Leveraging Large Language Models for Causal Discovery: a

Constraint-based, Argumentation-driven Approach

Zihao Li Fabrizio Russo

zihao.li24@[Link] fabrizio@[Link]

Department of Computing
arXiv:2602.16481v1 [[Link]] 18 Feb 2026

Imperial College London

Abstract records when constructing the causal graphs necessary

to perform causal inference (Cooper and Herskovits,
Causal discovery seeks to uncover causal re- 1992; Heckerman et al., 1995; Meek, 1995).
lations from data, typically represented as
Formulating such background knowledge is labour-
causal graphs, and is essential for predict-
intensive: domain experts must identify relevant vari-
ing the e!ects of interventions. While ex-
ables, specify admissible orientations, and articulate
pert knowledge is required to construct prin-
forbidden ancestral relations (Cooper and Herskovits,
cipled causal graphs, many statistical meth-
1992; Meek, 1995). Existing pipelines either encode
ods have been proposed to leverage obser-
these judgements as hard structural constraints, fix-
vational data with varying formal guaran-
ing or forbidding edges before discovery (Meek, 1995),
tees. Causal Assumption-based Argumenta-
or as Bayesian priors that weight the search over
tion (ABA) is a framework that uses sym-
graph structures (Cooper and Herskovits, 1992; Heck-
bolic reasoning to ensure correspondence be-
erman et al., 1995). On the one hand, hard con-
tween input constraints and output graphs,
straints commit to the specified structure irrespective
while o!ering a principled way to combine
of whether conditional independencies in the data cor-
data and expertise. We explore the use
roborate it, leaving no mechanism to expose conflicts
of large language models (LLMs) as imper-
between expertise and observations. Priors, on the
fect experts for Causal ABA, eliciting se-
other hand, require distributing probability mass over
mantic structural priors from variable names
super-exponential hypothesis spaces and still provide
and descriptions and integrating them with
limited feedback when they conflict with observations.
conditional-independence evidence. Experi-
ments on standard benchmarks and semanti- Causal Assumption-Based Argumentation (Causal
cally grounded synthetic graphs demonstrate ABA) o!ers a principled alternative by encoding can-
state-of-the-art performance, and we addi- didate causal relations as defeasible assumptions and
tionally introduce an evaluation protocol to confronting them with independence constraints de-
mitigate memorisation bias when assessing rived from data (Russo et al., 2024). ABA is a sym-
LLMs for causal discovery. bolic reasoning framework that constructs and evalu-
ates arguments for and against specific claims based
1 Introduction on a set of assumptions and rules (Bondarenko et al.,
1997; Cyras et al., 2017). Causal ABA encodes d-
Understanding the causal structure that governs ob- separation (Pearl, 2009) within this framework to en-
servational and interventional data is central to expla- sure that any DAG inferred from data is consistent
nation, prediction, and decision-making (Pearl, 2009; with (a defendable subset of) the input assumptions,
Peters et al., 2017; Spirtes et al., 2000). Constraint- whether in the form of independencies or structural
and score-based discovery algorithms translate statis- constraints. Additionally, Causal ABA’s rule-based
tical regularities into equivalence classes of directed structure o!ers the possibility to trace each accepted
acyclic graphs (DAGs), but their reliability hinges on or defeated claim back to the assumptions that sup-
sometimes strong assumptions or large sample sizes, port it. This exposes the provenance of inferred edges,
motivating practitioners to complement statistical ev- whether constraints are adopted or not, preserving
idence with mechanistic knowledge and experimental transparency and rigour in the integration.
 Leveraging Large Language Models for Causal Discovery

Large language models (LLMs) provide an attrac- 2 Related Work

tive yet imperfect source of knowledge. They encode
broad semantic and scientific information and can per- We build on three strands of literature: causal discov-
form structured reasoning when prompted appropri- ery from data, the integration of expert knowledge into
ately (Bommasani et al., 2021; Kojima et al., 2022). causal discovery, and the use of LLMs for knowledge
To leverage this information, we rely on semantically elicitation and causal discovery.
meaningful variable metadata (names and optional de-
scriptions rather than anonymised identifiers such as Causal discovery from data. Causal discovery al-
X1 , . . . , Xn ), which provides semantic signal unavail- gorithms aim to reconstruct causal graphs from ob-
able to data-only discovery pipelines. Yet, LLMs can servational and interventional data, often under the
hallucinate plausible-sounding but unsupported causal assumptions of acyclicity and faithfulness: that the
links (Ji et al., 2023), which goes against the rigour causal structure is a DAG and that all and only the
that causal discovery demands. Additionally, their conditional independencies implied by the DAG are
training on web-scale corpora, where many benchmark present in the data (Spirtes et al., 2000). Many meth-
causal discovery graphs are publicly available, makes ods (including ours) additionally assume causal su”-
it di”cult to discern whether a suggested causal de- ciency, i.e., no unobserved confounders (Spirtes et al.,
pendence stems from genuine reasoning or the ver- 2000). A rich literature has proposed various algo-
batim retrieval of a memorised solution. Perturbing rithms with di!erent assumptions, guarantees, and
descriptions beyond what the model has seen often trade-o!s between statistical and computational e”-
leads to sharp degradation in the predicted relations ciency (Glymour et al., 2019; Zanga et al., 2022).
(Carlini et al., 2021; Ji et al., 2023), underscoring the Constraint-based algorithms such as Peter-Clark (PC)
need for evaluation protocols that stress-test out-of- and Fast Causal Inference (Spirtes et al., 2000) infer
distribution generalisation. Markov equivalence classes through conditional inde-
These tensions raise two research questions: how can pendence (CI) tests, whereas score-based procedures
we exploit the rich priors embedded in LLMs without such as Greedy Equivalence Search (Chickering, 2002;
sacrificing rigour and transparency, and how can we Ramsey et al., 2017) trade combinatorial search for
certify that the resulting causal graphs do not merely statistical fit criteria. We use an improved version
reflect memorised benchmarks? of PC, Majority-PC (MPC) (Colombo and Maathuis,
2014), as the statistical engine to derive independence
We address these questions by treating LLMs as defea- constraints from data, but our approach is agnostic to
sible experts whose suggestions are scrutinised through the choice of the underlying discovery algorithm. MPC
argumentation. We make the following contributions: remains a baseline in our empirical evaluation, to-
gether with score-based Fast Greedy Search (Ramsey
• We design a robust LLM integration pipeline et al., 2017, FGS). We also benchmark against two re-
that elicits high-precision constraints on causal cent order-based methods: GRaSP (Lam et al., 2022),
direction (required and forbidden arrows), filters a greedy relaxation of the sparsest permutation crite-
them via a consensus mechanism across multiple rion, and BOSS (Andrews et al., 2023), an order-based
independent queries, and integrates them defea- score-search method that explores grow–shrink neigh-
sibly with data-derived conditional-independence bourhoods. Amongst the score-based category, contin-
evidence within a Causal ABA solver, improv- uous optimisation methods have recently gained trac-
ing tractability via data-driven search optimisa- tion by relaxing the combinatorial search over graphs
tion. into a continuous optimisation problem (Vowels et al.,
2022). A pioneering method is NOTEARS (Zheng
• We develop a novel synthetic evaluation pro- et al., 2018), which uses a smooth characterization of
tocol to mitigate memorisation concerns when acyclicity to enable gradient-based optimisation. We
evaluating LLMs for causal discovery. The proto- use NOTEARS-MLP (Zheng et al., 2020), its more
col creates random benchmarks by grounding ran- advanced, non-linear counterpart, as a baseline. 1
dom DAGs in the CauseNet (Heindorf et al., 2020)
Hybrid approaches tighten these relaxations by
knowledge graph via sub-graph isomorphism and
combining constraint- and score-based approaches
heuristic-guided selection, enabling a robust as-
(Tsamardinos et al., 2006) while logic-based ap-
sessment of LLM generalisation.
1
• We demonstrate empirically that our LLM- We include NOTEARS/NOTEARS-MLP as widely-
used structure-learning baselines from the continuous-
augmented Causal ABA pipeline achieves state- optimisation literature, without claiming causal identifi-
of-the-art performance on both standard bench- ability from observational data alone; see also cautionary
marks and our novel evaluation protocol. discussion in (Reisach et al., 2021).
 Li, Russo

proaches embed logical inference in the process e.g. via causal relations, yet we consolidate the queries into
SAT-based encodings (Hyttinen et al., 2014). Causal a single prompt for improved e”ciency and context.
ABA fits into the logical constraint-based literature Additionally, we adapt the consensus refinement strat-
but uses ABA (Bondarenko et al., 1997) to guarantee egy of (Cohrs et al., 2023) and benchmark against the
that every inferred orientation is backed by an explicit BFS-style LLM-only method of (Jiralerspong et al.,
chain of assumptions and rules (Russo et al., 2024). 2024), which iteratively expands a graph via pairwise
Our work complements these e!orts by integrating a queries.
scalable source of structured assumptions from LLMs.
Despite this promise, LLM responses remain fallible.
Integrating expert knowledge. Prior knowledge Benchmarks reveal gaps between memorised answers
has long been recognised as essential to trim the search and genuine causal reasoning (Jin et al., 2023; Wan
space and enforce domain restrictions in causal discov- et al., 2025; Zečević et al., 2023), and repeated studies
ery (Cooper and Herskovits, 1992; Heckerman et al., caution against using LLMs as sole decision-makers for
1995; Meek, 1995). Conventional approaches require causal discovery (Wu et al., 2025). Hallucinations of
experts to enumerate admissible edges, ancestral rela- LLMs are well documented (Ji et al., 2023) and, when
tions, or forbidden paths before running an algorithm. accepted as constraints, can eliminate true causal re-
More recent e!orts incorporate domain constraints lations (Chen et al., 2023). To mitigate these risks,
during search or as post-processing filters, yet they still we use schema-guided prompting (567-labs, 2025; Liu
rely on hand-crafted inputs and provide limited feed- et al., 2024b) to extract structured semantic con-
back when constraints conflict (Hyttinen et al., 2014). straints, filter them via consensus, and integrate them
By embedding LLM-sourced statements into the argu- with data-derived conditional-independence evidence
mentative machinery of Causal ABA, we retain trans- within Causal ABA; we also introduce robust bench-
parency about why and when constraints are overruled marks to evaluate LLM outputs beyond memorisation.
while creating a causal discovery pipeline that can use
both traditional and unstructured sources of expertise,
while blending them with data-driven insights. 3 Background
Let G = (V, E) be a graph with nodes V and edges
LLMs for causal discovery. LLMs are Transformer E → V ↑ V. Edges can be directed or undirected, the
architectures (Vaswani et al., 2017) pretrained on web- skeleton replaces every direction with an undirected
scale corpora and exhibit strong few-shot generalisa- link. A path path = x1 . . . xn is a sequence of dis-
tion (Brown et al., 2020). Their internalised knowledge tinct adjacent nodes; it is directed if (xi , xi+1 ) ↓ E for
can be elicited through prompting to recover common- all i, and cyclic if additionally x1 = xn . A directed
sense relations from variable-style descriptions (Zhao acyclic graph (DAG) has only directed edges and no
et al., 2023). For causal discovery, this ability enables cycles and will serve as our causal graph formalism.
LLMs to act as experts and translate metadata into A DAG is ascribed a causal semantics by interpreting
candidate structural constraints. nodes as random variables and edges as direct causal
A growing literature leverages LLMs in three comple- e!ects (Pearl, 2009). A Bayesian Network (BN) is a
mentary roles (Wan et al., 2025): (i) querying them pair (G, P) where P is a joint distribution over V that
to orient edges or assemble full causal graphs (Jiraler- satisfies the Markov condition relative to G: every
spong et al., 2024; Kıcıman et al., 2023; Vashishtha variable is independent of its non-descendants given
et al., 2025); (ii) using them as assistants that cri- its parents. For pairwise disjoint sets X, Y, Z → V
tique, refine, or document algorithmic outputs via re- we let (X↔↔Y | Z) indicate that X and Y are inde-
finement loops or natural-language analytics (Abdu- pendent given the conditioning set Z; (X↔↔Y | ↗) is
laal et al., 2024; Gkountouras et al., 2024; Khatibi simply written as (X↔↔Y) and singleton sets {x} are
et al., 2024; Liu et al., 2024a); and (iii) coupling tex- denoted by x (e.g., ({x}↔↔{y} | ↗) is written as (x↔↔y).
tual priors with statistical search through hard (Chen Also, (X ↘↔↔Y | Z) means that X and Y are dependent
et al., 2023) or soft constraints (Ban et al., 2023), or given Z. The link between DAG structure and CI is
even by replacing CI tests with conversational judg- captured by the d-separation criterion (Pearl, 2009).
ments within PC (Cohrs et al., 2023). In our approach, Causal ABA (Russo et al., 2024) is a framework that
MPC provides the CI statements from data, while combines computational argumentation (Dung, 1995;
the consensus-filtered LLM constraints are encoded Toni, 2014) with causal reasoning. An ABA frame-
as additional required/forbidden arrow facts and en- work (Bondarenko et al., 1997, ABAF) is a tuple
forced within the Causal ABA solver only when they ≃L, R, A, ⇐ where L is a formal language, R is a set
do not contradict strong statistical counter-evidence of inference rules of the form head ⇒ body where
(Section 4). As in (Jiralerspong et al., 2024; Kıcıman head ↓ L and body is a finite (possibly empty) set
et al., 2023; Vashishtha et al., 2025), we solicit direct of elements from L, A → L is a non-empty set of
 Leveraging Large Language Models for Causal Discovery

