Paper deep dive

Rules or Weights? Comparing User Understanding of Explainable AI Techniques with the Cognitive XAI-Adaptive Model

Louth Bin Rawshan, Zhuoyu Wang, Brian Y Lim

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 144

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%

Last extracted: 7/20/2026, 3:53:29 PM

Summary

The paper introduces CoXAM, a Cognitive XAI-Adaptive Model that uses computational rationality to simulate human reasoning when interpreting XAI techniques (Weights, Rules, and Hybrid). It compares forward and counterfactual decision tasks, finding that CoXAM aligns better with human behavior than baseline ML proxies, and that Rules are harder to recall than Weights, while counterfactual tasks are inherently more difficult.

Entities (9)

CoXAM · model · 98%Counterfactual Simulation · task · 95%Forward Simulation · task · 95%Mushroom Dataset · dataset · 95%Wine Quality Dataset · dataset · 95%XAI Schemas · concept · 92%Computational Rationality · theory · 90%Decision Tree · algorithm · 85%Linear Regression · algorithm · 85%

Relation Signals (6)

CoXAM → uses → Computational Rationality

confidence 95% · CoXAM employs computational rationality to choose among reasoning processes based on the trade-off in utility and reasoning time

Counterfactual Simulation → isharderthan → Forward Simulation

confidence 92% · The model successfully replicated and explained several key empirical findings, including that counterfactual tasks are inherently harder than forward tasks

Decision-tree rules → ishardertorecallthan → Linear Weights

confidence 90% · decision tree rules are harder to recall and apply than linear weights

CoXAM → simulates → Human Decision-Making

confidence 90% · CoXAM demonstrated a stronger alignment with human decision-making compared to baseline machine learning proxy models.

CoXAM → outperforms → Baseline ML Proxy Models

confidence 88% · CoXAM demonstrated a stronger alignment with human decision-making compared to baseline machine learning proxy models.

XAI Helpfulness → dependson → Application Data Context

confidence 85% · the helpfulness of XAI depends on the application data context

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Rules and Weights are popular XAI techniques for explaining AI decisions. Yet, it remains unclear how to choose between them, lacking a cognitive framework to compare their interpretability. In an elicitation user study on forward and counterfactual decision tasks, we identified 7 reasoning strategies of interpreting three XAI Schemas - weights, rules, and their hybrid. To analyze their capabilities, we propose CoXAM, a Cognitive XAI-Adaptive Model with shared memory representation to encode instance attributes, linear weights, and decision rules. CoXAM employs computational rationality to choose among reasoning processes based on the trade-off in utility and reasoning time, separately for forward or counterfactual decision tasks. In a validation study, CoXAM demonstrated a stronger alignment with human decision-making compared to baseline machine learning proxy models. The model successfully replicated and explained several key empirical findings, including that counterfactual tasks are inherently harder than forward tasks, decision tree rules are harder to recall and apply than linear weights, and the helpfulness of XAI depends on the application data context, alongside identifying which underlying reasoning strategies were most effective. With CoXAM, we contribute a cognitive basis to accelerate debugging and benchmarking disparate XAI techniques.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

142,257 characters extracted from source content.

