Early Insights into Argumentation-Guided Causal Evaluation with the Help of LLM

Early Insights into Argumentation-Guided Causal Evaluation with the Help of LLMs Pietro Baroni1,Federico Cerutti1,2,3,*,Massimiliano Giacomin1,Gian Franco Lamperti1and Marina Zanella1 1DII - Universitá di Brescia - Italy 2Cardiff University - UK 3University of Southampton - UK Abstract The rapid growth of Deep Neural Networks (DNNs) has brought substantial advances in artificial intelligence across domains such as vision, language, and recommendation systems. However, this progress comes at a steep energy cost, with model training and deployment contributing significantly to global computational energy consumption. Understanding what drives this energy demand requires more than empirical correlation — it demands causal explanations. In this work, we investigate the causal factors underlying energy use in DNN training, using structure learning algorithms such as the PC algorithm to derive candidate causal graphs. Recognising the limitations of such methods — particularly in terms of assumptions and finite data — we introduce a novel approach to evaluate each inferred link through formal argumentation. We treat each proposed causal relationship as a dialectical object, generating arguments and counterarguments that articulate its plausibility, underlying mechanisms, and possible confounders. We operationalise this reasoning using large language models in a zero-shot prompting setup, surfacing the evidential and conceptual assumptions behind each causal claim. This hybrid approach, combining causal discovery with structured argumentative evaluation, promotes interpretability and critical scrutiny in data-driven causal modelling. Preliminary results demonstrate its potential for rendering causal claims more transparent and contestable. Keywords LLM, Causality, Argumentation 1. Introduction Deep Neural Networks (DNNs) have become foundational to the current landscape of artificial intelligence (AI), enabling advances in fields as diverse as computer vision [ 1 ], natural language processing [ 2 ], personalised recommendation [ 3 ], and speech recognition [ 4 ]. These models are typically trained on large datasets using high-performance GPU clusters, often within large-scale data centres. As a consequence, the growth of deep learning research and applications has been accompanied by a substantial increase in the energy required to train and operate these models [5]. Recent analyses show that this trend, dubbed the “Red AI” era [ 6 , 7 , 8 ], presents a mounting ecological challenge. AI models are growing in size, complexity, and resource demand at an exponential rate. Despite ongoing improvements in chip efficiency and data centre cooling, the energy consumption and carbon emissions associated with training large models continue to double every 4–6 months. A projection [ 9 ] suggests that computing could account for up to 20.9% of global electricity demand by 2030. This includes contributions from both training and inference workloads. To better understand the drivers of such energy consumption, we need to move beyond simple empirical correlations and ask causal questions: What parameters truly influence energy use during model training? Does batch size cause higher energy demand, or is it a proxy for another latent factor? Addressing such questions calls for causal discovery — the process of inferring causal relationships from data. Algorithms such as the PC algorithm (Section 2) and its variants provide a data-driven means to construct causal graphs, identifying conditional independencies and orienting plausible causal directions. AI32025: 9th Workshop on Advances in Argumentation in Artificial Intelligence, September 12-13, 2025, Rende, Italy ©2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). Yet, as with any form of inference, these algorithmic outputs raise as many questions as they answer. Given the statistical limitations of finite data and the assumptions encoded in the discovery procedures, the resulting causal links should not be accepted uncritically. Instead, each proposed causal relationship demands a careful evaluation of its plausibility and justification — not only in light of data, but also in terms of domain knowledge, mechanisms, and conceptual coherence. This paper proposes that tools from formal argumentation theory [ 10 , 11 ] can help operationalise such evaluations (Section 3). In our setting, each candidate causal link is treated as an object of dialectical discussion: we construct arguments in favour of the link (e.g., based on observed regularities, known physical constraints, or plausible mechanisms) and contrast them with counterarguments (e.g., suggesting alternative explanations, questioning generalisability, or