assumptions, and is a total mapping from A into When MPC consume these tests, it returns the par-
L (contrary). In Causal ABA, the language L in- tially oriented DAG depicted above (right). Causal
cludes statements about graphical (causal) relation- ABA ingests the same CI statements, but reasons
ships (with (X, Y ) denoting a directed edge in the about their joint consistency. Each test result is ranked
causal graph), independencies and d-separations; the by credibility, and the framework searches for a sta-
rules R encode the principles of acyclicity and d- ble extension. Including both independence statements
separation, and the assumptions A represent candi- yields no stable extension: the d-separation rules make
date causal relations and/or independencies that can the two statements inconsistent, because E ↔↔ R | O
be accepted or rejected based on the evidence and their can hold together with E ↔↔ R only if R and E are
relations. An assumption a ↓ A attacks an assump- disconnected from the other variables—contradicting
tion b ↓ A if and only if a can be used to derive the many of the other CI tests. After progressive exclu-
contrary of b, i.e., b, using the rules in R. sions, the stable extension contains only the correct in-
dependence statement, yielding the ground-truth DAG.
An ABAF is evaluated using argumentation seman-
tics (Baroni et al., 2011), which determine which sets Having described our statistical and symbolic engines,
of assumptions (called extensions) can be collectively we now turn to the integration of LLM-derived knowl-
accepted based on their ability cohexist (conflict- edge into Causal ABA.
freeness) and defend against external attacks (see
(Russo et al., 2024; Toni, 2014) for formal definitions
and examples). A set of assumptions S → A will pro-
4 LLMs Knowledge for Causal ABA
duce no stable extension if it contains conflicts, i.e.,
Our methodology integrates the semantic knowledge of
if there exists a, b ↓ S such that a attacks b. Causal
LLMs with the symbolic rigour of Causal ABA. The
ABA employs stable semantics to guarantee a one-to-
core idea is to use an LLM as a proxy domain ex-
one correspondence between the accepted assumptions
pert to generate high-precision structural constraints,
and the d-separation relations induced by the output
which are then encoded as facts within the argumenta-
DAG (Russo et al., 2024).
tive framework to prune the vast search space of pos-
A bodyless rule (head ⇒) is called a fact and repre- sible causal graphs. This hybrid approach, depicted
sents an unconditionally true statement in the ABAF in Figure 1, aims to enhance both the e”ciency and
i.e., it is by default included in every extension. In the accuracy of the discovery process. The pipeline
(Russo et al., 2024) ABAPC is proposed as an in- consists of a method to elicit structured constraints
stantiation of Causal ABA using MPC (Colombo and from LLMs (left), and the formal integration of these
Maathuis, 2014), as the statistical engine to derive CI constraints into Causal ABA (right). Human expert
constraints from data. The CI facts from MPC are knowledge is included as optional, and not currently
ranked based on their associated p-values, and the al- part of the experiments, but can be readily integrated.
gorithm iteratively attempts to construct a stable ex-
tension by removing the least credible facts, until a 4.1 Constraint Elicitation Pipeline
stable extension is found. The resulting stable exten-
sion corresponds to a DAG that is consistent with the The quality of LLM-derived knowledge is highly de-
strongest possible (subset of) CI statements. pendent on the elicitation process (Anthropic, 2024).
We propose a robust pipeline to transform unstruc-
Example 3.1 Consider the simple four-variable DAG tured metadata into formal, high-precision constraints.
shown below (left): Education (E) and Race (R) influ-
ence Occupation (O), which in turn influences Income Eliciting Structural Priors. The process begins
(I); Education also directly a!ects Income (E ⇑ I). with a set of variables, each with a name and an op-
tional description. We developed a detailed prompt
True Graph Majority-PC that instructs a primary LLM (gemini-2.5-flash,
E E LLM 1 in Figure 1) to act as a domain expert and
O I O I to focus exclusively on high-precision judgments. The
prompt explicitly asks the LLM to generate two lists:
R R
• Required Directions: Causal relationships
From synthetic data sampled from this ground-truth
(X, Y ) that are required based on logic, tempo-
DAG we obtain 23 CI tests that become fact in Causal
ral precedence, or established principles.
ABA including E ↔↔ R (correct) and E ↔↔ R | O (a
finite-sample artefact). Conditioning on the collider O • Forbidden Directions: Causal relationships
should render E and R dependent, yet PC-style pro- (X, Y ) that are forbidden if they contradict es-
cedures treat both statements as equally trustworthy. tablished knowledge or logical consistency.
 Li, Russo

Figure 1: LLM Integration Pipeline: Given a set of variables with their names and (possibly) descriptions,
we prompt an LLM to generate pairwise causal statements. These statements are then parsed into structured
assumptions for Causal ABA, which combines them with data-derived independence or arrow constraints and
background knowledge to infer a set of causal graphs. Expert knowledge can be injected in the LLM prompts or
as defeasible facts. Variables descriptions and LLM parsing are optional components of the pipeline but enhance
the quality of the generated assumptions. Detailed prompts and parsing rules are provided in Appendix A.

The LLM is instructed to format its response us- spurious constraints (see Appendix A.3 for details).
ing clear headers and a line-by-line structure for each Consensus does not guarantee correctness; it is a prag-
prior, followed by a brief justification. The complete matic variance-reduction step aimed at improving pre-
prompt is provided in Appendix A. cision before constraints are passed to the solver.
Schema-Guided Output Extraction. To ensure
reliability and avoid constraining the LLM’s reason- 4.2 Integration into Causal ABA
ing process with rigid formatting requirements (Tam
We integrate the consensus-filtered LLM constraints
et al., 2024), we decouple the complex reasoning task
as required and forbidden arrow facts, which are en-
of prior generation from the simpler structured output
forced as constraints to further prune the search space.
extraction task. The free text output from the primary
To avoid forcing semantic constraints that directly
LLM is passed to a separate, dedicated process that
contradict strong statistical evidence, we apply them
uses a schema-enforcing tool (567-labs, 2025) to guide
only after the one-o! data-driven skeleton reduction
a second, lightweight LLM (gemini-2.5-flash-lite,
described below, and we discard any required arrow
LLM 3 in Figure 1) in robustly extracting and val-
whose corresponding undirected edge was removed by
idating the constraints. This second LLM is highly
this reduction. The remaining CI statements are han-
e!ective at interpreting the semi-structured text and
dled within Causal ABA as weighted constraints and
handling minor formatting inconsistencies, making the
are progressively relaxed until the program admits a
extraction process significantly more robust than tra-
stable extension (and thus a DAG).
ditional methods like regular expressions. This pro-
duces a structured list of constraints for the next stage. Data-Driven Skeleton Reduction. To improve
the scalability of Causal ABA we perform an initial,
Consensus for High Precision. To mitigate the
one-o! skeleton reduction: edges corresponding to a
stochasticity of LLM outputs and adhere to the high-
high-confidence independence test are removed from
precision requirement of causal discovery, we adapt the
the graph. This follows causal theory and the strat-
majority voting idea from (Cohrs et al., 2023) and in-
egy employed by both PC (Spirtes et al., 2000) and
troduce a consensus-based filtering step. We query the
Causal ABA, with the di!erence that, during the sub-
primary LLM five times independently with the same
sequent iterative relaxation of lower-confidence CI con-
prompt. The final sets of constraints consist only of
straints,2 these edges are not reintroduced. This modi-
the intersection of each of the forbidden and required
fication of Causal ABA avoids the computationally ex-
arrows sets, across all five parsed responses. This con-
2
servative approach ensures that only the most con- This is implemented in clingo (Gebser et al., 2019) by
sistently suggested constraints are passed to Causal changing from a hard constraint to a soft, weighted con-
ABA, significantly reducing the risk of incorporating straint that the solver can either satisfy or violate, remov-
ing its enforcing power without full program regeneration.
 Leveraging Large Language Models for Causal Discovery