Expand or collapse full text

by Rules or Weights? Comparing User Understanding of Explainable AI Techniques with the Cognitive XAI-Adaptive Model Louth Bin Rawshan Department of Computer ScienceNational University of SingaporeSingaporeSingapore e0546303@u.nus.edu , Zhuoyu Wang Department of Computer ScienceNational University of SingaporeSingaporeSingapore wang.zhuoyu@u.nus.edu and Brian Y Lim Department of Computer ScienceNational University of SingaporeSingaporeSingapore brianlim@nus.edu.sg (2026) Abstract. Rules and Weights are popular XAI techniques for explaining AI decisions. Yet, it remains unclear how to choose between them, lacking a cognitive framework to compare their interpretability. In an elicitation user study on forward and counterfactual decision tasks, we identified 7 reasoning strategies of interpreting three XAI Schemas—weights, rules, and their hybrid. To analyze their capabilities, we propose CoXAM, a Cognitive XAI-Adaptive Model with shared memory representation to encode instance attributes, linear weights, and decision rules. CoXAM employs computational rationality to choose among reasoning processes based on the trade-off in utility and reasoning time, separately for forward or counterfactual decision tasks. In a validation study, CoXAM demonstrated a stronger alignment with human decision-making compared to baseline machine learning proxy models. The model successfully replicated and explained several key empirical findings, including that counterfactual tasks are inherently harder than forward tasks, decision tree rules are harder to recall and apply than linear weights, and the helpfulness of XAI depends on the application data context, alongside identifying which underlying reasoning strategies were most effective. With CoXAM, we contribute a cognitive basis to accelerate debugging and benchmarking disparate XAI techniques. Explainable AI, cognitive modeling, user study †journalyear: 2026†copyright: c†conference: 31st International Conference on Intelligent User Interfaces; March 23–26, 2026; Paphos, Cyprus†booktitle: 31st International Conference on Intelligent User Interfaces (IUI ’26), March 23–26, 2026, Paphos, Cyprus†doi: 10.1145/3742413.3789147†isbn: 979-8-4007-1984-4/2026/03†submissionid: 5230†ccs: Human-centered computing Empirical studies in HCI†ccs: Computing methodologies Artificial intelligence 1. Introduction AI systems increasingly influence high-stakes decisions in finance, law, healthcare, and criminal justice, making explainable AI (XAI) essential for exposing a model’s decision process to end users. Although numerous XAI techniques have been developed (Mersha et al., 2024; Schwalbe and Finzel, 2024), popular explanations take the form of linear models (Ribeiro et al., 2016; Lundberg and Lee, 2017; Poursabzi-Sangdeh et al., 2021) or rules (Breiman et al., 1984; Ribeiro et al., 2018; Lakkaraju et al., 2016a). Despite their simplicity, these methods remain effective in critical domains such as healthcare (Caruana et al., 2015) and finance (Yeo et al., 2025b; Caterson et al., 2024). Perhaps this is because these structures align well with human reasoning schemas of factor weighting (Memelink and Hommel, 2013), and rule-based categorization (Goodman et al., 2008). Several user studies have compared the effectiveness of these explanation methods, but a consensus remains elusive, reporting conflicting results that linear models (Hase and Bansal, 2020) or rules (Ribeiro et al., 2018) are more helpful. We argue that a deeper examination of human cognition can help to elucidate when and why different XAI Schemas may be useful or misused. We draw inspiration from the successes of cognitive modeling to understand people and improve user experience and performance. Applications include: intelligent tutoring systems (Corbett and Anderson, 1994; Koedinger et al., 1997), programming tutors that infer novice error patterns (Piech et al., 2015; Rivers, 2017) and touchscreen typing (Jokinen et al., 2021; Shi et al., 2024). We argue that cognitive modeling of how users understand XAI can help developers to analyze the users’ reasoning strategies, and ultimately serve as a platform for rapid prototyping and early evaluation of XAI on simulated users. In this work, we focus on tabular data, where attributes are structured and semantically meaningful. Tabular data underpins many real-world applications in healthcare, finance, and operations, and remains a core setting for machine learning in practice (Caruana et al., 2015; Borisov et al., 2022; Grinsztajn et al., 2022). We use global explanations instead of instance-based explanations, since people use the same rules globally rather different rules per instance (Bruner et al., 1956; Shepard et al., 1961; Gigerenzer and Goldstein, 1996b; Maddox and Ashby, 2004). In this context, we study how users interpret two XAI Schemas—weight-based factors (Weights) and rule-based decision trees (Rules)—to understand the AI model’s decisions via forward and counterfactual simulation tasks111These tasks assess whether participants can anticipate an AI prediction (forward) and identify which attribute to change to reach an alternative target prediction (counterfactual) (Doshi-Velez and Kim, 2017).. From a formative user study with 24 participants, we elicited reasoning strategies, three for forward simulation (approximate calculation (Weights), feature attribution (Weights), attribute-threshold-attribute traversal (Rules)), and four for counterfactual simulation (inverse calculation (Weights), inverse feature attribution (Weights), root/leaf node threshold crossing (Rules), availability heuristic (any)). We propose CoXAM, a Cognitive XAI-Adaptive Model to represent a user who can learn and use linear-factor and rule-based explanations for AI simulation tasks. It is XAI-Adaptive by using a shared memory framework to encode attribute values, linear factors and rule nodes as chunks. Reasoning strategies for forward simulation were implemented with drift-diffusion decision dynamics (Ratcliff, 1993; Ratcliff et al., 2016) to model the decision probability and execution time, explicit heuristics determined from user reasoning steps, and several cognitive parameters to model human bounded rationality (e.g., diffusion noise, retrieval threshold). Assuming that each user could use any reasoning strategy, CoXAM chooses the ”best” strategy with computational rationality (Oulasvirta et al., 2022), based on the trade-off between the utility of the decision likely being correct and decision speed. We conducted a summative user study, across two application scenarios (Wine Quality, Mushroom Edibility), with 340 participants performing forward and counterfactual simulation tasks with Weights explanation, Rules explanation, or their Hybrid222We further introduce a Hybrid XAI condition that interleaves Weights and Rules explanations to test users’ ability to integrate or switch between reasoning styles. over 27600 trials. We evaluated CoXAM with the study data. We found that Weights explanation was most helpful for the Wine Quality scenario which had linear attributes, and Rules explanations for Mushroom which was more nonlinear. Unlike Weights explanation, Rules explanation was harder to recall than read, resulting in significantly lower forward simulation accuracy. User accuracy on counterfactual tasks was worse than on forward tasks, due to the inverse reasoning needed. These results were mostly replicated by CoXAM, which had the best fit to human decision responses compared to baseline proxies (KNN, Decision Tree, Linear Regression, SHAP) with the lowest NLL and BIC scores. Deeper analysis of reasoning strategies found that approximate calculation performed worse than feature attribution for forward tasks when values were larger, and availability heuristic was least effective for counterfactual tasks. Our contributions are: (1) Elicitation of user reasoning strategies for Rules and Weights explanations. (2) Evaluation of their (in)effectiveness for forward and counterfactual simulation tasks. (3) XAI-adaptive cognitive model of different XAI schemas for different decision tasks. CoXAM serves as a testbed for user behavior in tabular settings, allowing comparisons of Rule vs. Weights explanations while identifying the strategies users rely on. We include discussions on extensions to broader XAI families and tasks. 2. Related Work Our main focus is to investigate and model human reasoning in Rule-based and Weight-based XAI. We explore related work that evaluate and contrast rule- and weight-based XAI Schemas, examine user modeling in XAI domains, and draw upon cognitive modeling in more general domains. 2.1. Rules vs. Weights: A Core Debate A central question in tabular XAI is whether rule explanations (e.g., decision trees, decision sets, anchors) or weight explanations (e.g., linear models, local attributions) better support users. This debate predates XAI: classic judgment research documents strong performance for additive cue combination—including simple or equal-weight schemes—relative to informal rule formulations (Meehl, 1954; DAWES, 1979; Einhorn and Hogarth, 1975; Dawes and Corrigan, 1974). In parallel, work on compact, condition-based decision procedures showed that short, explicitly stated rules can rival regression under limited time or information (Gigerenzer and Goldstein, 1996a; Czerlinski et al., 1999). Modern XAI inherits this split and yields mixed head-to-head findings: Anchors (rules) improved precision and reduced effort in some forward simulation tasks (Ribeiro et al., 2018), whereas broader studies reported LIME (weights) improving simulatability over rule-based alternatives (Hase and Bansal, 2020). For counterfactual prediction, one study found no clear advantage for either family (Hase and Bansal, 2020) while another observed benefits for linear/weight information in a text-editing setting (Arora et al., 2022). A large-N comparison reported decision trees (rules) as more locally interpretable than logistic regression (weights) under their task representation (Slack et al., 2019). Overall, the literature does not support a universal winner. Outcomes appear sensitive to task framing (forward vs. counterfactual) and dataset properties. This background motivates examining, within concrete tasks, what information each form conveys and how users apply that information in different tasks. 