highlighting confounding factors). We implement this deliberative process (Section 4) using large language models (LLMs) in a zeroshot prompting setting [ 12 ]. For each candidate edge in a data-derived causal graph, the system generates structured natural language arguments and counterarguments. Rather than discarding links or altering the graph, we aim to surface the implicit reasoning behind each relation — bringing to light the assumptions, analogies, evidence patterns, and objections that might otherwise remain hidden. The contribution of this work is a method for making the structure and justification of causal claims explicit, interpretable, and open to contestation. In doing so, we offer a hybrid approach to causal inference that supplements algorithmic discovery with argumentative reasoning. We discuss our preliminary results in Section 5and address the limitations of the methodology in Section 6. 2. Causal Discovery Causal discovery aims to infer causal relationships from data, typically under structural assumptions. A foundational framework for causal reasoning is provided by Judea Pearl’s theory [ 13 , 14 , 15 ], formalised using directed graphical models and structural equations. 2.1. Structural Causal Models The formal model used in Pearl’s framework is typically referred to as a Structural Causal Model (SCM) or a Directed Graphical Causal Model (DGCM). It combines a directed acyclic graph (DAG) with a joint probability distribution that encodes assumptions about direct causal mechanisms and the independence structure of the system. Definition 2.1 (Structural Causal Model).AStructural Causal Model (SCM) is a tuple ℳ= ⟨𝒢,𝒳,ℱ, 𝑃 (𝑈)⟩where: •𝒢= (𝑉, 𝐸)is a directed acyclic graph (DAG), where each node 𝑋𝑖∈𝑉represents a variable. •𝒳is the set of possible values for the variables. •ℱ={𝑓𝑖: Pa(𝑋𝑖)×𝑈𝑖→𝑋𝑖} is a set of functions, one for each variable 𝑋𝑖 , where Pa(𝑋𝑖) denotes the set of parents of 𝑋𝑖in 𝒢and 𝑈𝑖is an exogenous noise variable. •𝑃(𝑈) = ∏︀𝑖𝑃(𝑈𝑖)is a product distribution over the exogenous variables. Each endogenous variable 𝑋𝑖is determined by 𝑋𝑖:= 𝑓𝑖(Pa(𝑋𝑖), 𝑈𝑖). Definition 2.2 (Intervention and do-calculus).Given a SCM ℳ and a variable 𝑋𝑖∈𝑉 , the effect of an intervention do(𝑋𝑖=𝑥) is a modified model ℳdo(𝑋𝑖=𝑥) where the structural function 𝑓𝑖 is replaced by the constant function 𝑓𝑖:= 𝑥, and the distribution over other variables is adjusted accordingly. Definition 2.3 (Causal Bayesian Network).ACausal Bayesian Network is a pair (𝒢, 𝑃)where: •𝒢is a DAG. •𝑃 is a joint probability distribution over the variables such that the Markov condition holds: each variable is independent of its non-descendants given its parents in the graph. 2.2. D-separation and Faithfulness To read conditional independencies from a graph, we rely on the notion of d-separation. Definition 2.4 (D-separation).Let 𝒢 be a DAG and let 𝑋 , 𝑌 , and 𝑍 be disjoint sets of nodes in 𝒢 . A path between a node in 𝑋and a node in 𝑌is said to be blocked by 𝑍if any of the following holds: • The path contains a non-collider node that is in 𝑍 . A node 𝑣 on a path is a non-collider if the path traverses it via either 𝑢→𝑣→𝑤,𝑢←𝑣←𝑤, or 𝑢←𝑣→𝑤. • The path contains a collider node 𝑣 (i.e., a node where the arrows on the path converge: 𝑢→ 𝑣←𝑤), and neither 𝑣nor any of its descendants are in 𝑍. We say that 𝑋 is d-separated from 𝑌 given 𝑍 if all paths between any node in 𝑋 and any node in 𝑌 are blocked by 𝑍. Definition 2.5 (Causal Markov Assumption).A distribution 𝑃 over a set of variables 𝑉 satisfies the Causal Markov Assumption with respect to a DAG 𝒢 if every variable is conditionally independent of its non-descendants given its parents. Definition 2.6 (Causal Faithfulness Assumption).A distribution 𝑃 is faithful to a DAG 𝒢 if every conditional independence relation that holds in 𝑃is entailed by d-separation in 𝒢. 2.3. Constraint-Based Causal Discovery and the PC Algorithm Under the Causal Markov and Faithfulness assumptions, we can discover aspects of the causal structure by testing for conditional independencies in the observed data. One of the most widely used methods in this class is the PC algorithm [16]. The PC algorithm is sound under the Causal Markov and Faithfulness assumptions, and pointwise consistent under large sample limits and no latent confounding variables. It outputs a Completed Partially Directed Acyclic Graph (CPDAG), which represents the Markov Equivalence Class (MEC) of all DAGs consistent with the observed independencies. Despite its widespread adoption in causal discovery research, the PC algorithm exhibits several important limitations that must be acknowledged. First, it operates under the assumption that there are no latent confounding variables or selection biases in the data. This means that the algorithm presumes all