pensive re-grounding of the logic program representing Our protocol, illustrated in Figure 2, embeds randomly
the ABAF, making the framework tractable for larger generated graph structures within CauseNet, a large-
problems, at the cost of being unable to fully assess scale knowledge graph of real-world cause-and-e!ect
the relations around these hard constraints. This scal- relationships (Heindorf et al., 2020). The pipeline con-
ability optimisation is applied to both ABAPC and sists of three main stages, as follows:
ABAPC-LLM in all experiments. Overall, ABAPC-
1. Structural Sca!olding. We begin by generat-
LLM inherits the CI-testing cost of PC/MPC but adds
ing a structural “sca!old”. To ensure structural diver-
overhead from solving the ABAF and from a one-o!
sity in our benchmarks, we first define a DAG topol-
set of LLM calls per dataset; empirical runtimes are
ogy using one of three methods: generating Erdős-
reported in Appendix C.3.3.
Rényi (Erdös and Rényi, 1959), Scale-Free (Barabási
LLM Constraints Enforcement. The LLM- and Albert, 1999) graphs, or directly constructing a
derived constraints from the consensus step are then DAG from a randomly populated lower triangular ad-
applied as additional hard constraints, further prun- jacency matrix (details in Appendix B). Once a DAG
ing the search space, while avoiding conflicts with the structure is sampled, we build a full BN by assign-
data-driven skeleton reduction. Specifically: ing randomly initialised conditional probability tables
to each node. This two-step process yields a fully de-
• A required arrow (encoded as (X, Y ) ⇒) is en-
fined BN that allows to sample synthetic observational
forced only if the edge X ⇓ Y was not already
data. We sample CPTs at random purely to generate
removed during the initial skeleton reduction.
data consistent with the sampled structure; since we
This prevents semantic constraints from conflict-
evaluate structural recovery rather than e!ect sizes,
ing with strong statistical counter-evidence.
semantic “strength” need not match synthetic e!ect
magnitudes, and the LLM is queried only about direc-
• A forbidden arrow (encoded as (X, Y ) ⇒) re-
tions (required/forbidden), not e!ect size.
stricts the solver from orienting a remaining edge
in a certain direction. 2. Semantic Grounding via Isomorphism. To as-
sign meaningful concepts to the nodes of the random
This two-stage process leverages statistical evidence DAG, we use CauseNet, a large-scale knowledge graph
to define the initial search space, then employs high- of cause-and-e!ect relationships extracted from web
precision semantic knowledge to further guide the dis- text (Heindorf et al., 2020). We search for an induced
covery process with information that is often orthogo- sub-graph isomorphism from the random DAG to
nal to what can be inferred from data alone. the CauseNet graph. Formally, given a random DAG
H = (VH , EH ) and CauseNet G = (VG , EG ), we find
5 CauseNet Synthetic DAGs an injective function f : VH ⇑ VG such that for
any pair of nodes u, v ↓ VH , an edge (u, v) ↓ EH
A fundamental challenge in evaluating LLMs for causal exists if and only if an edge (f (u), f (v)) ↓ EG ex-
discovery is the risk of data leakage (Balloccu et al., ists. This strategy ensures that the selected sub-
2024; Dong et al., 2024). Standard benchmarks, such graph from CauseNet has the exact same structure as
as Asia (Lauritzen and Spiegelhalter, 2018) or Sachs our random DAG. We employ the e”cient VF2 algo-
(Sachs et al., 2005) in the bnlearn repository (Scu- rithm (Cordella et al., 2004) for this search.
tari, 2010), are widely published and likely part of the
3. Heuristic-Guided Candidate Selection. A sin-
LLMs’ pre-training corpora. Consequently, high per-
gle random DAG can have thousands or even millions
formance on these benchmarks may reflect memorisa-
of isomorphic matches within CauseNet. To find the
tion rather than generalisable causal reasoning.
most plausible causal scenario, we score and rank all
candidate sub-graphs using a flexible, heuristic-based
Why Semantic Grounding Is Necessary. Be- system that combines three distinct cost terms into a
cause LLM elicitation relies on variable semantics, single objective function to be minimised.
anonymising variables (e.g., X1 , . . . , Xn ) would re-
move the signal and reduce the model to an unin-
formed oracle. We therefore ground randomly gen- • Semantic Compactness: Measures the the-
erated structures in real concepts via CauseNet to ob- matic coherence of the concepts in a candidate
tain semantically meaningful yet structurally diverse graph. Using pre-computed vector embeddings
graphs; while individual CauseNet facts may appear for all concepts, we calculate the average cosine
in pre-training corpora, composing them into random, distance of each concept’s vector from their geo-
previously unseen subgraphs makes verbatim retrieval metric centroid. A low cost indicates that the con-
of an entire target DAG implausible and helps distin- cepts form a tight semantic cluster (e.g., virus,
guish memorisation from generalisation. fever, fatigue).
 Li, Russo

Figure 2: Synthetic Evaluation Protocol: We generate random DAGs and ground them in CauseNet by finding
sub-graph isomorphisms. Given that there might be many possible isomorphisms, we use a heuristic composed
of three scores to select the most suitable one. The output is a semantically grounded DAG with variable names
that can be used to evaluate the robustness of LLMs in generating causal assumptions.

• Node Specificity: Penalises candidate graphs 6.1 Experimental Setup

that include overly general or ambiguous “hub”
nodes (e.g., illness, problem). The cost is the Data. We evaluate our approach on both standard
average logarithm of each node’s degree in the full causal discovery benchmarks and our novel synthetic
CauseNet graph, favouring sub-graphs composed datasets derived from CauseNet. For each dataset, we
of more specific, well-defined concepts. generate 5000 observational samples and repeat the
• Structural-Semantic Correlation: Ensures experiment 50 times with di!erent random seeds. The
the graph’s topology faithfully reflects the se- standard pseudo-real BNs used are from the bnlearn
mantic relationships between its concepts. We repository (Scutari, 2010). Our synthetic datasets
compute the rank correlation (Spearman, 1904) consist of 54 random DAGs with number of nodes
between all-pairs shortest-path distances in the |V| ↓ {5, 10, 15}, generated using the protocol de-
graph and the corresponding pairwise cosine dis- scribed in Section 5. We ensure that these synthetic
tances between semantic embeddings. Lower cost graphs are semantically grounded in CauseNet and are
(inverse of correlation) means that concepts di- novel to the LLMs by using random structures and
rectly connected in the graph are also semanti- a heuristic selection process. Detailed statistics of all
cally very close, ensuring holistic consistency. datasets are provided in Appendix B and the synthetic
BNs produced are available in our repository, together
The final objective is the candidate graph that
with the code to reproduce the experiments.
minimises a weighted sum of these three heuris-
tic cost terms. This pipeline produces high-quality, The choice of CI testing method is consistent
semantically-grounded causal graphs that serve as a across MPC, ABAPC, and ABAPC-LLM: Wilks G2
robust foundation for evaluating causal discovery in- likelihood-ratio test (Agresti, 2018), with significance
volving LLMs. We make our synthetic DAG generator level ω = 0.05. Within ABA-PC and ABA-PC-LLM,
publicly available as open-source software. MPC is only used to compute (an e”cient number of)
CI tests and their associated p-values.
6 Experimental Evaluation
Metadata Preparation and Prompting. We gen-
We now present an empirical evaluation of our LLM- erated variable descriptions for causal graphs using an
augmented Causal ABA framework. Our experiments LLM, providing it with the ground-truth graph as a
aim to assess the accuracy and robustness of the in- basis for defining each variable’s role. The prompt
ferred causal graphs, particularly in comparison to es- was carefully designed to ensure the resulting descrip-
tablished causal discovery algorithms and under vary- tions were semantically consistent with the under-
ing conditions of LLM input quality. Throughout, lying causal structure without explicitly stating any
we focus on purely observational data; consequently, of the causal links. The full prompt is provided
conditional-independence evidence alone identifies the in Appendix A. We then used these variable names
causal graph only up to a Markov equivalence class and descriptions as input to our constraint elicitation
(MEC). Semantic constraints can provide additional pipeline (Section 4) to generate the LLM-derived con-
orienting information within the MEC, and we evalu- straints. We used Google’s gemini-2.5-flash for the
ate their downstream impact via standard structural primary LLM and gemini-2.5-flash-lite for the
metrics against the ground-truth DAG. structured output extraction, leveraging Anthropic’s
 Leveraging Large Language Models for Causal Discovery

Figure 3: Normalised Structural Hamming Distance (SHD, left-axis) and F1-score (right-axis) of LLM-augmented
Causal ABA against baselines across synthetic datasets generated from the CauseNet Knowledge Graph (see
Section 5) grouped by number of nodes (|V| ↓ {5, 10, 15}). Error bars are standard deviations over 50 repetitions.

prompt improver (Anthropic, 2024) to iteratively re- 6.2 Results and Analysis
fine our prompts. For each dataset, the same set of
LLM-derived constraints was used across all 50 data We summarise the experimental results, comparing
repetitions. The same variable names are used for ABAPC-LLM with the baselines. We first examine
our method and the LLM-only baseline (Jiralerspong structural learning performance and then analyse how
et al., 2024) to ensure a fair comparison. LLM-derived constraints interact with data-driven in-
dependence tests. All details in Appendix C.
Metrics. We measure the accuracy of the inferred
causal graphs using three metrics: Structural Ham-
ming Distance (Tsamardinos et al., 2006, SHD), Struc- Structural Learning Performance. Figure 3
tural Intervention Distance (Peters and Bühlmann, compares the structural accuracy achieved by our
2015, SID), and F1-score. SHD quantifies the number method against the baselines across the synthetic
of edge modifications needed to transform the inferred CauseNet graphs.3 ABAPC-LLM attains the lowest
graph into the true graph, while SID measures the cor- normalised SHD and the highest F1-score for all graph
rectness of inferred causal e!ects under interventions. sizes, with particularly large margins for five- and
The F1-score provides a balanced measure of preci- fifteen-node problems over MPC, FGS, NOTEARS-
sion and recall for the presence of edges in the inferred MLP, GRaSP, BOSS, and the LLM-only baseline. The
graph. Additional details are in Appendix C.1. gains persist on ten-node graphs, although the gap
narrows as the statistical signal strengthens.
Baselines. We compare our LLM-augmented
ABAPC (ABAPC-LLM) against several baselines: BH-corrected (Benjamini and Hochberg, 1995) two-
ABAPC (Russo et al., 2024), MPC (Colombo sample, unequal variance t-tests (Appendix C.3.2)
and Maathuis, 2014), FGS (Ramsey et al., 2017), demonstrate that the SHD and F1 improvements over
NOTEARS-MLP (Zheng et al., 2020), GRaSP (Lam all baselines are significant for all graph sizes. Di!er-
et al., 2022), and BOSS (Andrews et al., 2023). These ences with the variant of Causal ABA that relies exclu-
algorithms represent a mix of constraint-based, score- sively on independence tests are smaller, but remain
based, and continuous optimisation approaches, pro- positive, indicating that the semantic priors comple-
viding a comprehensive benchmark for our method. ment rather than contradict the statistical evidence.
MPC is also the statistical engine underlying ABAPC, Runtime measurements are reported in Ap-
allowing us to isolate the impact of LLM-derived con- pendix C.3.3, showing that ABAPC-LLM remains
straints. Additionally, we include an LLM-only base- practical for graphs up to 15 nodes.
line (Jiralerspong et al., 2024) to gauge impact of our
data-LLM integration, and a random graph generation 3
SID, precision and recall curves are reported in Ap-
baseline to establish a lower bound on performance.
pendix C.3 and show the same ordering.
 Li, Russo