2.2. Modeling User Performance on XAI Several works have modeled human decisions in XAI, but typically with the aim of capturing behavioral trends, often neglecting underlying cognitive processes. This behavioral modeling is commonly achieved by training supervised machine learning models with user decision data to simulate human-like responses (Wang et al., 2022; Virgolin et al., 2020; Lage et al., 2018; Hilgard et al., 2021; Reichman et al., 2024). Recent works have further differentiated these modeling techniques: Ma et al. (Ma et al., 2023) trained a personalized and refinable decision tree based on user choices, while Mozannar et al. (Mozannar et al., 2022) utilized user feedback to model human rejection of AI erroneous suggestions. Conversely, Chen et al. (Chen et al., 2022) demonstrated that a decision proxy trained exclusively on AI explanations, without direct human input or labels, could still achieve reasonable alignment with human decisions. Fiori et al. (Fiori et al., 2024) describe that human understanding of natural language explanations could be loosely approximated by using large-language models as a proxy for the human. Unlike prior work that focuses on matching behavioral data, our work seeks to model the underlying human cognitive processes in consuming XAI. This would enable deeper insights into which strategies users adopt and how their reasoning is affected by XAI schema. We focus on AI-assisted decision making (Ma et al., 2023) where the user decides whether to delegate decisions to the AI or decide for themselves, instead of human-AI collaboration (Lai et al., 2022; Vaccaro et al., 2024) where the AI complements the user’s knowledge to support effective teaming. 2.3. Cognitive Modeling for Human Behavior Cognitive modeling has long been used to understand and predict human decision-making across domains such as user interface performance (John et al., 2004), robotics (Baron et al., 1994), autonomous driving (Kolekar et al., 2021; Choi et al., 2021; Bhattacharyya et al., 2022), the study of psychological disorders (Zeng et al., 2024), and agent-based simulations of collective behavior (Dobson et al., 2019; Kennedy, 2011). Traditional approaches often rely on data-driven imitation learning (Le Mero et al., 2022) or machine theory of mind (Rabinowitz et al., 2018), which learn to reproduce human behavior from large datasets but typically lack explicit representations of underlying cognitive processes. Reinforcement learning (Sutton, 2018; Cao and ZhiMin, 2019) can model sequential decision-making and adaptation, yet, by itself, it rarely accounts for intrinsic human limitations such as memory decay, attentional bottlenecks, or time–effort trade-offs (Fuchs et al., 2023). Recent work in cognitive science and HCI has instead moved toward computational rationality (Oulasvirta et al., 2022; Lieder and Griffiths, 2020), which extends the classic notion of bounded rationality (Simon, 1957) by describing human cognition as an optimization process constrained by limited computational resources. Under this view, people act approximately rationally—not to maximize objective accuracy or reward, but to optimize performance relative to their internal costs of computation, memory, and attention. Deviations from normative rationality are thus interpreted as adaptive trade-offs rather than biases or errors. Computational rationality provides a unifying lens for modeling cognition in interactive and decision-making contexts. In HCI, it has been used to explain how users allocate attention and choose interaction strategies that balance accuracy and effort in tasks such as touchscreen typing (Shi et al., 2024), menu search (Oulasvirta et al., 2022), and pedestrian decision-making under uncertainty (Wang et al., 2025). Within cognitive science, the framework has informed rational models of strategy selection and meta-reasoning (Lieder and Griffiths, 2017), demonstrating how people decide which cognitive strategy to deploy given resource constraints. In the context of XAI, computational rationality provides a principled way to describe user behaviors as adaptive: users aim to make sense of explanations efficiently, balancing the cognitive effort of understanding against the potential gain in decision quality. This framing connects variations in user performance with the underlying constraints of human cognition, without assuming that one explanation style is inherently superior. 3. Study Approach and Background To understand why explanations are effective (or not), we aim to simulate human reasoning when learning from and using two common XAI Schemas—Rules and Weights 333Note that the underlying surrogate model is logistic regression, but we call it ”Weights”, to illustrate the type of information shown to users.—by developing a cognitive model of user reasoning strategies. We focus on tabular data because it provides structured, discrete attributes and remains an integral component of many real-world applications in healthcare, finance, and operations (Caruana et al., 2015; Borisov et al., 2022; Grinsztajn et al., 2022). Our choice of using global explanations instead of instance-based ones is motivated by the natural process of reasoning with rules, where people tend to apply the same rules broadly rather than forming different rules for each instance (Bruner et al., 1956; Shepard et al., 1961; Gigerenzer and Goldstein, 1996b; Maddox and Ashby, 2004). We describe our overall approach, and introduce background concepts and context. 3.1. User-Centered Design for Cognitive Modeling To simulate human interpretation of different XAI Schemas, we have to identify cognitive processes and reasoning strategies from actual users, encode this in a cognitive model, and evaluate against real user decision data. Hence, we conducted three studies (illustrated in Fig. 1): I) Formative user study to elicit and characterize various reasoning strategies for Rule-based and Weight-based XAI interpretation. I) Modeling study to develop a cognitive model and evaluate it with baseline models for predictions of user (mis)interpretations, which may result in good or poor decisions and performances. I) Summative user study to collect data on user decisions on a large scale to associate with each reasoning strategy, and calibrate and evaluate our modeling. Figure 1. Overall approach to model users in XAI understanding across three studies: I) Formative study to elicit reasoning strategies, I) Modeling study to examine user reasoning, and I) Summative study to observe decisions. This figure illustrates the overall workflow of the research approach. The XAI Schemas inform two empirical components: a Formative Study that identifies participants’ reasoning strategies, and a Modeling Study (CoXAM) that formalizes these strategies to simulate user reasoning. The resulting model is then evaluated in a Summative Study using user decision data from forward and counterfactual tasks. 3.2. XAI Schemas: Rule-based Decision Tree and Weight-based Linear Regression Weight-based and rule-based explanations remain among the most popular explanation schemas of XAI methods for lay users. With weight-based explanations, users are informed which attributes are more influential toward a decision. However, the explanation is not actionable to know how to change a decision. Instead, rules (if-then) indicate explicit thresholds or antecedents when the result will change. Yet, it remains unsettled which XAI Schema is more interpretable for users. We focus on applying these XAI Schemas for surrogate global explanation and study the use of linear regression for Weights (Bo et al., 2024; Poursabzi-Sangdeh et al., 2021), and decision tree for Rules (Kozielski et al., 2025; Angelino et al., 2018a). Furthermore, we study the combined use of both Weights and Rules to investigate if the consolidated learning is beneficial. Through our formative study, we will identify reasoning strategies that users employ for each XAI Schema, and examine their efficacy and investigate why. To convey the instance attributes and decision task, we employed a tabular user interface (UI), similar to Bo et al. (Bo et al., 2024). Fig. 2 (a–b, e–f) shows the baseline components of the attribute values and prediction labels from the AI and XAI models. Fig. 2 (c–d) shows the Weights explanation from the Linear Regression XAI model. Fig. 3 (a–b) shows the Rules explanation from the Decision Tree XAI model. Figure 2. Weight-based XAI with Tabular UI components for: a) Attribute names of the instance. b) Value of each attribute with bar to indicate how high/low. c) Factors to multiply with each attribute value. d) Partial contributions of each attribute based on value x factor. e) Prediction of Weights explanation for current instance. f) Prediction of AI. a–b are always shown; c is only shown in trials with Weights explanation; d, e, f is not shown for decision testing. This figure shows the user interface for the Weights XAI condition in the tabular instance task. The interface presents six attributes with their corresponding values and visual bars indicating magnitude. For each attribute, the model displays its linear coefficient (Factor) and the resulting contribution (Partial Sum), computed as value × factor. The overall Factors prediction is displayed on the right alongside the AI’s predicted class label. Panels (a)–(b) are always visible, while panels (c)–(f) appear only when Weights explanation is provided. Figure 3. Rule-based XAI with Decision Tree UI for a) low (wine quality) and b) high complexity (mushrooms). The black outline highlights the path that corresponds to the current instance values. This figure shows the Rules explanation user interface used in the experiment for two example datasets. Panel (a) illustrates a low-complexity tree from the Wine Quality task, while panel (b) depicts a higher-complexity tree from the Mushrooms task. Each node represents an attribute and its decision threshold, with terminal leaves showing the predicted class labels. The black-highlighted path indicates the traversal corresponding to the current instance’s feature values, helping users visualize how the AI reaches its prediction. 3.3. Forward and Counterfactual Simulation Decision Tasks We focus on decision tasks that are commonly supported by XAI. Doshi-Velez and Kim (Doshi-Velez and Kim, 2017) defined human simulation tasks to assess users’ understanding of an AI model’s behavior. In this work, we focus on forward simulation, where the user estimates (simulates) the AI’s prediction from the instance’s attribute values and any accompanying XAI explanation; and counterfactual simulation, where the user changes an attribute to yield a different AI prediction. These tasks evaluate whether users have formed an accurate mental model of the AI’s decision logic, and their understanding of how specific attributes contribute to changes in the outcome. Formally, given an input =(x1,…,xn)⊤x=(x_1,…,x_n) , the AI predicts a binary classification (1) y^=fAI(), y=f_AI(x), where y^∈−1,1 y∈\-1,1\. This can be explained by an XAI model with human-interpretable prediction: (2) y~=gXAI(,y^,fAI). y=g_XAI(x, y,f_AI). In the forward simulation task, the user estimates the AI’s prediction based on the instance’s attributes and/or the explanation: (3) y˘=fAIuser(,gXAI), y=f_AI^user(x,g_XAI), where ideally y˘≈y y≈ y or y˘≈y~ y≈ y, indicating that the user’s simulated prediction aligns with the AI or XAI model respectively. For the counterfactual simulation task, the user modifies the value of an attribute r in x by Δxr x_r to obtain a new instance ′x , and the AI produces a new prediction: (4) ˘′ x =hCFuser(,y^,gXAI), =h_CF^user(x, y,g_XAI), (5) y^′ y =fAI(˘′). =f_AI( x ). A successful counterfactual ˘′ x should lead to the target AI outcome, i.e., y^′≠y y ≠ y. For this work, for user simplicity, we limit our scope to only changing one attribute in x. 3.4. Evaluation Context: Application Scenarios, Datasets, and AI Model For accessibility to lay users, we chose two publicly available datasets: the UCI Mushroom dataset (Wagner et al., 2021) (N = 61,068) featuring physical attributes for poisonous/edible classification and the Wine Quality dataset (Paulo et al., 2009) (N = 4,898), binarized for a low/high quality classification task based on chemical compositions. We limited both datasets to only 6 attributes. For both datasets, we trained a simple Multi-Layer Perceptron (MLP) model to predict the respective classification targets: wine quality and mushroom edibility. The models were trained on 80% of the data and tested on 20%, resulting in 85.0% test accuracy for the Wine Quality dataset and 78.2% for the Mushrooms dataset. 