relevant variables influencing the observed relationships have been measured and are included in the analysis. In real-world applications, especially in domains like medicine, economics, or machine learning system diagnostics, such assumptions are often violated. Hidden confounders may induce spurious dependencies or mask true causal links, thereby compromising the validity of the inferred causal structure. Second, the PC algorithm relies critically on the correctness of statistical tests for conditional independence. These tests serve as the foundation for removing edges and orienting v-structures during the graph construction process. However, in finite samples, such tests are prone to both Type I and Type II errors. A Type I error (false positive) occurs when the test incorrectly rejects the null hypothesis of conditional independence when it is in fact true, while a Type II error (false negative) occurs when the test fails to reject the null hypothesis despite the variables being dependent. These risks are exacerbated when conditioning on large sets of variables, which increases variance, or when the dependencies are weak or non-linear, making them harder to detect. As a consequence, the structure returned by the PC algorithm may be unstable or incorrect if the statistical decisions deviate from the population-level conditional independencies. Third, the algorithm cannot distinguish between different directed acyclic graphs (DAGs) that belong to the same Markov Equivalence Class (MEC). By design, it returns a Completed Partially Directed Acyclic Graph (CPDAG), which encodes a set of DAGs that share the same set of d-separation relations. While this is theoretically sound, it limits the informativeness of the output: many edges may remain undirected, and causal directionality is left ambiguous without additional assumptions or interventional data. Finally, and perhaps most crucially from the perspective of formal argumentation, the PC algorithm is entirely syntactic in its use of statistical dependencies; it provides no account of the semantic content of the edges in the resulting graph. Each link is treated purely as an artefact of probabilistic association (or lack thereof) rather than as a meaningful causal hypothesis embedded in domain knowledge. This poses a substantial challenge in applications where interpreting, justifying, or contesting individual links is essential. Therefore, beyond purely algorithmic discovery, there is a pressing need for frameworks that allow for the articulation and critical evaluation of causal claims — link by link — using structured arguments, explanations, and counter-arguments. Such argumentative approaches can complement statistical discovery methods by making the assumptions, interpretations, and domain-specific implications of each inferred causal relationship explicit and debatable. 3. A Structured Account of Arguments For and Against Causal Claims Causal arguments are central in many domains, from scientific inquiry to policy and engineering, where actors seek to justify why one event (the effect) follows from another (the cause). This section provides a structured typology of argumentative forms both in support of and against causal claims, building on top of [10,11]. 3.1. Arguments in Favour of Causal Claims Arguments supporting causal links can be grouped by the type of justification they provide. We distinguish three principal classes: circumstantial evidence, contrastive evidence, and causal explanations. 3.1.1. Circumstantial Evidence Circumstantial evidence relies on regularities or proximity relations between two events, without offering a mechanistic explanation of counterfactual analysis. The reasoning typically takes the form: “ 𝐴 caused 𝐵 because 𝐴 regularly precedes 𝐵 , co-occurs with 𝐵 , or resembles other known causes of 𝐵.” We distinguish several sub-types of circumstantial evidence: Spatio-temporal contiguity: The cause and effect occur in close spatial or temporal proximity. This supports the intuition that the proximity of events may suggest a causal link. Repeated co-occurrence: The purported cause and effect consistently appear together across multiple instances. This statistical regularity, though not sufficient for causation, can signal a potentially robust association worth investigating. Analogical similarity: The situation under analysis resembles other known causal scenarios. If 𝐴 and 𝐴′ share relevant features, and 𝐴′ is known to cause 𝐵 , then by analogy, 𝐴 might be assumed to cause 𝐵as well. These forms of evidence are inherently speculative and typically serve as initial heuristics to guide hypothesis generation or further empirical testing. 