LLM Constraints Quality and Interaction with

Data. To understand how our LLM-augmented
pipeline improves structural accuracy we inspect the
quality of the priors entering the solver in relation
to the data-driven constraints. For every dataset we
compute the F1-score of the LLM-derived constraints
against the ground-truth graph, the F1-score of the
retained CI statements, and the resulting change in
DAG F1 after adding the semantic constraints (#F1).
The heatmap in Figure 4 shows a clear synergy: high-
quality constraints translate into consistent F1 gains
when the underlying conditional independence infor-
mation is also reliable, whereas noisy LLM outputs
are either ignored or mildly detrimental when the sta- Figure 4: CauseNet Synthetic: Heatmap showing
tistical evidence is weak. This interaction highlights how the quality of LLM-derived and data-derived con-
the role of our structured integration, which privileges straints relates to changes in final graph reconstruction
the more trustworthy source while still exploiting com- accuracy after integration (#F1 against the true DAG,
plementary information when both signals agree. The in brackets the number of constraints n).
bnlearn benchmarks counterpart are reported in Ap-
pendix C.3.4 and show the same pattern, albeit with a
narrower range of LLM constraint quality due to po- Ahmed Abdulaal, adamos hadjivasiliou, Nina
tential memorisation. Splits by DAG characteristics Montana-Brown, Tiantian He, Ayodeji Ijishakin,
are reported in Appendix C.4 and show that the im- Ivana Drobnjak, Daniel C. Castro, and Daniel C.
provements are consistent across structural properties. Alexander. Causal modelling agents: Causal graph
discovery through synergising metadata- and data-
7 Conclusion driven reasoning. In The Twelfth International Con-
We introduced a hybrid causal discovery framework ference on Learning Representations, 2024. URL
that injects LLM-derived semantic knowledge into [Link]
Causal ABA, combining symbolic guarantees with rich Alan Agresti. An Introduction to Categorical Data
and unstructured semantic information. A dedicated Analysis. Wiley, 3 edition, 2018.
elicitation pipeline extracts high-precision structural
constraints from natural-language metadata, while our Bryan Andrews, Joseph Ramsey, Ruben
CauseNet-grounded synthetic benchmarks provide a Sanchez Romero, Jazmin Camchong, and Erich
setting where memorisation is unlikely and improve- Kummerfeld. Fast scalable and accurate discovery
ments in structural, interventional and precision–recall of DAGs using the best order score search and
metrics become apparent. grow–shrink trees. Advances in Neural Information
Processing Systems, 36, 2023.
Our findings show that consensus-filtered semantic
Anthropic. Prompt improver, 2024. URL https://
constraints can boost structural accuracy complement-
[Link]/news/prompt-improver. Ac-
ing statistical evidence, and they motivate several av-
cessed: 2025-09-29.
enues for future work. In particular, we plan to en-
rich the priors with knowledge mined from scientific Simone Balloccu, Patrı́cia Schmidtová, Mateusz
corpora, extend the dialogue with LLMs to reason Lango, and Ondřej Dušek. Leak, cheat, repeat:
about potential unobserved confounders, and endow Data contamination and evaluation malpractices in
the solver with adaptive weighting schemes that cali- closed-source llms, 2024. URL [Link]
brate LLM assertions using confidence scores that can org/abs/2402.03927.
be used to rank against data-driven learning. Explor- Taiyu Ban, Lyvzhou Chen, Xiangyu Wang, and Huan-
ing alternative models and prompt designs will fur- huan Chen. From query tools to causal archi-
ther clarify how di!erent LLM capabilities influence tects: Harnessing large language models for ad-
the trade-o! between coverage and precision. vanced causal discovery from data, 2023.
Albert-László Barabási and Réka Albert. Emer-
References gence of scaling in random networks. Science, 286
567-labs. Instructor: Structured outputs for llms. (5439):509–512, 1999. doi: 10.1126/science.286.
[Link] 5439.509. URL [Link]
2025. GitHub repository. abs/10.1126/science.286.5439.509.
 Leveraging Large Language Models for Causal Discovery

Pietro Baroni, Martin Caminada, and Massimil- 2023. URL [Link]

iano Giacomin. An introduction to argumenta- NEAoZRWHPN.
tion semantics. Knowl. Eng. Rev., 26(4):365– Diego Colombo and Marloes H. Maathuis. Order-
410, 2011. URL [Link] independent constraint-based causal structure
S0269888911000166. learning. J. Mach. Learn. Res., 15(1):3741–3782,
Yoav Benjamini and Yosef Hochberg. Controlling the 2014. URL [Link]
false discovery rate: a practical and powerful ap- 2627435.2750365.
proach to multiple testing. Journal of the Royal Gregory F. Cooper and Edward Herskovits. A
Statistical Society: Series B (Methodological), 57(1): bayesian method for the induction of probabilistic
289–300, 1995. networks from data. Machine Learning, 9(4):309–
Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, 347, 1992.
Russ Altman, Sanjeev Arora, Sydney von Arx, et al. L.P. Cordella, P. Foggia, C. Sansone, and M. Vento.
On the opportunities and risks of foundation mod- A (sub)graph isomorphism algorithm for matching
els. arXiv preprint arXiv:2108.07258, 2021. large graphs. IEEE Transactions on Pattern Anal-
Andrei Bondarenko, Phan Minh Dung, Robert A. ysis and Machine Intelligence, 26(10):1367–1372,
Kowalski, and Francesca Toni. An abstract, 2004. doi: 10.1109/TPAMI.2004.75.
argumentation-theoretic approach to default reason- Kristijonas Cyras, Xiuyi Fan, Claudia Schulz, and
ing. Artif. Intell., 93:63–101, 1997. doi: 10.1016/ Francesca Toni. Assumption-based argumentation:
S0004-3702(97)00015-5. URL [Link] Disputes, explanations, preferences. FLAP, 4(8),
10.1016/S0004-3702(97)00015-5. 2017. URL [Link]
Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie [Link]/downloads/[Link].
Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Yihong Dong, Xue Jiang, Huanyu Liu, Zhi Jin, Bin
Neelakantan, Pranav Shyam, Girish Sastry, Amanda Gu, Mengfei Yang, and Ge Li. Generalization or
Askell, Sandhini Agarwal, Ariel Herbert-Voss, memorization: Data contamination and trustworthy
Gretchen Krueger, Tom Henighan, Rewon Child, evaluation for large language models, 2024. URL
Aditya Ramesh, Daniel M. Ziegler, Je!rey Wu, [Link]
Clemens Winter, Christopher Hesse, Mark Chen,
Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Phan Minh Dung. On the acceptability of arguments
Chess, Jack Clark, Christopher Berner, Sam Mc- and its fundamental role in nonmonotonic reasoning,
Candlish, Alec Radford, Ilya Sutskever, and Dario logic programming and n-person games. Artif. In-
Amodei. Language models are few-shot learners, tell., 77(2):321–358, 1995. URL [Link]
2020. URL [Link] 10.1016/0004-3702(94)00041-X.

Nicholas Carlini, Florian Tramer, Eric Wallace, P Erdös and A Rényi. On random graphs i. Publica-
Matthew Jagielski, Ariel Herbert-Voss, Katherine tiones Mathematicae Debrecen, 6:290–297, 1959.
Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Martin Gebser, Roland Kaminski, Benjamin Kauf-
Erlingsson, Alina Oprea, and Nicolas Papernot. Ex- mann, and Torsten Schaub. Multi-shot ASP solv-
tracting training data from large language models. ing with clingo. Theory Pract. Log. Program., 19
In Proceedings of the 30th USENIX Security Sym- (1):27–82, 2019. URL [Link]
posium, pages 2633–2650, 2021. S1471068418000054.
Lyuzhou Chen, Taiyu Ban, Xiangyu Wang, Derui Lyu, John Gkountouras, Matthias Lindemann, Phillip
and Huanhuan Chen. Mitigating prior errors in Lippe, Efstratios Gavves, and Ivan Titov. Lan-
causal structure learning: Towards llm driven prior guage agents meet causality – bridging llms and
knowledge, 2023. URL [Link] causal world models, 2024. URL [Link]
2306.07032. org/abs/2410.19923.
David Maxwell Chickering. Optimal structure identi- Clark Glymour, Kun Zhang, and Peter Spirtes. Re-
fication with greedy search. J. Mach. Learn. Res., view of causal discovery methods based on graphical
3:507–554, 2002. URL [Link] models. Front. Genet., 04 June 2019 Sec. Statisti-
v3/[Link]. cal Genetics and Methodology, 10:524, 2019. URL
Kai-Hendrik Cohrs, Emiliano Diaz, Vasileios Sitokon- [Link]
stantinou, Gherardo Varando, and Gustau Camps- David Heckerman, Dan Geiger, and David M. Chick-
Valls. Large language models for constrained-based ering. Learning Bayesian Networks: The Combi-
causal discovery. In AAAI 2024 Workshop on ”Are nation of Knowledge and Statistical Data. Ma-
Large Language Models Simply Causal Parrots?”, chine Learning, 20(3):197–243, 1995. ISSN 1573-
 Li, Russo