4. Formative Study of User Reasoning on XAI We conducted a formative user study to elicit user reasoning strategies and mental models when using or recalling different XAI Schemas (Weights or Rules) for the forward and counterfactual simulation tasks (Task Type), under varying test conditions (with or without XAI). 4.1. Experiment Method and Procedure Each participant was randomly assigned to one Application Scenario (between-subject: Mushroom, or Wine Quality). To reduce bias and the impact of prior knowledge, we anonymized the prediction labels as Type 1 and Type 2, instead of their original connotations like ”poisonous” or ”poor quality”. After consenting to the study, the participant went through the following procedure. For each of two XAI Schemas (randomly ordered, within-subject), (1) Introduction to the Study (see Appendix Fig. A6). (2) Consent to participate. This study was approved by the university institutional review board (IRB). (3) Tutorial on user interface (Fig. A10 for Weights, Fig. A11 for Rules). (4) Forward simulation session (x10 trials) (a) Predict w/o XAI. The UI shows only values and a visualization slider (Fig. 2b). Participants can hover over the slider bars to view attribute ranges in the dataset. They then make a binary choice (Type 1 vs. Type 2) for the AI System’s predicted output (Fig. A14). (b) Predict w/ XAI. Participants saw either factors for the Weights explanation (Fig. A15) or a decision tree for Rules explanation (Fig. A17, without the instance path highlighted). The XAI prediction itself was hidden to encourage reading the explanation. The thresholds and factors were shown to 3 significant figures on hover. (c) Feedback of AI prediction. Participants reviewed their answers against the AI System’s prediction (Fig. A16 and A18). Frequent review supported learning from mistakes and greater focus on the Weights or Rules explanations. (5) Counterfactual simulation session (x10 trials). No AI feedback on the edited instances were shown during this session. (a) Change attribute value w/o XAI. Participants were instructed to ”select one attribute and change it as little as possible” (via sliders; Fig. A20) to alter the AI output. (b) Change attribute value w/ XAI. With either the Weights (Fig. A21) or Rules (Fig. A22) explanation shown, participants reviewed the explanation, then chose and adjusted an attribute using the slider. We asked participants to think aloud and share their reasoning process as they used the XAI to perform forward and counterfactual simulations. This included identifying what parts of the UI they focused on, how they interpreted the explanations, their preferences of the XAI Schemas, and any points of confusion. 4.2. Findings We recruited 24 participants from a local university, with an average age of 22, and 11 of whom were female (13 male). All were undergraduate students from various majors. With participant consent, we recorded both the interactions with the XAI UI and the conversation audio. Each participant was compensated with $7.80 USD in local currency for 30 min of participation time. As participants were partly guided by the experimenters, their performance on the task is not reported. Focusing on our objective of understanding how users interpret rule-based or weight-based XAIs, we center our qualitative analysis on their reasoning strategies across XAI Schemas (Weights, Rules) and under varying test conditions (with or without XAI). We thematically coded the transcripts of the think-aloud study following the principles of grounded theory analysis (Charmaz, 2014; Strauss et al., 1990; Glaser and Strauss, 2017). First, the first and second co-authors conducted open coding using in-vivo codes, with codes generated directly from participant utterances (e.g., “multiplying the value and factor”, “changing the bottom attribute”). Next, these authors performed axial coding to identify relationships among the open codes and organized them into coherent reasoning strategies. These two stages were conducted through multiple rounds of iteration. Then, selective coding was used to merge similar strategies, including those observed across with- and without-XAI trials, but separated by XAI Schema. Finally, the identified reasoning strategies were reviewed by the last senior co-author and refined. Throughout this process, we followed the grounded-theory heuristic of constant comparison to ensure that important distinctions between strategies were not overlooked. 4.2.1. Forward Simulation Task For the forward simulation task, we identified two common strategies for the Weights explanation and one for the Rules explanation. Approximate Calculation (Weights). Some participants attempted to follow the mathematical calculations of the XAI explicitly, by multiplying the factors with the attribute values, and summing up the partial contributions. They applied two mental shortcuts: 1) focusing on attributes that had larger effects on the sum, and 2) using approximations instead of exact calculations when determining the sign of the sum (for a binary decision). Regarding focused attributes, P5 dismissed small partial sums of pH and SO2 as “just fractions of the values”, and P2 “mainly looked at the [mushroom] Height and Width, plus the adjustment [intercept]”. Regarding approximate calculations, P12 “didn’t do calculation to exact value, but I do [it] approximately. Like I just see if the number [partial sum] is big or small compared to each other, and do a very rough estimate”. However, approximations were insufficient for ambiguous cases, e.g., P3 found that “the exact math involved is too complicated for me, so when the partial sum [total sum] is near to the boundary of 0, it is hard to decide”. When the XAI was not shown, participants recalled the factor values and carried out the same process. Confident in his memory, P11 felt that “there is no difference if I see the explanation because I can remember the whole thing [the factors] anyway”. But others were less certain: P3 “forgot the factors, I just remember the first one [for Bruises] is minus 500.” Feature Attribution (Weights). Some participants avoided doing the multiplication math and instead conceived approximations of the partial sums (feature attribution = factor × value) based on the visual representation of the values (Fig. 2). They neglected the factor numbers, except for noting the direction (sign) toward or against the decision. P16 mentioned “the [attribute] width value is very high [based on the visual representation], so maybe it is Type 1. I always remember this [Width factor] value is minus and this one [Height factor] is positive so I mainly do prediction by them” . P2 remarked that “if I can see that the Vinegar Taint and pH on the lower side and Alcohol on the higher side, then the [total sum] number will be positive, then [the AI prediction] will be Type 2 and vice versa”. Participants may calibrate their weight of influence for attributes as they correct their thinking across trials, e.g., P13 realized that “Sulphates is high, maybe, so it contributed to Type 2 more than I thought”. Attribute-Threshold-Attribute Traversal (Rules). When reading the Rules explanation, participants traced the path from the root by navigating based on the thresholds at each node; e.g., P7 “was able to follow the path, where the Alcohol [attribute value] now is 13, more than threshold of 11, and then Vinegar Taint is less than 0.56, it leads me to Sulphates in this case the value now is 0.77, more than 0.5, so it should be Type 2”. However, when deciding without XAI, participants recalled the root attribute and its threshold, applied the threshold, and recalled the subsequent attribute and its threshold, recursively. This allowed them to follow a trace along recalled branches. For example, P7 “remember[ed] the Alcohol cutoff is at 11, so it’s the left branch, then I remember for Vinegar Taint it is less than 0.33, and then Alcohol is tested again and it is more than 9.8, so it is Type 2”. However, participants may recall the wrong attribute or threshold: e.g., P14 said she “tried remembering the model [Decision Tree], and recalling it …, [she] sometimes could not figure out the variables [attribute names]”. P10 “remembered the wrong [threshold] number, I compared it with 0.56, but I should compare with 0.72”; 4.2.2. Counterfactual Simulation Task For the counterfactual task trials, we identified: two strategies for Weights explanation, two for Rules explanation exclusively, and one common strategy for both XAI Schemas. Inverse Calculation (Weights). When XAI was provided, participants would decide how to change one attribute with the following steps: 1) determine how much to change the predicted score to cross the decision threshold (y=0y=0), 2) select the most sensitive attribute (with highest factor magnitude), 3) derive how much to change by dividing the score change needed by the factor number. For example, P11 thought “maybe I reduce Height by 4, because I need to make it [the total sum] a negative number, …, so it should be 290 divided by 90 [Height factor] which is 3 point something”. Some participants kept the new attribute values within realistic bounds, but sometimes found that this was not sufficient to achieve the target prediction, e.g., P3 “was trying to make the partial sum negative to make it to Type 1, and previously I changed Alcohol values to the minimum value, but realized the factor size is too low and it won’t contribute to the negative partial sum”. However, when no XAI was shown, participants were unable to sensibly determine the counterfactual change. P11 recalled “first trying to think what is like the magnitude [of the total sum] like, … and I am not sure, am I supposed to calculate the sum, because that is quite challenging. I will probably not calculate everything”. Inverse Feature Attribution (Weights). Rather than explicitly perform Inverse Calculation, participants approximated partial sums using a Feature Attribution heuristic. They then intuitively estimated how much to adjust an attribute to bring the perceived total contribution in their mental model back to zero. Attribute selection was again biased toward attributes with larger weights in the participant’s internal model. A particularly pronounced case occurred when