3.1.2. Contrastive Evidence Contrastive evidence draws on observed differences in outcomes across varying conditions. Such arguments typically follow the structure: “𝐴caused 𝐵because 𝐵occurs under 𝐴but not under ¬𝐴.” We identify several sub-types of contrastive evidence: Statistical covariation: A measurable difference in outcomes is observed between groups or conditions, and this difference persists even after adjusting for potential confounding variables. The contrast is interpreted as supporting a causal role for the varying factor. Before-after comparison: An intervention or change is introduced, and a corresponding shift in outcomes is observed. If other factors remain stable, the contrast in outcome is attributed to the intervention. Controlled experiment: All conditions are held constant except for the variable of interest. A consistent difference in outcome across conditions is then attributed to the manipulated variable. 3.1.3. Causal Explanation Causal explanations articulate a mechanism that connects the cause to the effect. These arguments are typically stronger than purely correlational or contrastive forms due to their explanatory depth. They are often structured as: “𝐴causes 𝐵because 𝐴initiates a sequence of intermediate steps leading to 𝐵.” We distinguish several sub-types of causal explanation: Mechanistic explanation: The argument identifies a specific sequence of processes or interactions through which the cause produces the effect. This often involves reference to known physical, computational, or biological mechanisms. Elimination of alternatives: A causal claim is supported by ruling out other plausible explanations. If the observed effect coincides only with changes in 𝐴 , and other variables are held constant, 𝐴 is inferred to be the cause. Typicality of effect: The observed outcome matches the expected pattern associated with similar causes in comparable contexts. This reinforces the plausibility of the proposed mechanism. 3.2. Arguments Against Causal Claims Arguments challenging causal claims fall into two categories: those questioning the plausibility of the causal relation and those attacking the logical structure or sufficiency of the justification. 3.2.1. Plausibility Challenges These arguments suggest that the proposed causal link is implausible in light of available evidence. We distinguish several sub-types: Wrong temporal order: The effect is observed before the supposed cause. Since causes must precede their effects, this undermines the causal interpretation. No connection: The proposed cause and effect belong to unrelated domains, or no plausible pathway links them. Without a credible mechanism, the causal claim lacks support. Free decision: The outcome results from an independent choice or intervention that is not causally determined by the proposed factor. The cause is incidental rather than explanatory. Insufficient cause: The proposed factor occurs without reliably producing the effect. This suggests that it alone cannot account for the outcome and may require additional conditions. Unnecessary cause: The effect can be fully explained by other causes. The proposed factor is therefore not needed to account for the outcome, weakening its causal relevance. 3.2.2. Logical Objections Logical objections target the inferential structure of a causal argument, highlighting weak reasoning or offering superior alternatives. We distinguish several sub-types: Alternative cause: The observed association between the proposed cause and effect can be better explained by a third, unaccounted-for variable that influences both. Post hoc fallacy: The argument infers causality merely from temporal succession—assuming that because 𝐵followed 𝐴,𝐴must have caused 𝐵—without further justification. Low statistical support: The effect is observed inconsistently or weakly across instances. A low base rate or limited correlation challenges the robustness of the causal claim. Anecdotal evidence: The argument relies on a single or highly atypical case, which is insufficient for generalisation and vulnerable to noise or confounding. Unknown mechanism: No explanation is given for how the proposed cause leads to the effect. Without a plausible mechanism, the claim remains speculative. 3.2.3. Qualifying Causal Claims Some arguments do not reject the existence of a causal link but instead refine or constrain its interpretation. These qualifications help clarify the nature, strength, or context of the causal relationship. We distinguish several sub-types: Partial cause: The proposed factor contributes to the outcome but is not the sole or primary cause. Other influences play a more significant role. Indirect cause: The effect arises through a chain of intermediate steps rather than a direct influence. The causal link is mediated by other variables. Common cause: Both the proposed cause and the observed effect result from a shared underlying factor. The deeper cause lies elsewhere. Interaction: The proposed factor produces the effect only in combination with other conditions. On its own, it may have little or no causal impact. Reversed causality: The direction of influence is the opposite of what is claimed. What is presented as the cause is actually a response to the effect. Accidental cause: The observed causal link is coincidental, resulting from an unrelated or unanticipated event that occurred simultaneously. 