acceptodds
Under review as a conference paper at ICLR 2027

Causal Contracts: Constraint-Verified Decoding for Hallucination-Resistant LLMs

Abstract

Large language models generate fluent and confident text, but their training does not explicitly enforce causal validity. This limitation is critical in high-stakes domains, where residual hallucinations often involve incorrect causal or temporal relations rather than missing knowledge. We therefore define hallucination relative to a domain structural causal model (SCM). A claim is considered inconsistent when its associational, interventional, or counterfactual relation is not supported by the corresponding causal structure. We formulate generation as a Lagrangian-relaxed constrained optimization problem and introduce a projected primal-dual procedure. Under standard smoothness assumptions, we establish a non-asymptotic stationarity guarantee. We also derive a PAC-style bound on residual constraint violations as a function of verifier queries. Finally, we define an evaluation protocol spanning clinical QA, scientific explanation, and causal reasoning. The protocol combines causal-consistency verification, human annotation, inter-annotator agreement, and paired significance testing. Preliminary experiments on real LLM outputs demonstrate the feasibility of graph-based causal verification and identify causal claims that conventional factuality measures can miss.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
large language models, hallucination detection, structural causal models, neuro-symbolic reasoning, constrained decoding, causal inference, Lagrangian optimization, clinical natural language processing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_34HbOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a constraint-based generation framework using domain-specific SCMs to mitigate causal hallucinations in LLMs. Unlike existing hallucination mitigation methods that mainly focus on factual correctness or output consistency, the proposed framework verifies causal claims against predefined causal graphs and iteratively revises generated responses based on detected violations. The authors further formulate this process using Lagrangian relaxation and primal-dual optimization, along with a theoretical analysis. The main contribution of this work is extending conventional fact-based hallucination mitigation to causal validity verification and incorporating it into the LLM generation process.

Critical strengths
  1. The paper presents a well-motivated problem formulation focusing on causal hallucinations and clearly identifies a fundamental limitation of current LLMs, which is their inability to fully guarantee the validity of generated causal claims.

  2. Although causality in LLMs has been extensively studied, it is interesting that this work connects causality with hallucinations by treating them as violations of constraints defined by domain-specific SCM.

  3. The paper clearly presents its research objectives and motivations.

  4. The authors provide detailed discussions of failure cases, limitations, and implementation bugs encountered during the experiments, which improves the transparency of the study and helps readers better assess the reliability of the findings.

  5. An interesting aspect of this work is that it connects causal verification to the actual process of revising generated responses, rather than using the verifier solely for post-hoc error detection.

Critical weaknesses
  1. Section 3.1 explains that reference graphs can be constructed from clinical guidelines or knowledge bases, and Appendix A.3 provides the topics and node-edge structures of the graphs used. However, it is unclear what sources and criteria were used to define each relationship and how the validity of these relationships was verified.

  2. The paper lacks concrete examples showing how causal errors in generated responses are identified and corrected through the proposed framework. I think including qualitative examples in the main text would help readers better understand the proposed approach and improve the overall presentation of the paper.

  3. The appendix acknowledges an inconsistency between the step-size assumption and the claimed convergence rate, and states that the theorem's assumptions or conclusions need to be revised. Since this directly affects the correctness of the theoretical contribution, I believe the authors should revise the theorem in Section 3.4 to reflect the corrected assumptions and guarantees, rather than leaving this issue only in the appendix.

  4. It appears that the proposed intervention-effect verification and primal-dual optimization have not been directly implemented or evaluated on actual LLMs. Therefore, the current experiments appear insufficient to demonstrate the practical effectiveness of the full proposed framework, and its generalizability to broader hallucination scenarios remains unclear.

Questions for authors

Although Section 3.2 distinguishes between ASSOC and CAUSE, Section 5.1 appears to apply the same directed-path criterion to both types. However, two variables can be associated through a common cause without having a directed causal path between them. Could the authors clarify how the proposed method avoids incorrectly identifying such valid associations as causal errors?

Reviewer_PM56OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates hallucinations in LLM-generated content that violate causal relationships, arguing that traditional factuality-checking methods often fail to detect such relational errors—a significant limitation in domains like medicine or scientific explanation, where accurate causal reasoning is essential. Consequently, the authors define hallucinations in relation to a domain-specific Structural Causal Model (SCM). However, the actual experiments are limited to the extraction of CAUSE/ASSOC assertions, the verification of directed paths within causal graphs, and the revision of text based on critical feedback. Techniques such as DO-EFFECT, INDEP, continuous relaxation optimization, and primal-dual decoding were not included in the primary experiments.

Critical strengths
  1. Problem localization holds a certain degree of research significance.

  2. Incorporating external causal structures as constraints is of some value.

  3. The authors report negative results in a relatively transparent manner.

Critical weaknesses

1.Excessive mathematical formalism obscures the actual method, and much of it is not implemented

The current abundance of mathematical content does not help answer this question, but rather obscures the differences between methods and experiments.


2.The theoretical guarantees contain unresolved inconsistencies


3.Graph reachability is not sufficient to verify causal correctness

The existence of a directed path does not mean that the specific causal assertion is semantically valid; the absence of a directed path does not mean that the original text must be wrong.


4.The experiments do not demonstrate that the proposed method reduces hallucinations

The results for "ScienceQA + photosynthesis" in Table 3 fail to demonstrate that the proposed causal verifier is more effective at reducing hallucinations than other methods. Furthermore, the FA (hallucination frequency/rate) value of 1.00 is derived from a very small sample size, and the authors provide no robust statistical evidence of significant differences between the baselines.

Questions for authors

same as weaknesses

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

This paper proposes Causal Contracts, a framework that extracts causal claims from LLM outputs, checks them against a domain causal graph, and asks the model to revise unsupported claims. It formulates generation as a constrained optimization problem and provides theoretical convergence and residual-violation bounds. Additionally, experiments on clinical and scientific QA datasets evaluate factual accuracy, hallucination rate, and causal consistency.

Critical strengths

The paper addresses an important limitation of existing hallucination methods by explicitly checking causal claims against domain knowledge. Its graph-based verification framework makes causal errors interpretable and easy to locate, which is valuable for high-stakes domains.

Critical weaknesses
  1. This paper lacks a clear end-to-end framework figure that shows how causal parsing, graph verification, primal-dual optimisation, and constrained decoding interact. Moreover, the only pipeline figure appears in the appendix and covers only the implemented scoring component, excluding several core parts of the proposed method. This makes both the method and the gap between the proposed and evaluated systems difficult to understand.

  2. The framework critically depends on mapping free-form causal concepts to nodes in the reference graph, but this grounding process is not sufficiently validated. Semantically equivalent terms may fail to match, while incorrect mappings may produce false violations. Since unresolved assertions are excluded from CCS, the method can report a high score even when most claims are not checked.

  3. The experiments evaluate only a simplified generate–extract–path-check–revise pipeline rather than the full proposed framework. The DO-effect, d-separation, and primal-dual decoding components are not implemented, while path existence alone cannot fully validate a causal claim. Therefore, the experiments do not sufficiently support the paper’s main methodological claims.

  4. The method relies on small hand-curated causal graphs, but their sources and expert validation are not clearly reported. Moreover, the experiments use only graph nodes and edges rather than the structural equations and probability distributions required by the stated SCM. Therefore, the system mainly checks consistency with a manually constructed graph rather than fully verifying causal validity.

Questions for authors

Please refer to the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.