0565. doi: 10.1023/A:1022623210503. URL https: and their application to expert systems. Jour-
//[Link]/10.1023/A:1022623210503. nal of the Royal Statistical Society: Series B
Stefan Heindorf, Yan Scholten, Henning Wachsmuth, (Methodological), 50(2):157–194, 12 2018. ISSN
Axel-Cyrille Ngonga Ngomo, and Martin Potthast. 0035-9246. doi: 10.1111/j.2517-6161.1988.tb01721.
Causenet: Towards a causality graph extracted from x. URL [Link]
the web. In Proceedings of the 29th ACM In- 1988.tb01721.x.
ternational Conference on Information & Knowl- Chenxi Liu, Yongqiang Chen, Tongliang Liu, Ming-
edge Management, CIKM ’20, page 3023–3030, New ming Gong, James Cheng, Bo Han, and Kun Zhang.
York, NY, USA, 2020. Association for Comput- Discovery of the hidden world with large language
ing Machinery. ISBN 9781450368599. doi: 10. models, 2024a. URL [Link]
1145/3340531.3412763. URL [Link] 2402.03941.
10.1145/3340531.3412763.
Michael Xieyang Liu, Frederick Liu, Alexander J. Fi-
Antti Hyttinen, Frederick Eberhardt, and Matti annaca, Terry Koo, Lucas Dixon, Michael Terry, and
Järvisalo. Constraint-based causal discovery: Con- Carrie J. Cai. ”we need structured output”: To-
flict resolution with answer set programming. In wards user-centered constraints on large language
Proceedings of the Thirtieth Conference on Uncer- model output. In Extended Abstracts of the CHI
tainty in Artificial Intelligence, UAI 2014, Quebec Conference on Human Factors in Computing Sys-
City, Quebec, Canada, July 23-27, 2014, pages 340– tems, CHI ’24, pages 1–9. ACM, may 2024b. doi:
349. AUAI Press, 2014. 10.1145/3613905.3650756. URL [Link]
Ziwei Ji, Nanyun Lee, Tianyi Yu, Qihui He, Ruibo org/10.1145/3613905.3650756.
Zhang, Xiang Ren, et al. A survey on hallucination
Christopher Meek. Causal inference and causal ex-
in natural language generation. ACM Computing
planation with background knowledge. In UAI
Surveys, 55(12):1–38, 2023.
’95: Proceedings of the Eleventh Annual Confer-
Zhijing Jin, Yuen Chen, Felix Leeb, Luigi Gresele, ence on Uncertainty in Artificial Intelligence, Mon-
Ojasv Kamal, Zhiheng Lyu, Kevin Blin, Fernando treal, Quebec, Canada, August 18-20, 1995, pages
Gonzalez, Max Kleiman-Weiner, Mrinmaya Sachan, 403–410. Morgan Kaufmann, 1995. URL https:
and Bernhard Schölkopf. CLadder: Assessing //[Link]/10.48550/arXiv.1302.4972.
causal reasoning in language models. In NeurIPS,
Judea Pearl. Causality: Models, Reasoning, and Infer-
2023. URL [Link]
ence. Cambridge University Press, 2 edition, 2009.
e2wtjx0Yqu.
Thomas Jiralerspong, Xiaoyin Chen, Yash More, J. Peters, D. Janzing, and B. Schölkopf.
Vedant Shah, and Yoshua Bengio. E”cient causal Elements of Causal Inference: Founda-
graph discovery using large language models. In tions and Learning Algorithms. MIT
ICLR 2024 Workshop: How Far Are We From AGI, Press, Cambridge, MA, USA, 2017. URL
2024. URL [Link] [Link]
5RBUTx75yr. elements-of-causal-inference/.
Elahe Khatibi, Mahyar Abbasian, Zhongqi Yang, Iman Jonas Peters and Peter Bühlmann. Structural inter-
Azimi, and Amir M. Rahmani. Alcm: Autonomous vention distance for evaluating causal graphs. Neu-
llm-augmented causal discovery framework, 2024. ral Comput., 27(3):771–799, 2015. URL https:
URL [Link] //[Link]/10.1162/NECO_a_00708.
Emre Kıcıman, Robert Ness, Amit Sharma, and Chen- Joseph D. Ramsey, Madelyn Glymour, Ruben
hao Tan. Causal reasoning and large language mod- Sanchez-Romero, and Clark Glymour. A mil-
els: Opening a new frontier for causality, 2023. lion variables and more: the fast greedy equiva-
Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yu- lence search algorithm for learning high-dimensional
taka Matsuo, and Yusuke Iwasawa. Large language graphical causal models, with an application to func-
models are zero-shot reasoners. In Advances in Neu- tional magnetic resonance images. Int. J. Data Sci.
ral Information Processing Systems, 2022. Anal., 3(2):121–129, 2017. URL [Link]
10.1007/s41060-016-0032-z.
Wai-Yin Lam, Bryan Andrews, and Joseph Ramsey.
Greedy relaxations of the sparsest permutation al- Alexander G. Reisach, Christof Seiler, and Sebastian
gorithm. In Proceedings of the 38th Conference on Weichwald. Beware of the simulated DAG! causal
Uncertainty in Artificial Intelligence (UAI), 2022. discovery benchmarks may be easy to game, 2021.
S. L. Lauritzen and D. J. Spiegelhalter. Local com- Fabrizio Russo, Anna Rapberger, and Francesca Toni.
putations with probabilities on graphical structures Argumentative causal discovery. In Proc. of KR
 2024, 2024. doi: 10.24963/KR.2024/88. URL Guangya Wan, Yunsheng Lu, Yuqi Wu, Mengxuan Hu,
[Link] and Sheng Li. Large language models for causal dis-
covery: Current landscape and future directions, feb
Karen Sachs, Omar Perez, Dana Pe’er, Dou-
2025. URL [Link]
glas A. Lau!enburger, and Garry P. Nolan.
Causal protein-signaling networks derived from Xingyu Wu, Kui Yu, Jibin Wu, and Kay Chen Tan.
multiparameter single-cell data. Science, 308 Llm cannot discover causality, and should be re-
(5721):523–529, 2005. doi: 10.1126/science. stricted to non-decisional support in causal discov-
1105809. URL [Link] ery, 2025. URL [Link]
abs/10.1126/science.1105809. 00844.
Marco Scutari. Learning bayesian networks with the Alessio Zanga, Elif Ozkirimli, and Fabio Stella. A
bnlearn R package. Journal of Statistical Software, survey on causal discovery: Theory and practice.
35(3):1–22, 2010. doi: 10.18637/jss.v035.i03. Int. J. Approx. Reason., 151:101–129, 2022. URL
[Link]
C. Spearman. The proof and measurement of associ-
Matej Zečević, Moritz Willig, Devendra Singh Dhami,
ation between two things. The American Journal
and Kristian Kersting. Causal parrots: Large lan-
of Psychology, 15(1):72–101, 1904. ISSN 00029556.
guage models may talk causality but are not causal.
URL [Link]
Transactions on Machine Learning Research, 2023.
Peter Spirtes, Clark Glymour, and Richard Scheines. ISSN 2835-8856. URL [Link]
Causation, Prediction, and Search. MIT Press, forum?id=tv46tCzs83.
2000. Zirui Zhao, Wee Sun Lee, and David Hsu. Large lan-
Zhi Rui Tam, Cheng-Kuang Wu, Yi-Lin Tsai, Chieh- guage models as commonsense knowledge for large-
Yen Lin, Hung-yi Lee, and Yun-Nung Chen. Let scale task planning, 2023. URL [Link]
me speak freely? a study on the impact of for- org/abs/2305.14078.
mat restrictions on large language model perfor- Xun Zheng, Bryon Aragam, Pradeep Ravikumar, and
mance. In Proc. of EMNLP 2024, pages 1218– Eric P. Xing. Dags with NO TEARS: continu-
1236, 2024. URL [Link] ous imization for structure learning. In Proc. of
[Link]-industry.91/. NeurIPS 2018, pages 9492–9503, 2018. URL https:
Francesca Toni. A tutorial on assumption-based argu- //[Link]/paper/2018/hash/
mentation. Argument Comput., 5(1):89–117, 2014. e347c51419ffb23ca3fd5050202f9c3d-Abstract.
URL [Link] html.
869878. Xun Zheng, Chen Dan, Bryon Aragam, Pradeep
Ioannis Tsamardinos, Laura E. Brown, and Con- Ravikumar, and Eric P. Xing. Learning sparse
stantin F. Aliferis. The max-min hill-climbing nonparametric dags. In Proc. of AISTATS 2020,
bayesian network structure learning algorithm. volume 108 of PMLR, pages 3414–3425. PMLR,
Mach. Learn., 65(1):31–78, 2006. URL https: 2020. URL [Link]
//[Link]/10.1007/s10994-006-6889-7. v108/[Link].

Aniket Vashishtha, Abbavaram Gowtham Reddy,

Abhinav Kumar, Saketh Bachu, Vineeth N. Bala-
subramanian, and Amit Sharma. Causal order: The
key to leveraging imperfect experts in causal infer-
ence. In 2025 International Conference on Learning
Representations, April 2025. URL [Link]
[Link]/en-us/research/publication/
causal-order-the-key-to-leveraging-imperfect-experts-in-causal-inference/.
Ashish Vaswani, Matthew Shardlow, et al. Attention
is all you need. Proceedings of the 34th International
Conference on Machine Learning, 70(1), 2017.
Matthew J. Vowels, Necati Cihan Camgoz, and
Richard Bowden. D’ya Like DAGs? A Survey on
Structure Learning and Causal Discovery. ACM
Comput. Surv., 55(4), November 2022. ISSN 0360-
0300. URL [Link]
 Li, Russo

Leveraging Large Language Models for Causal Discovery: a

Constraint-based, Argumentation-driven Approach
Supplementary Materials

In this supplementary material we provide details omitted from the main text due to space limitations. In
particular, we describe the prompts used for metadata enrichment and constraint elicitation (Appendix A),
the consensus strategy for stabilising LLM outputs, the generation and semantic grounding of the synthetic
CauseNet datasets (Appendix B), and the full experimental protocol, metrics, baselines, ablations, and additional
quantitative analyses (Appendix C).

A Prompts

This sections covers the details on the prompts used for the di!erent tasks in which LLMs are involved in. See
Figure 1 in the main text for the pipeline overview. The following prompts were built with various iterations of
Anthropic’s prompt improver tool (Anthropic, 2024).

A.1 Metadata Enrichment Prompt

Below, we provide the prompt used for eliciting the variable descriptions before interrogating the main LLM
(LLM 1 in Figure 1) about causal relations amongst variables. The following prompt is related to LLM 2 in
Figure 1, which we deem an optional block as its necessity may depend on context. Ablations on the importance

! "
of variable descriptions are in Section 4.

You are tasked with generating descriptions for causal variables in a randomly
generated causal graph used for synthetic dataset evaluations of causal discovery
algorithms . Your goal is to create meaningful descriptions for each variable that
accurately reflect their role in the graph without revealing their relationships
with other variables .

Here is the description of the randomly generated causal graph :

< causal_graph >

{ C A U SA L _ G RA P H _ D E S C R IP T I O N }
</ causal_graph >

To complete this task , follow these steps :

1. Carefully read and analyze the causal graph description .

2. For each variable mentioned in the graph , create a description that :

a ) Precisely describes its meaning within the context of the graph
b ) Does NOT spoil its relationships with other variables , either explicitly or
implicitly
c ) Maintains a consistent context or scenario across all variable descriptions

3. If a variable has different contextual meanings in its relationships with other

variables , provide a general description that could encompass these different
meanings without revealing the specific relationships .

4. After generating all variable descriptions , assess the quality of the random causal
graph based on how well the variables ’ meanings align with each other and how
realistic the overall scenario is .

5. Present your output in the following format :

 Leveraging Large Language Models for Causal Discovery

< title >

[ Provide a concise title for the causal graph ]
</ title >

< variable_descriptions >

[ List each variable and its description ]
</ variable_descriptions >

< graph_quality_assessment >

[ Provide your assessment of the graph quality , including how well the variables ’
meanings align and how realistic the overall scenario is ]
</ graph_quality_assessment >

Remember , your primary goal is to create meaningful descriptions without revealing any
causal relationships . Be creative in developing a consistent context that could
# $
plausibly connect all the variables .

A.2 LLM constraints elicitation prompt

Below, we provide the principal prompt used in our proposed LLM elicitation pipeline. Note that this prompt
significantly di!ers from prior work on LLMs and causal discovery, e.g. (Jiralerspong et al., 2024; Kıcıman et al.,
2023), since it involves a single call for all variables in the dataset. This is enabled by the much longer context
window of modern LLMs (e.g., over 1 million tokens for gemini-2.5-flash). This single-call approach is not
only more computationally e”cient but also allows the LLM to reason holistically about the entire system of

! "
variables, potentially capturing higher-order interactions that pairwise queries might miss.

You will be analyzing a set of causal variables to determine the required and
forbidden causal directions between them . Your goal is to achieve very high
precision in identifying these relationships .

< causal_variables >

{ CAUSAL_VARIABLES }
</ causal_variables >

Your task is to analyze these causal variables and generate two sets :
1. ** Required directions **: Causal relationships that must exist based on logical
necessity , temporal ordering , or fundamental causal principles
2. ** Forbidden directions **: Causal relationships that cannot exist due to logical
impossibility , temporal constraints , or definitional contradictions

Here are the key principles to follow :

** Required directions ** should include :

- Relationships where one variable definitionally or logically must cause another
- Temporal precedence relationships ( causes must precede effects )
- Relationships where the causal mechanism is well - established and unavoidable

** Forbidden directions ** should include :

- Relationships that would violate temporal ordering ( effects cannot cause their own
causes )
- Relationships that are logically contradictory or definitionally impossible
- Relationships where the direction would violate established causal mechanisms

** Important guidelines :**

- Only include relationships where you have very high confidence
- When in doubt about a relationship , do not include it in either set
- Consider both direct and indirect causal pathways
- Be precise about the direction of causality ( A → B is different from B → A )

Format your final answer as follows :

** Required Directions :**

 Li, Russo

- [ Variable A ] → [ Variable B ]: [ Brief justification ]

- [ Continue for all required directions ]

** Forbidden Directions :**

- [ Variable C ] → [ Variable D ]: [ Brief justification ]
- [ Continue for all forbidden directions ]

Your final answer should only include the Required Directions and Forbidden Directions
sections with their respective causal relationships and justifications . Aim for
very high precision - only include relationships where you are highly confident in
# $
the causal direction requirement or prohibition .