4. Argumentative Evaluation of Tentative Causal Links via Language Models Once tentative causal relationships have been identified using structure learning algorithms such as the PC algorithm (see Section 2), we proceed to critically evaluate these links using argumentative reasoning. To this end, we leverage large language models (LLMs) to generate structured arguments in favour of or against each causal claim, based on the typology of causal argumentation patterns introduced earlier (cf. Section 3and [10,11]). The prompts we employ are explicitly designed to elicit responses that map onto macro-families of causal argumentation — circumstantial evidence,contrastive evidence, and causal explanation for supporting claims, and plausibility challenges,logical objections, and qualifications for opposing ones. These prompt templates are provided in Listings 1and 2. Listing 1 Prompt Template: Argue in Favour of a Causal Link 1Given the causal claim: [INSERT CAUSAL CLAIM], 2produce a structured argument *in support* of this claim. Choose only one among the following argumentation strategies: 3 41. Circumstantial Evidence: 5Provide at least one argument using spatio-temporal contiguity, repeated cooccurrence, or similarity to known causes. 6 72. Contrastive Evidence: 8Provide at least one argument based on statistical covariation, before-after comparison, or a controlled experiment. 9 10 3. Causal Explanation: 11 - Provide at least one mechanistic explanation, argument from no alternative explanation, or an argument based on typical effect. 12 13 Each argument must: 14 - Clearly state which macro-family and specific subtype it belongs to. 15 - Be logically self-contained and persuasive. 16 - Refer explicitly to the observed or hypothesised phenomena. 17 - Be concise. Listing 2 Prompt Template: Argue Against a Causal Link 1Given the causal claim: [INSERT CAUSAL CLAIM], 2produce a structured argument *against* this claim. Choose only one among the following argumentation strategies: 3 41. Plausibility Challenges: 5Include at least one argument based on wrong temporal order, no plausible connection, free decision, 6insufficient cause, or unnecessary cause. 7 82. Logical Objections: 9Include at least one argument from alternative cause, post hoc fallacy, low statistical support, 10 anecdotal evidence, or unknown mechanism. 11 12 3. Qualifying Claims: 13 - Optionally include partial, indirect, common cause, interaction, reversed causality, or accidental cause qualifications. 14 15 Each argument must: 16 - Identify the macro-family and the specific argumentation subtype. 17 - Be articulated as a counterpoint to a potential or actual supporting argument. 18 - Indicate whether the causal claim is to be rejected, weakened, or reformulated. 19 - Be concise. For instance, suppose the PC algorithm outputs the link batch_size →power . We then instantiate the favour prompt (see Listing 1) as: “Given the causal claim: ’Large batch sizes cause increased GPU power consumption’, produce a structured argument in support of this claim...” This would result in the LLM producing arguments such as: Listing 3 Prompt Template: Final Judgement on a Causal Link 1You are given two arguments regarding the same causal claim. 2 3Causal Claim: 4[INSERT CAUSAL CLAIM] 5 6Argument in Favour: 7[INSERT PRO-CASUAL ARGUMENT HERE] 8 9Argument Against: 10 [INSERT COUNTER-ARGUMENT HERE] 11 12 Task: 13 Evaluate the overall credibility of the causal claim based on the two arguments. Your judgment must: 14 15 1. Reference the strength and relevance of the macro-family and subtype for each argument. 16 2. Indicate whether the causal claim should be: 17 - Accepted as likely 18 - Tentatively accepted with caveats 19 - Undecided or requiring more evidence 20 - Rejected 21 3. Justify your decision with explicit reference to the comparative argumentative strength (\eg directness, generalisability, mechanistic plausibility, alternative explanations). 22 4. Avoid introducing new arguments - focus only on evaluating the two provided. 