participants focused solely on the most influential factor where P18 reasoned that the “Width has always been a big positive factor, so I’l just change it until [I think] the prediction flips”. Crucially, this strategy remained effective without XAI, since participants relied on internalized models of Feature Attribution rather than the displayed total sum. Root Threshold Crossing (Rules). Some participants changed the value of the attribute corresponding to the root node, by moving it across the corresponding threshold value. This was a straightforward and cognitively simple strategy with P7 identifying that the same attribute can be changed consistently “rather than changing by following the branches [of the Decision Tree]”. This strategy is robust, even without XAI, where P7 “remembered the threshold value and just hovered around [crossed] it”. Leaf Threshold Crossing (Rules) Participants often modified the values of the attributes at the leaf nodes. This was an intuitive strategy, as in most cases, the neighboring node had the opposite label (though occasionally the neighbor was an internal node; see Fig. 3b, right-most node). For example, P5 “followed the paths, … changed the Sulphate [leaf node] to less than 0.5, it will change to Type 1.” Without XAI, participants attempted to simulate traversal from memory, where P14 “recalled the graph [Decision Tree] mentally and went through it to decide”. However, incomplete recall of deeper nodes sometimes caused participants to stop prematurely, making the strategy less consistent. Availability Heuristic. Participants reused changes they had applied in earlier trials. For example, P5 mentioned they could just use “the previous strategy of increasing the alcohol value” while P11 “did the same change of Sulphates as last time”. Though cognitively cheap, this strategy was flawed because the previous change might not be applicable to the current instance. 5. CoXAM: Cognitive XAI-Adaptive Model We aim to understand, via modeling, how people can learn about, interpret, and decide on various XAI Schemas. We propose CoXAM, a unified cognitive model that can adapt to interpreting different XAI Schemas—Rules and Weights—across decision tasks—forward simulation and counterfactual simulation. Not only does this enable a fair comparison of these different XAI Schemas and decision tasks on the same platform, it facilitates the modeling of a single user learning from and interpreting multiple XAI Schemas together. We combine i) a memory model that determines what can be recalled, i) a drift-diffusion process that converts an evidence number into a choice and decision time, and i) a computational-rational controller that allocates effort and selects strategies to trade accuracy for time. We instantiate three forward simulation strategies and four counterfactual strategies within this framework. 5.1. Memory Model Our memory model follows the account of declarative memory in ACT-R (Anderson et al., 2004) that defines memory chunks and how they are retrieved. Chunks are modeled as structured records of slot-value pairs (e.g., Attribute: Alcohol, Value: 10). Given a retrieval cue q specified as a target slot value (e.g., Attribute: Alcohol), chunks that were used more often, more recently, and that match the cue are more likely to be recalled. Concretely, for a retrieval request at time, t, each chunk, i, is given an activation value, Ai(,t)A_i(q,t), and the chunk with the highest activation is recalled if it exceeds a retrieval threshold, κ; otherwise, retrieval fails. For each chunk, these criteria are computed with an activation at retrieval time t: (6) Ai(,t)=ln⁡(∑j=1ni(Δtij)−0.5)−mi()+ε,A_i(q,t)= \! ( _j=1^n_i( t_ij)^-0.5 )-m_i(q)+ , where nin_i is the number of past uses of the chunk, Δtij t_ij is the time since the j-th use of the chunk, mi()m_i(q) is the count of slot mismatches with the cue, ε∼Logistic(0,ζ) (0,ζ) is an error noise following a Logistic distribution, and ζ>0ζ>0 is the activation noise scale. The chunk will be retrieved if the activation exceeds a retrieval threshold, i.e., Ai(,t)>κA_i(q,t)>κ. 5.1.1. Memory Chunks for Weights explanation Each factor is stored as its own chunk and retrieved using the attribute as the cue. Attribute: name, Factor: value describes a template chunk, which would be recalled with a cue: Attribute: name. Each time the factor is read or recalled, the chunk’s usage count is incremented, which increases its activation and subsequent recall likelihood. Strategy-specific variants may add slots to the chunk and are noted later. 5.1.2. Memory Chunks for Rules explanation We represent the Rules explanation using four chunk types, with each node keyed by a unique Node id. Repeated usage increases activation similar to the Factor chunks. (1) Attribute: stores the node’s attribute, retrieved by the cue of the node id. Example: Node id: i, Attribute: name. (2) Threshold: stores the node’s threshold, with retrieval cue of node id and attribute. Using both as cues can increase mix-ups when multiple nodes share an attribute. Example: Node id: i, Attribute: name, Threshold: τ. (3) Child: stores each outgoing branch, with retrieval cue of node id and branch. Example: Node id: i, Branch: Left/Right, Child node id: j. (4) Leaf Label: stores the class at a leaf (no threshold/children). Example: Node id: i, Label: +1/-1. 5.2. Forward Simulation Strategies We formalize the three reasoning strategies elicited in our formative study: Approximate Calculation and Feature Attribution on Weights explanation, and Attribute-Threshold-Attribute Traversal for Rules explanation. We model how each reasoning strategy supports or refutes a decision. Given a binary decision y∈=−1,+1y =\-1,+1\, each reasoning strategy (described in Sections 5.2 and 5.3) will compute an evidence e for (+1+1) or against (−1-1) the decision. We use a Drift-Diffusion Model (DDM) (Ratcliff, 1978; Ratcliff and Smith, 2004; Bogacz et al., 2006; Wiecki and Frank, 2013) to estimate the probability of deciding y=1y=1 with the reasoning strategy that produces evidence e, and the expected execution time T. Specifically, we use the standard logistic approximation (Bogacz et al., 2006; Navarro and Fuss, 2009): (7) P(y=+1|e) P(y=+1|e) ≈σ(2aeκ2), ≈σ\! ( 2a\,eκ^2 ), (8) [T(e)] [T(e)] ≈a|e|tanh⁡(a|e|κ2), ≈ a|e|\, \! ( a|e|κ^2 ), where σ is the sigmoid function, a>0a>0 is an effort parameter that controls the time spent and consequent probability, and ζ>0ζ>0 is a diffusion noise parameter that slows the speed of reasoning. a is set dynamically using the controller described in section 5.4.3, while ζ is tuned as a cognitive parameter of each participant. 5.2.1. Approximate Calculation (Weights) For each trial, the strategy chooses to look at a subset of the attributes, X, which is again chosen by the controller in section 5.4.3. It multiplies each of the selected attribute values xrx_r with the corresponding factor wrw_r to get a feature attribution ”partial sum” ar=wrxra_r=w_rx_r, sums the total y˘=∑r∈ar y= _r a_r, determines whether the total sum y˘ y is positive or negative, and makes the binary decision sgn(y˘)sgn( y). We define the DDM evidence e of the strategy based on whether the total sum |y˘|| y| is distinct (far from 0) or ambiguous (close to 0), and whether the feature attributions |ar||a_r| are concordant (mostly in the same direction) or ambivalent (many on either direction). Formally, (9) e=|∑r∈wrxr|∑r∈|wrxr|e= | _r w_rx_r| _r |w_rx_r| 5.2.2. Feature Attribution (Weights) This strategy shares some similarities with the Approximate Calculation strategy, but employs mental simplifications for an intuitive approach rather than the rational one of the latter (Kahneman, 2011). For the rrth attribute, instead of reading or recalling the factor wrw_r this strategy uses a nebulous mental representation ωr _r instead. At retrieval, following (Karelaia and Hogarth, 2008), we define this as a random sample from a Normal distribution: (10) ωr∼(μr,σr2) _r ( _r, _r^2) where μr _r and σr _r are the mean and standard deviation parameters, respectively, of the mental model’s guess of the factor. The μr _r and σr _r are stored as individual chunks in memory replacing the explicit factor values for the Approximate Calculation strategy. For the first trial, we initialize each μ to either +1+1 or −1-1 depending on the sign of the provided factor, and each σ to 1. These parameters are updated as more trials are encountered using the provided AI prediction y y as feedback, with a Laplace update for a linear regression model under a gaussian prior (Bishop and Nasrabadi, 2006; Murphy, 2012; MacKay, 1992) (refer to Appendix A.1.1). The DDM evidence is estimated similarly to Eq. 9 but with the latent mental factors: (11) e=|∑r∈ωrxr|∑r∈|ωrxr|e= | _r _rx_r| _r | _rx_r| Note that the calculations are only performed in the cognitive model as an implicit, intuitive mechanism and does not represent an explicit rational calculation by the user. 5.2.3. Attribute-Threshold-Attribute Traversal (Rules) When the Rules explanation is provided, starting from the root node, this reasoning strategy reads the node η, evaluates the attribute value xηx_η against the threshold value τ, and decides whether to follow the left (lesser xη<τηx_η< _η) or right (greater xη≥τηx_η≥ _η) branch. It then recursively applies this comparison at each node until it reaches a leaf. Without XAI, the strategy recalls the Rules explanation by recalling the root node, evaluates its threshold to select the subsequent branch, then recalls the next node and recurses. We define the DDM evidence e of the strategy based on whether each attribute was distinctly clear from the threshold for all traced nodes, i.e., (12) e=∑η|xη−τη|e= _η|x_η- _η| 5.3. Counterfactual Simulation Strategies We formalize the four reasoning strategies elicited in our formative study: Inverse Calculation and Inverse Feature Attribution for Weights explanation, and a merged Node Threshold Crossing and Availability Heuristic for Rules explanation. Unlike the forward simulation task, counterfactual simulation is not a binary decision task, so we cannot use DDM to model probabilities of the choice. We do not model the execution timing of the different strategies, though we include reading and calculation timing measures based on Section 5.4.2. 5.3.1. Inverse Calculation (Weights) With XAI provided, this strategy selects one attribute ρ based on a probability distribution proportional to the factor magnitudes |wr||w_r| across attributes, i.e., attributes with larger factors are more likely to be selected, i.e., (13) P(ρ|y~)=|wρ|∑r|wr|P(ρ| y)= |w_ρ| _r|w_r| With the ρ attribute selected, the strategy adjusts its value in the opposite direction of the current XAI Total Sum y~=∑rwrxr y= _rw_rx_r, just enough to cross 0 to the other prediction label, normalized by the attribute’s factor wρw_ρ i.e., (14) Δxρ=−(∑rwrxrwρ+ε) x_ρ=- ( _rw_rx_rw_ρ+ ) where ε>0 >0 is a margin parameter to cross the y=0y=0 decision boundary. Without XAI, this strategy was not used by participants in the formative study, and would not work. 5.3.2. Inverse Feature Attribution (Weights) This strategy is the same as the Inverse Calculation, but uses internal weights wr(int)w_r^(int) and the calculated internal sum instead. Like the Feature Attribution strategy for forward simulating on Weights explanation, this strategy is the mentally intuitive simplification of Inverse Calculation, where explicit factors wrw_r are replaced by mental representation ωr _r, i.e., (15) Δxρ=−(∑rωrxrωρ+ε) x_ρ=- ( _r _rx_r _ρ+ ) The mental factors are the same as those in Eq. 10, since the same cognitive model can handle both forward and counterfactual simulation tasks. 