A.3 LLM Consensus Details

The consensus mechanism is designed to generate high-precision causal constraints by aggregating the outputs
from multiple independent LLM queries. This is due to the inherent stochasticity of LLMs, and inspired by the
majority voting strategy proposed by (Cohrs et al., 2023).
For each causal discovery problem, the LLM is queried five times, yielding five distinct sets of constraints. Let
Ri and Fi represent the set of “required” and “forbidden” arrows, respectively, from run i ↓ {1, . . . , 5}. The final
!5
high-confidence consensus sets are obtained by taking the intersection across the five runs: Rconsensus = i=1 Ri
!5
and Fconsensus = i=1 Fi . This guarantees that an arrow appears in the final constraints only if it was proposed
unanimously across all independent evaluations. The repetition count of five was chosen empirically to balance
precision and recall, with enough repeats to filter out stochastic or low-confidence suggestions while avoiding
excessive computational cost. The conservative intersection rule therefore substantially improves the precision
of the resulting constraints, as empirically validated in Appendix C.4.

B Graph and Data Generation Details

The generation of our synthetic datasets is a two-stage process designed to create structurally diverse and
semantically plausible causal graphs. First, we generate a structural sca!old (a Directed Acyclic Graph, or
DAG) using one of three di!erent topology generation methods. Second, we ground this abstract structure in
real-world concepts by finding a matching sub-graph within the CauseNet knowledge graph, guided by specific
heuristics.

B.1 Structural Sca!olding and Graph Types

We generate the initial DAG topologies using three distinct methods to ensure a variety of graph structures:

• Erdős-Rényi (ER) and Scale-Free (SF): These two methods are adapted from the well-established
NOTEARS codebase (Zheng et al., 2018). The ER model (Erdös and Rényi, 1959) produces graphs with a
random, uniform edge distribution, while the SF model (Barabási and Albert, 1999) generates graphs with
power-law degree distributions, characterized by a few high-degree “hub” nodes.

• Lower Triangle (LT): This custom method constructs a connected DAG by directly populating a lower
triangular adjacency matrix. The algorithm proceeds in two stages. First, to guarantee connectivity, it
ensures every node from 1 to n ⇓ 1 has at least one parent by randomly adding a single incoming edge
for each from a node with a lower index. Second, it distributes the remaining specified number of edges
by randomly placing them in the unoccupied positions of the lower triangle. This two-step process ensures
both connectivity and the desired edge density.

While the ER and SF implementations from NOTEARS can produce graphs with a slightly varied number of
nodes and edges, the LT method allows for precise control over these parameters.
 Leveraging Large Language Models for Causal Discovery

B.2 Semantic Grounding and Heuristics

Once a structural sca!old is generated, we imbue it with semantic meaning by finding an isomorphic sub-graph
within our CauseNet knowledge graph. Since multiple isomorphic sub-graphs can exist, we employ one of three
heuristics to select the most suitable candidate match.

• none: This baseline heuristic serves as a control. It performs no optimization and simply selects the first
isomorphic sub-graph returned by the search algorithm.
• degrees: This heuristic aims to find the most specific and least central concepts. It selects the candidate sub-
graph that minimises the sum of the in-degrees and out-degrees of all its nodes within the larger CauseNet
graph. The intuition is that nodes with lower total degrees represent more niche concepts.
• semantics: This composite heuristic, as detailed in the main paper, selects the candidate sub-graph that
minimizes a weighted quality score combining three metrics. For a given candidate subgraph with concept
(node) set Sc , the metrics are:
– Semantic Compactness (Hcompact ): Measures thematic coherence by calculating the average cosine
distance of each concept’s vector embedding from their geometric centroid:
" #
Hcompact (Sc ) = avgv→Sc dcos (emb(v), µSc )
where emb(v) is the vector embedding of concept v, µSc is the centroid vector of all concept embeddings
in the set Sc , and dcos is the cosine distance.
– Node Specificity (Hspec ): Penalises overly general “hub” nodes using the average log-degree of each
node in the full CauseNet graph:
Hspec (Sc ) = avgv→Sc (log(deg(v) + 1))
where deg(v) is the degree (sum of in- and out-degrees) of node v in the full CauseNet graph. The
logarithm dampens the penalty for extremely high-degree nodes.
– Structural-Semantic Correlation (Hcorr ): Ensures alignment between graph topology and semantic
relationships. The score to be minimised is based on the Spearman’s rank correlation:
Hcorr (Sc ) = 1 ⇓ SpearmanCorr(dgraph , dsem )
where dgraph is the set of all pairwise shortest-path distances between nodes in the candidate graph,
and dsem is the corresponding set of pairwise cosine distances between their semantic embeddings. A
high correlation (low score) indicates that structurally close nodes are also semantically close.
The final score to be minimised is a weighted sum Hf inal = w1 Hcompact + w2 Hspec + w3 Hcorr , where the
weights wi are customisable.

Additionally, our implementation supports user-defined heuristics, allowing researchers to specify custom scoring
functions to rank candidate sub-graphs according to their own criteria or domain-specific preferences.

B.3 Dataset Schema

The full suite of 54 synthetic datasets was generated by creating one graph for each unique combination of the
following parameters:

• Number of Nodes: {5, 10, 15}

• Graph Density: Two levels of edge counts for each node size:
– 5 nodes: 5 and 7 edges
– 10 nodes: 10 and 15 edges
– 15 nodes: 15 and 22 edges
• Graph Type: {ER, SF, LT}
• Heuristic: {none, degrees, semantics}

Some examples of the generated graphs are shown in Figure 5.

 Li, Russo

high_blood_pressure

tumors eye_problems osteoporosis

dysfunction brain_damage birth_defect degeneration

sleep_apnea chronic_pain blindness spinal_stenosis

irritability neck_pain

dag_5_nodes_5_edges_semantics_SF

numbness

dag_10_nodes_15_edges_none_ER

chronic_pain lack_of_appropriate_supervision responsibilities lack_of_sleep poor_sleep infertility

despair abandoned_child_syndrome burnout behavioral_problems emotional_problems

alcohol_abuse substance_abuse drug_abuse

social_problems

dag_15_nodes_15_edges_degrees_random

Figure 5: Examples of semantically grounded DAGs from our synthetic data pipeline, generated with di!erent
structural methods (ER, SF, LT) and semantic heuristics (none, degrees, semantics).

B.3.1 CPT and Data Generation

After a semantically grounded DAG is finalised, we use the pyAgrum library to construct the corresponding
Bayesian Network. For each variable in the network, the domain size is set to 2, making all variables binary.
The Conditional Probability Tables (CPTs) for the BN are randomly generated during the initialisation process.
The full collection of the 54 generated Bayesian Networks in BIFXML and PNG format is available in the
project’s public repository, allowing for full reproducibility of our experiments.
The randomly initialised CPTs are not intended to reflect realistic e!ect sizes or context-specific independencies.
Their sole purpose is to generate observational data that respects the sampled graph structure. Our evaluation
targets structural recovery (presence and orientation of edges), not the calibration of causal e!ect magnitudes.
Mismatches between semantic strength and statistical e!ect size are therefore expected and mirror low-signal
regimes encountered in practice.

C Details on Experiments

We provide here additional details on the experimental setup, metrics, and results that were omitted from the
main text due to space constraints.
 Leveraging Large Language Models for Causal Discovery

C.1 Metrics Definitions

We evaluate the performance of causal discovery algorithms using standard metrics that capture di!erent aspects
of structural accuracy. Let Gtrue = (V, Etrue ) be the ground truth and Gest = (V, Eest ) the estimated DAG.

• Structural Hamming Distance (SHD): The SHD (Tsamardinos et al., 2006) measures the structural
di!erence between two graphs. It is defined as the number of edge operations (additions, deletions, or
reversals) required to transform Gest into Gtrue . A lower SHD indicates a better structural match. In our
plots, we report the normalised SHD (NSHD), which is the SHD divided by the number of edges in the true
graph, to facilitate comparison across graphs of di!erent sizes.
• Structural Intervention Distance (SID): The SID (Peters and Bühlmann, 2015) is a metric that quan-
tifies the di!erence in the interventional distributions implied by two causal graphs. It counts the number
of pairs (i, j) for which the causal e!ect of intervening on node i on node j is incorrectly predicted by Gest
compared to Gtrue . A lower SID indicates a better match in terms of causal predictions. We report the
normalised SID, dividing by the total number of possible interventions.
• Precision, Recall, and F1-Score: These metrics evaluate the accuracy of the learned graph skeleton (the
set of adjacencies, ignoring direction). Let TP (True Positives) be the number of correctly identified edges,
FP (False Positives) be the number of incorrectly identified edges, and FN (False Negatives) be the number
of missed edges.
– Precision is the fraction of correctly identified edges among all edges in the estimated graph:
TP
Precision = TP+FP .
TP
– Recall is the fraction of true edges that were correctly identified: Recall = TP+FN .
– F1-Score is the harmonic mean of Precision and Recall, providing a single balanced measure: F 1 =
Precision·Recall
2 · Precision+Recall .

C.2 Baselines

We used the following six baselines with respective implementations (see Section 6 for context):

• A Random baseline (RND) as in (Russo et al., 2024), by just sampling 10 random graphs with the same
number of nodes as the ground truth.
• Fast Greedy equivalence Search4 (FGS) (Ramsey et al., 2017) is a score-based Causal Discovery algorithm.
It is a fast implementation of GES (Chickering, 2002) where graphs are evaluated using the Bayesian Infor-
mation Criterion (BIC) upon addition or deletion of an edge, in a greedy fashion, involving the evaluation
of insertion and removal of edges in a forward and backward fashion.
• NOTEARS-MLP5 (NT) (Zheng et al., 2020) learns a non-linear SEM via continuous optimisation. Having
a Multi-Layer Perceptron (MLP) at its core, this method should adapt to di!erent functional dependen-
cies among the variables. The optimisation is carried out via augmented Lagrangian with a continuous
formulation of acyclicity (Zheng et al., 2018).
• GRaSP (Lam et al., 2022) is an order-based method based on greedy relaxations of the sparsest permutation
criterion, providing a recent and competitive baseline for structure learning.
• BOSS (Andrews et al., 2023) is an order-based score-search method that combines order search with grow–
shrink-style neighbourhood exploration for fast DAG discovery.
• Majority-PC6 (MPC) (Colombo and Maathuis, 2014) is a constraint-based causal discovery algorithm. It
uses independence tests and graphical rules based on d-separation to extract a CPDAG from the data.
MPC is an improved version of the original Peter-Clark (PC) algorithm (Spirtes et al., 2000) that renders
it order-independent while maintaining soundness and completeness with infinite data. This constitutes the
statistical engine underlying ABAPC, isolating the impact of Causal ABA.
4
[Link]
5
[Link]
6
[Link]
 Li, Russo

Figure 6: Precision and Recall on Synthetic Datasets. Bar plots comparing the Precision and Recall of LLM-
augmented Causal ABA against baselines across synthetic datasets generated from the CauseNet Knowledge
Graph, grouped by number of nodes (|V| ↓ {5, 10, 15}). Error bars represent standard deviations over 50
repetitions.