23 24 Output format: 25 - Summary judgment (one sentence) 26 - Justification (3-5 sentences) • Circumstantial evidence (repeated co-occurrence): “In over a dozen training configurations, large batch sizes were consistently associated with high energy usage.” • Contrastive evidence (controlled experiment): “When only the batch size was varied, all other factors held constant, energy draw increased with larger batches.” • Causal explanation (mechanistic): “Larger batch sizes lead to greater parallelisation, saturating GPU resources and increasing power draw.” This argumentative evaluation serves two purposes. First, it adds interpretability and justifiability to data-driven causal claims. Second, it exposes weak or unsupported inferences by juxtaposing strong and weak forms of justification within the same causal hypothesis. To complement the generation of individual arguments for and against causal hypotheses, we introduce a final reasoning stage where an LLM is prompted to adjudicate between competing views. This dialectical evaluation phase simulates the structure of critical discussion in causal reasoning. Given one argument supporting a causal link and one refuting it, the model is tasked with determining the overall plausibility of the claim. The prompt is designed to ensure that the response explicitly references both the argumentative structure and the relative strength of the opposing macro-families. The complete template is shown in Listing 3. This stage enforces a structured comparison. The LLM is instructed to identify the macro-family and subtype of each input argument, evaluate their logical and evidential strength, and conclude with one of four options: the causal claim is (i) accepted as likely, (ii) tentatively accepted with caveats, (iii) undecided and requiring more evidence, or (iv) rejected. The rationale behind this design is to integrate argumentative quality into the causal assessment pipeline, preventing premature commitment to claims generated by data-driven discovery alone. Consider again the hypothesis that “large batch sizes cause increased GPU power consumption”. Suppose we input the following arguments: • Argument in favour (Contrastive evidence - controlled experiment): “When all training parameters were kept constant except for batch size, power draw increased with larger batches.” • Argument against (Logical objection - alternative cause): “Both batch size and energy usage correlate with model complexity, which is the more likely common cause.” The LLM, following the prompt in Listing 3, may respond: Summary judgment: Tentatively accepted with caveats. Justification: The argument in favour is based on a controlled experiment, which provides strong contrastive evidence and isolates batch size as a plausible causal factor. The counterargument rightly identifies a potential confound (model complexity), but does not negate the observed contrastive pattern. While further control for model size is advisable, the causal link is tentatively supported. 5. Preliminary Results 5.1. Dataset: The BUTTER-E Benchmark To support the empirical evaluation of causal discovery methods in energy-sensitive machine learning pipelines, we employ the BUTTER-E dataset [ 17 , 18 ], a large-scale benchmark of real-world energy consumption measurements associated with Deep Neural Network (DNN) training. Specifically, BUTTER-E extends the primary sweep of the original BUTTER dataset by adding detailed energy usage data collected on both CPU and GPU hardware. The resulting dataset comprises over 41,000 distinct training runs of fully connected multi-layer perceptrons (MLPs), each corresponding to a unique configuration of hyperparameters. The training runs in BUTTER-E span a diverse space of model sizes, shapes, and depths. The number of trainable parameters ranges from 25 to 225 , distributed across 2 to 20 hidden layers. Eight architectural shapes were explored, including: •rectangle and rectangle_residual (uniform width, with or without residual connections), •trapezoid (width decreases linearly with depth), •exponential (exponential decay in width), • and wide_first_𝑛x configurations with 𝑛∈ {2,4,8,16} , where the first hidden layer is 𝑛 times wider than the subsequent layers. The training datasets were selected from the Penn Machine Learning Benchmark (PMLB) repository [ 19 , 20 ], and include: 201_pol , 294_satellite_image , 529_pollen , 537_houses , adult , banana,connect_4,MNIST,nursery,sleep,splice, and wine_quality_white. The energy measurements were obtained by re-executing the primary sweep of BUTTER on the Eagle high-performance computing (HPC) cluster at NREL. Each run was assigned to a dedicated CPU or dualGPU compute node. CPU training was executed using Intel’s OneDNN-accelerated TensorFlow, while GPU training used cuDNN and TensorFlow with NVIDIA-recommended settings. Instantaneous power consumption was recorded at one-minute intervals using the Hewlett-Packard Enterprise Integrated Lights-Out (iLO) monitoring system embedded in each node.