5.3.3. Node Threshold Crossing (Rules) With XAI, along the trace T for an instance’s attributes, this strategy chooses the tree node at depth d to change the attribute value xdx_d toward the node’s threshold τd _d. In our formative study, we observed that participants chose to change the attribute at the root (d=dmin=0d=d_ =0) or leaf (d=d||d=d_|T|). We generalize this by allowing any other depth d∈[0,d||]d∈[0,d_|T|] to be predetermined as a cognitive parameter. This strategy assumes that the choice of node η is probabilistic following the truncated Normal distribution with normal mean μ=dμ=d and standard deviation σ, i.e., (16) P(η|d)=1σ(ρ−dσ)Φ(d−dσ)−Φ(dmin−dσ)P(η|d)= 1σ N( ρ-dσ) ( d_T-dσ)- ( d_ -dσ) where N is the probability density function of the standard normal distribution, and Φ is its cumulative distribution function. Given a selected node η, this strategy shifts the attribute value xηx_η in the direction toward the threshold τη _η and exceeding by a margin ε , i.e., the counterfactual change in attribute value is (17) Δxη x_η =−(xη−τη) =-(x_η- _η) (18) Δxρ x_ρ =−(Δxη+sgn(Δxη)ε) =-( x_η+sgn( x_η) ) where sgn(∗)=+1sgn(*)=+1 if ∗≥0*≥ 0 and −1-1 otherwise is the sign function, Δxη x_η is the measured difference between attribute value and threshold for the considered node η, Δxρ x_ρ is the amount to change the selected attribute ρ. Without XAI, the strategy forward simulates the corresponding path based on its memory but might be forced to stop an earlier depth if it is unable to recall the full path. 5.3.4. Availability Heuristic Since performing the inverse reasoning is mentally demanding, participants may take a mental shortcut of reusing their recent decisions. This strategy models this by storing each recently used strategy as a memory chunk, recording the change made. It then recalls a chunk probabilistically and repeats it, using the retrieval cue of Target Class: -1, +1. 5.4. Reasoning Strategy Selection We introduce a method to select between the 3 strategies for forward simulation and the 4 strategies for counterfactual simulation, unifying the model into a single combined framework. The strategy selection is based on the concept of computational rationality, where the reasoning time is compromised with the utility of each strategy. We first describe how we modify the probability distribution from the above strategies to account for random errors, then describe how total reasoning time is computed for the strategies and subsequently describe the computational rationality formulation and our specific implementation of it using a Reinforcement Learning (RL) model. 5.4.1. Choice Lapsing To account for occasional random choices, which could occur due to mistakes we do not model such as motor or visual lapses, we adjust the probability P of each reasoning strategy deciding y=+1y=+1 as a combination of the strategy-based probability, and random chance for choosing y=+1y=+1, i.e., (19) P(y~=+1)←(1−λ)P(y=+1)+λ/||P( y=+1)←(1-λ)P(y=+1)+λ/|Y| where 0≤λ<10≤λ<1 is the lapse rate to choose randomly (Wichmann and Hill, 2001; Prins, 2012; Fründ et al., 2011), normalized by number of possible decisions ||=2|Y|=2 for binary decisions. We fix λ to be .05.05, based on prior works (Prins, 2012; Witton et al., 2017; Ratcliff and McKoon, 2008). 5.4.2. Reasoning Time In addition to the execution time T(e)T(e), we add coarse times for reading and explicit mental calculation. We use a simplistic model where each read takes an expected time of 1s, to account for visual search over ∼ 15–20 items (≈ 0.4–0.8s), and ∼ 200ms to internalize the number or text (Wolfe, 2021; Rayner, 1998). Each mental calculation takes an expected time of 2s, consistent with timing for two-significant-figure arithmetic (Ashcraft, 1992). The expected total time for the strategy, is, (20) [T(s)]=[T(e)]+[Tr(s)]+[Tc(s)].E[T(s)]=E[T(e)]+E[T_r(s)]+E[T_c(s)]. where [Tr(s)]E[T_r(s)] and [Tc(s)]E[T_c(s)] are the total times taken for reading and explicit calculations respectively. Explicit calculations are only used in the Approximate Calculation and Inverse Calculation strategies. Our objective is not to match response timings precisely, but rather coarsely account for the variation for time between strategies, which is an important feature of the strategy selection. 5.4.3. Computational Rationality for Strategy Selection Following Lieder and Griffiths (Lieder and Griffiths, 2017), we model the selection of a reasoning strategy s as a rational choice. For each strategy, this determines the Value V of the strategy based on the Utility U of the decision and the Cost of the expected reasoning time T to arrive at the decision. For forward simulation trials, the utility is based on whether the decision x˘ x from the strategy correctly matches the AI label y y, i.e., (21) U(y˘=y^|y^)=P(y˘=+1), if y^=+11−P(y˘=+1), if y^=−1U( y= y| y)= casesP( y=+1)&, if $ y=+1$\\ 1-P( y=+1)&, if $ y=-1$ cases The utility and time varies based on the reasoning strategy s chosen and its cognitive parameters θs _s (e.g., effort a, tree trace depth d). Hence, the Value of the reasoning strategy is: (22) V(s)=U(y˘=y^,s,θs)−γ[T(s,θs)],V(s)=U( y= y,s, _s)- [T(s, _s)], where γ is the opportunity-cost parameter to determine whether the cognitive model prioritizes utility or speed. For counterfactual trials, the utility term is instead instantiated with the probability of changing the AI prediction under the proposed edit distribution. (23) U(y^′≠y^,s,θs)=Prρ∼P⁡[fAI(x+Δxρ)≠y^].U( y ≠ y,s, _s)= _ρ P\! [\,f_AI\! (x+ x_ρ )≠ y\, ]. where P is the strategy output distribution for selecting the attributes and fAI(x+Δxρ)f_AI(x+ x_ρ) is the AI prediction for edited counterfactual of the instance. 5.4.4. Controller and Learning a Strategy Policy We use a multi-layer perceptron (MLP) trained using a reinforcement learning (RL) framework, to model how a participant might adaptively choose reasoning strategies over time. We call this model a controller. In reinforcement learning, an agent learns through trial and error: it observes a state of the environment, takes an action, and receives a numerical reward indicating how good that choice was. At each trial t, the controller observes a continuous state vector which concatenates three components: 1) cognitive parameters of retrieval threshold, α, diffusion noise, ν and opportunity cost, γ; 2) strategy history which includes summary statistics of each strategy’s past use for the current episode (mean time, success rate, and frequency); and 3) mean partial sum values of each feature. This state captures what the controller currently “knows” about their own performance and the task context. The controller then chooses an action at=(s,a,,d)∈,a_t=(s,a,X,d) , where s is a discrete choice of reasoning strategy, a is a the effort level, X is an optional subset of attended attributes for Approximate Calculation and Feature Attribution strategies and d is an optional trace depth for the Node Threshold Crossing strategy. After taking an action, the controller receives a Reward R=V(s)R=V(s) where V(s)V(s) is the Value of the selected strategy. The controller learns a stochastic policy πϕ(at∣st) _φ(a_t s_t) that increases the likelihood of actions leading to higher long-term reward. We optimize this policy using proximal policy optimization (PPO) (Schulman et al., 2017), a standard RL algorithm that balances improvement and stability. Because the controller cannot know in advance how accurate or time-consuming each strategy will be on a new dataset, training it only on the target task would make it overfit and start with poor expectations. We therefore first pre-train the controller on several auxiliary tabular datasets to learn general priors about strategy success rates and timing. During inference on the target domain, these priors allow the controller to make adaptive choices under uncertainty, approximating how a participant might draw on past experience when facing a new task. 5.5. Cognitive Parameters We therefore tune four core cognitive parameters to match each individual participant: (1) Diffusion noise ν>0ν>0 (only for forward simulation): Higher ν shrinks the effective evidence, yielding slower and less accurate choices for the same boundary a. (2) Retrieval threshold κ: Higher κ increases retrieval failures from memory. (3) Opportunity cost γ>0γ>0: Larger γ favors faster strategies (lower a, smaller |||X|) at the expense of accuracy. (4) Margin ε>0 >0 (only for counterfactual simulation): Larger m pushes edits further beyond the decision boundary. 6. Summative Study We conducted a summative user study to investigate the effect of the three different XAI Schemas on user understanding. These collected responses also provide a validation basis for CoXAM. 6.1. Experimental Design We manipulated three independent variables: i) XAI Schemas (Rules, Weights, and Hybrid; between subjects), i) XAI complexity (Low and High; between subjects), and i) Tested with XAI (with or without XAI; within subjects). Each participant was assigned to one of three XAI Schemas: Rules, Weighs or Hybrid. In the Hybrid condition, the XAI Schema was randomly chosen on each trial between Rules and Weights, thereby exposing participants to both forms of XAI. This condition provides a stronger test of our model’s XAI adaptability, as participants could flexibly employ strategies suited to either explanation type. We also varied XAI complexity at two levels (between subjects): Low, featuring a Decision Tree of depth 2 or a sparsified Linear Regression with 3 non-zero factors, and High, featuring a Decision Tree of depth 3 or a full Linear Regression with 6 non-zero factors. These manipulations allowed us to better evaluate the model’s generalizability. For both the forward simulation and counterfactual phases, we controlled the instance selection to satisfy (i) prediction labels of the AI are balanced to be 50% for each label type and (i) each XAI prediction matched the AI prediction for 90% of the instances. For each trial with XAI shown, participants using the Hybrid XAI are randomly shown one of Weights or Rules explanations. We measured dependent variables: (1) Forward Simulation Accuracy: proportion of matches between user labels and AI predictions. (24) Forward Accuracy =1N∑k=1N(y˘k=y^k)Forward Accuracy = 1N _k=1^N 1( y_k= y_k) where () 1() is the indicator function, y˘k y_k is the user response and y^k y_k is the indicator function. Since the fidelity of each XAI was controlled to be 90%, we do not report the redundant comparison between user labels and XAI predictions. (2) Counterfactual Simulation Accuracy: proportion of trials in which participants successfully changed the AI’s prediction by modifying an attribute, creating new instance ˘′ x . (25) Counterfactual Accuracy=1N∑k=1N(fAI(x˘′)≠y^k)Counterfactual Accuracy= 1N _k=1^N 1(f_AI( x )≠ y_k) where fAI(x˘′)f_AI( x ) is the AI prediction for the new instance. 