• ABAPC7 (ABAPC) (Russo et al., 2024) is an instantiation of Causal ABA, a logic- and constraint-based
Causal Discovery algorithm that uses an Assumption-Based Argumentation framework to resolve conflicts
and ambiguities within observational data. The method uses the CI tests from the MPC algorithm as inputs
into the Causal ABA (hence ABAPC) and excludes a progressively bigger amount of low-confidence tests
to find a stable extension of the ABAF that contains a DAG which is guaranteed to imply the d-separations
supported by the retained tests. This is the non-LLM variant of our proposed method, allowing us to isolate
the impact of LLM-derived constraints.
• LLM-BFS8 (BFS) (Jiralerspong et al., 2024) is a breadth-first search-based approach that leverages LLMs for
causal discovery. It explores the graph structure by iteratively expanding the search space and incorporating
LLM-generated insights to build the causal graph, only leveraging variable names and descriptions. This
method is the LLM-only baseline, allowing us to isolate the impact of data-driven constraints.

C.3 Additional Results

The following sections provide additional results and analyses that complement the findings presented in the
main text. We include further performance metrics, statistical test results, and detailed evaluations on the
bnlearn benchmark datasets.

C.3.1 Additional Metrics

Figures 6 and 7 complement Figure 3 from the main text by showing the performance on the synthetic CauseNet
datasets with respect to Precision, Recall, and SID. The results confirm the trend observed with SHD and
F1-score: our proposed ABAPC-LLM method consistently outperforms the baselines across all graph sizes,
demonstrating superior precision and recall in edge detection and greater accuracy in predicting interventional
e!ects.

C.3.2 Statistical Test Results

Here we provide details for the statistical tests used to measure the significance of the di!erence in the results
presented in Figure 3 in the main text. In Table 2 we provide t-statistics and p-values for all comparisons against
7
[Link]
8
[Link]
 Leveraging Large Language Models for Causal Discovery

Figure 7: Structural Intervention Distance on Synthetic Datasets. Bar plots comparing the normalised Structural
Intervention Distance (SID) of LLM-augmented Causal ABA against baselines across synthetic datasets generated
from the CauseNet Knowledge Graph, grouped by number of nodes (|V| ↓ {5, 10, 15}). Error bars represent
standard deviations over 50 repetitions.

Table 1: Average runtime in seconds on synthetic CauseNet datasets (mean±std over repetitions). LLM-BFS is
omitted because its wall-clock time is dominated by external API calls.

Method |V| = 5 |V| = 10 |V| = 15

Random 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00
FGS 0.08 ± 0.00 0.09 ± 0.00 0.10 ± 0.01
MPC 0.00 ± 0.01 0.02 ± 0.01 0.03 ± 0.01
ABAPC 0.16 ± 0.11 3.31 ± 3.70 5.39 ± 5.19
ABAPC-LLM 0.18 ± 0.14 3.28 ± 3.47 5.31 ± 5.14
NOTEARS-MLP 0.11 ± 0.01 0.27 ± 0.05 0.45 ± 0.04
GRaSP 1.09 ± 0.10 5.45 ± 2.67 14.06 ± 4.80
BOSS 1.29 ± 0.16 7.22 ± 3.20 18.63 ± 7.05

our proposed method, ABAPC-LLM, on the synthetic CauseNet datasets. LLM-BFS is not included in these
tests as it was only run once (due to its high computational cost).
For all families of pairwise comparisons, we apply the Benjamini–Hochberg procedure to control the false discovery
rate. Corrected p-values are reported in the pBH column of Tables 2 and subsequent tables. All significance claims
in the main text refer to BH-corrected results.
In the table we present pairwise comparisons of means, for F1 and Normalised SHD, scores presented in Figure 3
of the main text. We use two-sample, unequal variance t-tests, with degrees of freedom of 53 (10 seeds and
4 noise distributions, minus 1). The null hypothesis is that the means of the two samples are equal, and the
alternative hypothesis is that they are not equal. We reject the null hypothesis for p-values below 0.05, indicating
a statistically significant di!erence between the two methods being compared.
All comparisons show statistically significant di!erences (p < 0.001) between ABAPC-LLM and all baselines,
across all graph sizes and both metrics.

C.3.3 Runtime Analysis

Table 1 reports average wall-clock runtime (in seconds) for all methods on the synthetic datasets, averaged over
50 repetitions. To focus on algorithmic scaling, we report runtimes excluding external LLM API latency; LLM
 Li, Russo

Table 2: Two-sample, unequal variance t-tests (ABAPC-LLM vs others) on Synthetic Data for |V | ↓ {5, 10, 15}
with nobs = 900.0. Metrics: F1 (higher is better) and NSHD (lower is better). Significance levels: 0 ’***’ 0.001
’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ns’ 1. BH-corrected p-values are also reported to account for multiple comparisons.
Dataset Metric Methods Means±Std t p-value pBH
ABAPC-LLM vs ABAPC 0.66 ± 0.09 vs 0.63 ± 0.11 8.05 1.55e-15 *** 1.92e-15 ***
ABAPC-LLM vs FGS 0.66 ± 0.09 vs 0.17 ± 0.07 134.20 0 *** 0 ***
F1 ABAPC-LLM vs MPC 0.66 ± 0.09 vs 0.36 ± 0.13 58.89 0 *** 0 ***
ABAPC-LLM vs NOTEARS 0.66 ± 0.09 vs 0.16 ± 0.00 169.62 0 *** 0 ***
ABAPC-LLM vs Random 0.66 ± 0.09 vs 0.19 ± 0.08 114.96 0 *** 0 ***
|V | = 10
ABAPC-LLM vs ABAPC 0.44 ± 0.10 vs 0.47 ± 0.11 ↑6.29 3.92e-10 *** 4.34e-10 ***
ABAPC-LLM vs FGS 0.44 ± 0.10 vs 1.31 ± 0.15 ↑146.55 0 *** 0 ***
NSHD ABAPC-LLM vs MPC 0.44 ± 0.10 vs 0.88 ± 0.14 ↑77.21 0 *** 0 ***
ABAPC-LLM vs NOTEARS 0.44 ± 0.10 vs 1.00 ± 0.01 ↑174.07 0 *** 0 ***
ABAPC-LLM vs Random 0.44 ± 0.10 vs 2.34 ± 0.53 ↑105.08 0 *** 0 ***
ABAPC-LLM vs ABAPC 0.68 ± 0.08 vs 0.62 ± 0.10 14.28 8.84e-44 *** 1.16e-43 ***
ABAPC-LLM vs FGS 0.68 ± 0.08 vs 0.11 ± 0.04 196.72 0 *** 0 ***
F1 ABAPC-LLM vs MPC 0.68 ± 0.08 vs 0.29 ± 0.12 84.89 0 *** 0 ***
ABAPC-LLM vs NOTEARS 0.68 ± 0.08 vs 0.12 ± 0.00 224.45 0 *** 0 ***
ABAPC-LLM vs Random 0.68 ± 0.08 vs 0.13 ± 0.05 183.38 0 *** 0 ***
|V | = 15
ABAPC-LLM vs ABAPC 0.38 ± 0.07 vs 0.43 ± 0.09 ↑12.87 2.93e-36 *** 3.23e-36 ***
ABAPC-LLM vs FGS 0.38 ± 0.07 vs 1.33 ± 0.12 ↑201.95 0 *** 0 ***
NSHD ABAPC-LLM vs MPC 0.38 ± 0.07 vs 0.84 ± 0.10 ↑107.23 0 *** 0 ***
ABAPC-LLM vs NOTEARS 0.38 ± 0.07 vs 0.93 ± 0.01 ↑222.07 0 *** 0 ***
ABAPC-LLM vs Random 0.38 ± 0.07 vs 3.56 ± 1.09 ↑87.39 0 *** 0 ***
ABAPC-LLM vs ABAPC 0.81 ± 0.07 vs 0.68 ± 0.08 37.41 5.92e-226 *** 8.88e-226 ***
ABAPC-LLM vs FGS 0.81 ± 0.07 vs 0.32 ± 0.10 120.03 0 *** 0 ***
F1 ABAPC-LLM vs MPC 0.81 ± 0.07 vs 0.45 ± 0.14 68.68 0 *** 0 ***
ABAPC-LLM vs NOTEARS 0.81 ± 0.07 vs 0.34 ± 0.00 198.82 0 *** 0 ***
ABAPC-LLM vs Random 0.81 ± 0.07 vs 0.34 ± 0.15 87.12 0 *** 0 ***
|V | = 5
ABAPC-LLM vs ABAPC 0.21 ± 0.08 vs 0.33 ± 0.08 ↑31.88 6.47e-177 *** 9.06e-177 ***
ABAPC-LLM vs FGS 0.21 ± 0.08 vs 0.96 ± 0.18 ↑112.64 0 *** 0 ***
NSHD ABAPC-LLM vs MPC 0.21 ± 0.08 vs 0.83 ± 0.12 ↑132.23 0 *** 0 ***
ABAPC-LLM vs NOTEARS 0.21 ± 0.08 vs 0.89 ± 0.01 ↑264.60 0 *** 0 ***
ABAPC-LLM vs Random 0.21 ± 0.08 vs 1.18 ± 0.25 ↑110.54 0 *** 0 ***

inference is performed once per dataset and amortised across repetitions.

C.3.4 Performance on bnlearn Benchmarks

Figures 8, 9, and 10 show the performance of all methods on the standard benchmark datasets from the bnlearn
repository. As discussed in the main text, these well-known datasets are likely part of the LLMs’ training data,
which may lead to memorisation e!ects. This is reflected in the strong performance of both ABAPC-LLM and
the LLM-only baseline (LLM-BFS) on several datasets like SACHS, ASIA, and CANCER. Nonetheless, our
method remains competitive or superior across all metrics, and its performance on datasets like CHILD, which
are more complex, demonstrates that the combination of data-driven reasoning and semantic priors is robust.
Interesting to note is the 100% recall achieved by LLM-BFS on four of the six datasets (Figure 9), indicating
that the LLM was able to identify all true edges in these cases, albeit with lower precision. This suggests that
while the LLM can e!ectively capture the presence of causal relationships, it may also introduce spurious edges,
highlighting the importance of combining LLM insights with data-driven methods for balanced performance.
Additionally, the SID score of LLM-BFS on these same datasets is 0 (no errors in interventional predictions),
which raises suspitions about the performance on these common benchmarks.

