval coverage, bound error, incompleteness-reason validity, and certificate validity are reported as secondary analyses because they diagnose different failure modes and would make the multi-backbone table unnecessarily wide.' but no such metrics are to be found and thus leaves the bounded and unresolved certificates unevaluated. The oracle guarantees also do not transfer to the SFT-trained model.
Questions for authors

I am curious how the authors see this work advancing on Vashishtha et al. (2026). Also there appears to be a discrepancy in Table 6. Under the definitions, I would expect CertValidity < IdentAcc for the Unresolved case but I may be misunderstanding.

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

Related papers

Discussion (0)

Sign in to comment.