6.1.1. Experiment Procedure Each participant performed the following procedure: Trials in the forward and counterfactual sessions are similar to the formative user study (refer to Section 4), (1) Introduction to the study. (2) Consent to participate. (3) Tutorial on the user interface. (4) Tutorial on the XAI UI. Following which, we asked three screening questions about the UI, and the two XAI explanations (see in Appendix, Figs. A9, A10, A11). (5) Session 1: Forward Simulation (×40× 40 trials) (a) View the interface with attribute values (with or without XAI) and estimate the AI prediction (see Fig. A12 for without XAI; Fig. A15–A17 for with XAI). (b) Receive feedback on the AI prediction [with XAI (Figs. A16–A18) or without XAI (Fig. A14)]. (6) Session 2: Counterfactual Simulation (×40× 40 trials): Select an attribute and change its value [with or without XAI] (Fig. A20-A21) (7) Answer demographic questions. We recruited 340 participants (202 female, median age 36) via Prolific with the minimum requirement of 20 participants per condition. An additional 300 participants were excluded based on a straightforward comprehension screening (Figs. A10, A11). Each was compensated £7.00 for an average completion time of 51 mins. The variation in complexity and datasets, is intended to test the generalization of CoXAM and whether the previous elicited strategies persist in a broader online population. 6.2. Results Figure 4. User performance of forward simulation task in a), and b) and counterfactual simulation task in c), and d), across datasets: Wine Quality (left), and Mushrooms (right) across XAI Schemas, and testing conditions (w/o XAI, w/ XAI). Error bars are 95% CIs. Random performance for the forward simulation task would be 50%. Panels (a) and (b) show forward simulation accuracy for the Wine Quality and Mushrooms datasets, respectively, while panels (c) and (d) present counterfactual simulation accuracy for the same datasets. Results are reported for three XAI Schemas (Rules, Weights, and Hybrid) under two testing conditions: without explanations (red, w/o XAI) and with explanations (blue, w/ XAI). Error bars represent 95% confidence intervals. Forward simulation accuracy is consistently above the random baseline of 50%, though the relative benefits of explanations differ across datasets and model types. Counterfactual simulation accuracy is lower overall, with modest improvements depending on model type and explanation condition. We performed a linear mixed effects model fit on each dependent variable as the response, Complexity nested within XAI Schemas, Trial Index and Tested with XAI (with or without XAI trial) as fixed effects, an interaction effect between the XAI Schemas and Tested with XAI, and Participant as a random effect. Complexity was found to have insignificant fixed and interaction effects, and hence we omitted it from our analysis for clarity. We briefly discuss significant results as these findings are not the main contribution of our work. We only report pairwise comparisons using contrast t-tests that are highly significant, p << .005. 6.3. Forward Simulation Participants in the Rules condition performed significantly better in With-XAI trials compared to Without-XAI trials across both datasets (Fig. 4a–b), Mushrooms (78.7% vs. 63.9%, contrast test, p<.001p<.001) and Wine Quality (74.8% vs. 63.9%, p<.001p<.001). In contrast, participants in the Weights and Hybrid conditions showed no significant difference between with- and without-XAI trials. This suggests that Rules explanations are harder to recall than to read, whereas Weights explanations can be more easily internalized. Across XAI Schemas, the Wine Quality dataset appeared more amenable to Weights explanation than to Rules explanation with a significant improvement in accuracy in Without-XAI trials (72.3% vs. 63.9%, contrast test: p<.005p<.005). Conversely, for the Mushrooms dataset, Weights had lower accuracy in with-XAI trials relative to Rules (65.6% vs. 78.7%, contrast test: p<.0001p<.0001). Two factors likely contribute to this increased performance of Rules: (1) the Decision Tree for the Mushrooms data is slightly smaller, containing two fewer leaf nodes (see Fig. 3b), and (2) the Mushrooms dataset exhibits more nonlinear relationships than Wine Quality. 6.4. Counterfactual Simulation Counterfactual performance showed a similar pattern as shown in Fig. 4c–d, with Weights outperforming Rules (43.6% vs. 26.8%, contrast test: p<.001p<.001) on the Wine Quality dataset. For Mushrooms, no significant differences were found across XAI Schemas, possibly because the dataset’s structure aligns well with the Decision Tree representation, mitigating its typical disadvantages for counterfactual reasoning. Participants using the Rules explanation changed values more conservatively (in Fig. A2), likely because they relied on recalling explicit threshold values when identifying attribute changes. 7. Modeling Analysis To understand why certain XAI and reasoning strategies are more effective compared to others, we use CoXAM to simulate participant behavior and explain performance and trends using underlying cognitive causes. 7.1. Fitting CoXAM for Forward Simulation Trials Table 1. Comparison of Negative Log-Likelihood (NLL) and Bayesian Information Criterion (BIC) for CoXAM against baseline proxies for forward simulation task for the Wine Quality and Mushrooms datasets. Lower values are better. Bold is best fit. This table compares model fits for the forward simulation task using Negative Log-Likelihood (NLL) and Bayesian Information Criterion (BIC) across two datasets (Wine Quality and Mushrooms) and three explanation conditions (Rules, Weights, and Hybrid). Lower values indicate better model fit. Overall, CoXAM achieves the best fit across most conditions, showing lower NLL and BIC values than baseline models (Decision Tree, Linear Regression, and KNN without XAI). An exception occurs in the Mushrooms dataset under the Rules condition, where the Decision Tree baseline attains a slightly lower BIC, suggesting close performance for that specific case. Wine Quality Mushrooms Rules Weights Hybrid Rules Weights Hybrid Models NLL BIC NLL BIC NLL BIC NLL BIC NLL BIC NLL BIC Decision Tree 26.8 57.3 — — 26.2 56.1 21.7 47.1 — — 29.7 63.1 Linear Regression — — 28.0 59.7 26.8 57.3 — — 37.0 77.7 35.5 74.7 KNN w/o XAI 29.7 66.8 30.1 67.6 29.9 67.2 24.0 55.4 28.1 63.6 27.2 61.8 CoXAM 18.9 48.8 19.9 50.9 20.2 51.5 20.7 52.5 21.5 54.1 20.8 52.7 To validate our modeling approach, we fit CoXAM to the 40 forward simulation trials for each participant individually with three tunable parameters: retrieval threshold, κ, opportunity cost, γ, and diffusion noise, ν. The fitting was done using a Gaussian process-based optimization (Snoek et al., 2012). Due to the difficulty of integrating a static, global XAI explanation directly into a naïve machine-learning model as suggested by Chen et al. (Chen et al., 2022), we implemented three alternative baseline proxies to approximate human predictions: (1) K-Nearest Neighbors (no XAI): Each participant is represented by a KNN model that incrementally updates its training set with the instances encountered so far, without incorporating any XAI information. (2) Decision Tree Proxy: For participants in the Rules and Hybrid conditions, we used the corresponding Decision Tree XAI model to generate predictions. To better match participants’ stochastic errors, we applied a single smoothing parameter that blends the model’s predicted probability with a uniform distribution over possible outcomes. This accounted for occasional lapses or uncertainty in human responses. (3) Linear Regression Proxy: Analogous to the Decision Tree proxy, we used the Linear-Regression XAI to approximate participant behavior in the Weights and Hybrid conditions. Table 1 shows that CoXAM provides a more faithful account of participant responses than all three proxy models. The only exception occurs for the Decision Tree proxy on the Mushrooms dataset, which attains a slightly lower BIC despite a higher NLL, due to having less free parameters. Figure 5. Results of summative user study (orange) compared to CoXAM (blue) of accuracy in the forward simulation task. Error bars are 95% CI. Forward simulation accuracy for human participants (orange) and the CoXAM model (blue) across explanation types and datasets, evaluated with and without XAI support. Results are reported for the Wine Quality dataset (a–b) and the Mushrooms dataset (c–d), under testing conditions without XAI (a, c) and with XAI (b, d). Explanation types include Rules, Weights, and Hybrid representations. Accuracy is computed with respect to the AI system’s ground-truth predictions, and error bars denote 95% confidence intervals. Overall, CoXAM closely mirrors human performance patterns across explanation types and conditions, capturing both relative performance differences among explanation formats and the impact of XAI availability. Figure 6. Forward simulation accuracy of two Weights and one Rules strategies for with and without XAI trials across datasets. Forward simulation accuracy across explanation strategies, datasets, and XAI conditions. The figure reports mean forward simulation accuracy (y-axis, %) with error bars indicating variability across participants. Results are shown for two datasets—Wine Quality (a–b) and Mushrooms (c–d)—under conditions without XAI (a, c) and with XAI (b, d). Three explanation strategies are compared: Approximate Calculation, Feature Attribution, and Attribute–Threshold–Attribute (ATA) Traversal. For the Wine Quality dataset, Feature Attribution and ATA Traversal outperform Approximate Calculation in both conditions, with a consistent improvement when XAI is available. For the Mushrooms dataset, ATA Traversal achieves the highest accuracy with XAI, while Feature Attribution shows reduced performance relative to the other strategies. Overall, the results indicate that explanation strategy and the presence of XAI substantially influence users’ ability to accurately simulate model behavior, with rule-based ATA Traversal benefiting most from XAI support. Figure 7. Strategy prevalence by XAI Schema for forward simulation. This figure shows the prevalence of different reasoning strategies used by participants in the forward simulation task, grouped by XAI Schema and dataset. Each cell represents the proportion of trials where a given strategy was applied. For the Wine Quality dataset, participants predominantly used the Attribute Threshold/Traversal strategy with Rules and the Feature Attribution strategy with Weights XAI. In contrast, for the Mushrooms dataset, the Approximate Calculation strategy was dominant for Weights XAI, while Attribute Threshold/Traversal remained consistent for Rules. Hybrid