C.4 LLM Constraints Evaluation

We provide here additional analyses on the quality of the causal constraints generated by the LLM, as well as
ablation studies to assess the impact of di!erent components of our approach. In particular, we investigate the
role of variable descriptions in enhancing the LLM’s understanding of the causal relationships, and we evaluate
 Leveraging Large Language Models for Causal Discovery

Figure 8: Structural Learning Performance on bnlearn benchmarks. Bar plots comparing the normalised Struc-
tural Hamming Distance (NSHD, left-axis) and F1-score (right-axis) of our method against baselines across
standard benchmark datasets. Error bars represent standard deviations over 50 repetitions.

the e!ectiveness of our consensus mechanism in improving the reliability of the generated constraints. Finally
we provide additional heatmaps to illustrate the interaction between constraint quality and structural learning
performance.
Table 3 summarises the performance of the LLM in generating causal constraints for bnlearn and our CauseNet
synthetic datasets, comparing the consensus-based approach with a simple average of single runs, and evaluating
the impact of providing variable descriptions.

Table 3: LLM-derived constraint quality, including average number of constraints, under di!erent elicitation
strategies. We compare consensus vs. the average of single runs, and with vs. without descriptions on bnlearn
and bnlearn. Bold = higher of Average vs. Consensus within each dataset and description column; ⇔ “with
desc” > “w/o desc” (same method), ↖ otherwise. Green = overall best across all settings per row.

bnlearn CauseNet
Metric Average Consensus Average Consensus
w/o desc with desc w/o desc with desc w/o desc with desc w/o desc with desc
Forbidden constraints
# constraints 19.80 ± 13.90 25.47 ± 24.39 8.00 ± 6.60 8.17 ± 1.72 17.59 ± 13.02 20.64 ± 16.60 4.85 ± 4.83 4.07 ± 3.42
Precision 0.927 ± 0.186 0.973 ± 0.041⇔ 0.833 ± 0.408 0.948 ± 0.085⇔ 0.952 ± 0.132 0.942 ± 0.155↖ 0.855 ± 0.339 0.877 ± 0.317⇔
Recall 0.409 ± 0.234 0.472 ± 0.219⇔ 0.241 ± 0.240 0.275 ± 0.216⇔ 0.293 ± 0.216 0.312 ± 0.221⇔ 0.123 ± 0.164 0.115 ± 0.168↖
F1 0.533 ± 0.260 0.606 ± 0.224⇔ 0.338 ± 0.316 0.392 ± 0.281⇔ 0.409 ± 0.232 0.431 ± 0.234⇔ 0.189 ± 0.216 0.174 ± 0.218↖
Required constraints
# constraints 8.27 ± 6.97 10.80 ± 11.71 4.50 ± 3.73 3.67 ± 4.13 12.57 ± 9.60 14.56 ± 11.52 4.11 ± 3.29 4.74 ± 3.33
Precision 0.468 ± 0.288 0.597 ± 0.235⇔ 0.375 ± 0.327 0.505 ± 0.422⇔ 0.487 ± 0.249 0.464 ± 0.230↖ 0.568 ± 0.353 0.514 ± 0.335↖
Recall 0.445 ± 0.352 0.611 ± 0.328⇔ 0.303 ± 0.388 0.281 ± 0.375↖ 0.486 ± 0.255 0.516 ± 0.242⇔ 0.244 ± 0.210 0.278 ± 0.244⇔
F1 0.401 ± 0.240 0.555 ± 0.232⇔ 0.284 ± 0.284 0.321 ± 0.337⇔ 0.450 ± 0.206 0.453 ± 0.190⇔ 0.314 ± 0.228 0.334 ± 0.249⇔

The consensus approach is conservative and often results in a smaller set of constraints. In some cases, it may
yield an empty set, which can artificially lower the aggregated performance metrics since they are set to 0 in such
instances. To provide a clearer view of the performance when constraints are generated, Table 4 presents the
same metrics but excludes cases where the number of generated constraints is zero. This is particularly relevant
for the consensus method, which is more prone to this scenario. Based on the results in these tables, we can
draw the following observations:
 Li, Russo

Figure 9: Precision and Recall on bnlearn benchmarks. Bar plots comparing the Precision and Recall of our
method against baselines across standard benchmark datasets. Error bars represent standard deviations over 50
repetitions.

Table 4: LLM-derived constraint quality excluding cases with zero constraints. We include average number of
constraints, under di!erent elicitation strategies. We compare consensus vs. the average of single runs, and with
vs. without descriptions on bnlearn and bnlearn. Bold = higher of Average vs. Consensus within each dataset
and description column; ⇔ “with desc” > “w/o desc” (same method), ↖ otherwise. Green = overall best across
all settings per row.

bnlearn CauseNet
Metric Average Consensus Average Consensus
w/o desc with desc w/o desc with desc w/o desc with desc w/o desc with desc
Forbidden constraints
# constraints 20.48 ± 13.63 25.47 ± 24.39 9.60 ± 5.94 8.17 ± 1.72 17.86 ± 12.94 20.79 ± 16.57 5.57 ± 4.78 4.58 ± 3.29
Precision 0.959 ± 0.062 0.973 ± 0.041⇔ 1.000 0.948 ± 0.085↖ 0.966 ± 0.061 0.949 ± 0.132↖ 0.982 ± 0.069 0.986 ± 0.053⇔
Recall 0.423 ± 0.225 0.472 ± 0.219⇔ 0.289 ± 0.234 0.275 ± 0.216↖ 0.297 ± 0.215 0.314 ± 0.220⇔ 0.141 ± 0.168 0.129 ± 0.172↖
F1 0.551 ± 0.244 0.606 ± 0.224⇔ 0.405 ± 0.300 0.392 ± 0.281↖ 0.415 ± 0.228 0.434 ± 0.231⇔ 0.217 ± 0.218 0.196 ± 0.222↖
Required constraints
# constraints 9.92 ± 6.45 11.17 ± 11.74 6.75 ± 1.71 5.50 ± 3.87 12.66 ± 9.57 14.66 ± 11.50 4.53 ± 3.17 5.02 ± 3.22
Precision 0.562 ± 0.213 0.618 ± 0.210⇔ 0.562 ± 0.195 0.757 ± 0.206⇔ 0.491 ± 0.247 0.468 ± 0.227↖ 0.626 ± 0.317 0.544 ± 0.320↖
Recall 0.535 ± 0.316 0.632 ± 0.312⇔ 0.454 ± 0.399 0.422 ± 0.394↖ 0.490 ± 0.252 0.520 ± 0.239⇔ 0.269 ± 0.204 0.294 ± 0.241⇔
F1 0.482 ± 0.172 0.574 ± 0.211⇔ 0.426 ± 0.233 0.481 ± 0.294⇔ 0.454 ± 0.203 0.457 ± 0.187⇔ 0.346 ± 0.214 0.354 ± 0.242⇔

• Impact of Descriptions: Providing semantic descriptions alongside variable names generally improves
the performance of the LLM in generating both forbidden and required constraints. As shown in Table 3,
for forbidden constraints, descriptions lead to a notable increase in F1-score for both bnlearn (from 0.53 to
0.61 for average, and 0.34 to 0.39 for consensus) and synthetic datasets. A similar, even more pronounced,
improvement is observed for required constraints, where the F1-score on bnlearn jumps from 0.4 to 0.56 for
the average method. This suggests that richer context allows LLMs to make more accurate causal judgments.
• Consensus vs. Average of Single Runs: The consensus mechanism acts as a high-precision filter. It
generates a smaller number of constraints compared to the average of single runs, but with significantly
higher precision when zero-constraint cases are excluded (Table 4). For instance, for forbidden constraints
on bnlearn without descriptions, consensus precision reaches 1. However, this comes at the cost of lower
recall, resulting in a lower F1-score overall compared to the average method. This highlights a trade-o!
between precision and recall: consensus is preferable when the priority is to avoid false positives, while
averaging single runs yields a broader, higher-recall set of constraints.
 Leveraging Large Language Models for Causal Discovery

Figure 10: Interventional Accuracy on bnlearn benchmarks. Bar plots comparing the normalised Structural
Intervention Distance (SID) of our method against baselines across standard benchmark datasets. Error bars
represent standard deviations over 50 repetitions.

• bnlearn vs. CauseNet Datasets: The LLM generally performs better on the bnlearn datasets compared to
the synthetic CauseNet ones, especially shown in terms of F1 in Table 4. This is likely due to the memorisa-
tion of common benchmarks present in the LLM’s training data. The performance on the synthetic datasets,
while lower, provides a more realistic measure of the LLM’s generalisable causal reasoning capabilities, and
shows that the LLM is still capable of generating high-quality constraints for novel problems.

C.4.1 Interaction Between Constraint Quality and Structural Performance

Figures 12 and 13 show heatmaps of the interaction between the F1-score of LLM-derived constraints and data-
derived independence tests, and their impact on the final DAG F1-score. These complement Figure 4 from the
main text by providing a more granular view of how the quality of constraints influences structural learning
performance across di!erent dataset characteristics.
The results indicate that high-quality constraints from the LLM can significantly enhance the performance of
Causal ABA, especially when combined with reliable data-driven tests. The e!ects do not show distinctive trends
across sizes or graph types, as shown in Figures 12 and 13, respectively.
LLMs performance on CauseNet vs bnlearn. Figure 11 shows a side-by-side comparison of the interaction
heatmaps on the CauseNet and bnlearn datasets. Figure 11a is the same as the one in the main text, while
Figure 11b shows the results for the bnlearn benchmarks. We can see that the trends are consistent across both
synthetic and real-world datasets, but the overall accuracy of LLM-derived constraints is higher in the bnlearn
benchmarks, likely due to memorisation e!ects.
Figure 14 provides a breakdown of the interaction heatmap for each individual bnlearn dataset using the same
binning thresholds as the global heatmap in Figure 11b: [0,0.33), [0.33,0.66), [0.66,1]. This is the same as
Figure 12 and 13 for the synthetic datasets. We can notice a much lower diversity of LLM-derived constraint
quality within datasets, with high concentration in one or two bins. This is likely due to the smaller number of
variables in these datasets, which makes it easier for the LLM to memorise the relationships.
Finally, Figure 15 shows the same breakdown but using quantiles for each dataset to define the bins. This
allows us to see more clearly the interaction between constraint quality and structural performance within each
dataset, without being a!ected by the overall accuracy of the LLM on that dataset. We can see that even in
datasets where the LLM show lower overall performance (e.g. CHILD), there are still instances where high-quality
constraints lead to improved structural learning performance.
 Li, Russo

(a) CauseNet synthetic datasets. (b) bnlearn benchmarks.

Figure 11: Interaction between the F1-score of LLM-derived constraints and data-derived independence tests;
color denotes the mean change in the final DAG F1-score.

Figure 12: CauseNet Synthetic: Heatmaps of the interaction between the F1 of LLM-derived and data-derived
constraints, split by number of nodes.

Figure 13: CauseNet Synthetic: Heatmaps of the interaction between the F1 of LLM-derived and data-derived
constraints, split by graph type.
 Leveraging Large Language Models for Causal Discovery

Figure 14: bnlearn Benchmarks: Heatmap of the interaction between the F1-score of LLM-derived constraints
and data-derived independence tests per dataset using the same binning threshold as the global heatmap in
Figure 11b: [0,0.33), [0.33,0.66), [0.66,1].

Figure 15: bnlearn Benchmarks: Heatmap of the interaction between the F1-score of LLM-derived constraints
and data-derived independence tests per dataset using quantiles for each dataset to define the bins.