XAI conditions show mixed usage patterns, indicating flexible strategy selection depending on task complexity. Table 2. Comparison of Negative Log-Likelihood (NLL) and Bayesian Information Criterion (BIC) for CoXAM against baseline proxies for participant attribute-selection in the counterfactual task for the Wine Quality and Mushrooms datasets.Lower is better. Bold is best fit. This table compares the Negative Log-Likelihood (NLL) and Bayesian Information Criterion (BIC) of CoXAM against two baseline models—Random and Global SHAP—across the Wine Quality and Mushrooms datasets, for three explanation types (Rules, Weights, and Hybrid). Lower values indicate better model fit. CoXAM consistently achieves the lowest NLL and BIC across all conditions, showing strong agreement with user behavior. While Global SHAP improves upon the Random baseline, it remains less predictive of participant decisions compared to CoXAM. Wine Quality Mushrooms Rules Weights Hybrid Rules Weights Hybrid Model NLL BIC NLL BIC NLL BIC NLL BIC NLL BIC NLL BIC Random 71.67 143.34 71.67 143.34 71.67 143.34 71.67 143.34 71.67 143.34 71.67 143.34 Global SHAP 42.1 84.2 56.3 112.6 54.7 106.4 50.2 100.4 55.1 110.2 51.0 102.0 CoXAM 35.6 71.2 51.8 103.6 45.0 90.0 45.8 91.6 50.6 101.2 46.6 93.2 Figure 8. Results of summative user study (orange) compared to CoXAM (blue) of accuracy in counterfactual simulation task. Error bars are 95% CI. Counterfactual simulation accuracy for human participants (orange) and the CoXAM model (blue) across explanation types and datasets, evaluated with and without XAI support. Results are shown for the Wine Quality dataset (a–b) and the Mushrooms dataset (c–d), under testing conditions without XAI (a, c) and with XAI (b, d). Explanation types include Rules, Weights, and Hybrid representations. Accuracy is computed with respect to the AI system’s prediction change, and error bars denote 95% confidence intervals. Overall, CoXAM closely tracks human performance across datasets, explanation types, and XAI conditions, reproducing relative performance differences and condition-dependent trends observed in human counterfactual reasoning. Figure 9. Results of CoXAM simulated user accuracy in the counterfactual simulation task, for different strategies. Error bars are 95% CI. Figure 9. Counterfactual simulation accuracy of CoXAM-simulated users across two datasets (Wine Quality and Mushrooms), with and without XAI support. Results are shown for four reasoning strategies: Inverse Calculation, Node Threshold Crossing, Inverse Feature Attribution, and Availability Heuristic. Error bars represent 95% confidence intervals. Accuracy is highest for Inverse Calculation and Node Threshold Crossing, while Availability Heuristic consistently yields lower performance. Figure 10. Strategy prevalence by XAI Schema for counterfactual simulation. This figure shows how often different reasoning strategies were used in the counterfactual simulation task, grouped by XAI Schema and dataset. The rows list the four strategies: Inverse Calculation, Inverse Feature Attribution, Node Threshold Crossing, and Availability Heuristic. The columns represent the Wine Quality and Mushrooms datasets under Rules, Weights, and Hybrid explanation conditions. Each cell shows the percentage of trials in which that strategy was used. For Wine Quality, participants most often used Node Threshold Crossing with Rules explanations and Inverse Feature Attribution with Weights explanation. For Mushrooms, Availability Heuristic and Inverse Calculation were most common under Weights explanations, while Node Threshold Crossing remained frequent under Rules and Hybrid conditions. We sampled responses from CoXAM using parameter distributions fitted per-participant to reproduce human forward-simulation performance across 12 experimental conditions (2 datasets × 2 conditions [with vs.without XAI] × 3 XAI Schemas). The resulting model accuracies (Fig. 5) were highly correlated with human performance (Pearson r=.91r=.91, p<.0001p<.0001) and closely matched in magnitude (RMSE = 1.8%), indicating that CoXAM effectively captured both the relative trends and absolute levels of participant accuracy. 7.2. Analysis of Reasoning Strategies for Forward Simulation We analyzed the strategy selection distribution of CoXAM in the forward simulation task (Fig. 7). Under the Weights explanation condition, CoXAM primarily used the Feature Attribution strategy for the Wine Quality dataset, while favoring the Approximate Calculation strategy for the Mushrooms dataset. When analyzing strategy performance separately (Fig. 6), the model achieved higher accuracy with Feature Attribution (73.3%) than with Approximate Calculation (48.1%) on Wine Quality (contrast t-test, p<.001p<.001), prompting the controller to prefer Feature Attribution. In contrast, for Mushrooms, Approximate Calculation outperformed Feature Attribution (62.7% vs. 52.8%; p<.001p<.001), reflecting adaptation to dataset properties. These differences align with task structure. Factor magnitudes in Mushrooms are smaller, so approximate mental arithmetic is easier, which benefits Approximate Calculation and supports the observed preference. By comparison, Wine Quality involves larger-valued attributes that make mental sums and differences harder, which reduces the effectiveness of approximate arithmetic and favors Feature Attribution under Weights explanations. For the Hybrid XAI Schema, CoXAM favored the Attribute-Threshold-Attribute Traversal method across datasets. Although Feature Attribution slightly outperformed Attribute–Threshold –Attribute Traversal strategy for the Wine Quality dataset, CoXAM preferred the latter due to its lower predicted execution time (9.1s versus 11.1s for Feature Attribution), indicating that the controller optimizes accuracy and reasoning time rather than accuracy alone. Detailed accuracy and timing per strategy are reported in Table 5 in the Appendix. We then examined sensitivity to cognitive parameters. Increasing the opportunity cost led the controller to select a smaller subset of attributes under Approximate Calculation (Fig. A3b), with a similar trend for Feature Attribution, indicating that higher perceived effort shifts behavior toward fewer, more diagnostic attributes. Accuracy declined with higher opportunity cost, diffusion noise, and retrieval threshold (Fig. A3a). Notably, retrieval threshold had the weakest effect on Feature Attribution, suggesting that this strategy is comparatively robust to memory limitations, while traversal and approximate arithmetic depend more on successful recall of intermediate quantities and path information. 7.3. Fitting CoXAM for Counterfactual Simulation Trials We also fit CoXAM to the counterfactual simulation trials of participants with three parameters, analogous to the forward-simulation fitting, but replacing the diffusion noise parameter ν with the margin parameter ε . CoXAM was optimized to minimize the sum of (i) the negative log-likelihood of the user’s attribute-selection choices and (i) the mean absolute error (MAE) between the participant’s change magnitude and CoXAM’s predicted change. To evaluate how well CoXAM reproduces participants’ attribute-selection behavior, we compared it against two baseline models: (1) Random: A simple baseline assuming that each of the six attributes has an equal probability of being selected. (2) Global SHAP: A knowledge-based baseline assuming participants can identify and select attributes important to the underlying AI model. Each attribute’s selection probability is set proportional to its global importance, computed as the mean absolute SHAP value of the attribute across all instances in the training dataset. The comparison results demonstrate that CoXAM provides a closer approximation to human behavior than either baseline. The overall distribution of attribute selections is reported in the Appendix A5. Sampling from the attribute distribution output by CoXAM with its parameters fit on the users, we compare CoXAM’s overall accuracies against human results (Fig. 8). Again, CoXAM aligns well with user performance across the 12 conditions (Pearson correlation r=72.1%, RMSE=5.49%). 7.4. Analysis of Reasoning Strategies for Counterfactual Simulation Analyzing strategy selection revealed patterns consistent with the forward-simulation modeling while also highlighting counterfactual-specific effects. Under the Weights explanation, CoXAM favored Inverse Feature Attribution more strongly for the Wine Quality dataset, whereas strategy use was more balanced between Inverse Feature Attribution and Inverse Calculation for Mushrooms. Despite this difference in selection, both strategies achieved similar accuracies (Wine Quality: 50.3% vs. 47.3%; Mushrooms: 44.9% vs. 48.9%), indicating no clear performance advantage. A substantial proportion of trials relied on the Availability Heuristic, which consistently yielded the lowest accuracy, but was selected due to its speed of use. In the Hybrid XAI condition, CoXAM strongly preferred the Node Threshold Crossing strategy, although the Availability Heuristic still appeared in a notable fraction of trials. For the Rules exlanation, CoXAM predicts that about 43% of trials relied on the Availability Heuristic across both datasets, underscoring its frequent use despite poor accuracy. Additionally, the Node Threshold Crossing method was very ineffective for the Wine Quality dataset, signalling that the decision threshold was not properly represented by the Rules explanation. Sensitivity analysis of the cognitive parameters (Fig. A4) showed weaker and less consistent effects than in the forward-simulation task. Even for Node Threshold Crossing, which requires a forward traversal of the decision tree before counterfactual editing, accuracy was only modestly influenced by the retrieval threshold, suggesting that participants had largely internalized the XAI through prior forward-simulation exposure. A clearer effect emerged for the margin parameter ε , where counterfactual accuracy increased with larger margins across strategies. This trend indicates that the explainer boundaries differ substantially from those of the underlying AI, such that optimal counterfactual edits often require discounting the explainer’s boundary. 8. Discussion We discuss our key findings about Rules and Weights explanation understanding, and the generalization and limitations of CoXAM. 8.1. Rules vs. Weights Our experimental findings show that trends are dataset and strategy dependent even if the fidelity and cognitive load of the XAI is kept consistent (i.e., similar depth trees and values rounded to similar significant numbers in the interface). We saw that participants are able to flexibly identify the more successful strategy and use it, with the Approximate Calculation and Feature Attribution strategies being preferred differently across the two datasets. On the other hand, we found a consistent choice of favoring Rule-based strategies (Attribute-Threshold-Attribute Traversal and Node Threshold Crossing) in the Hybrid XAI condition for both forward and counterfactual tasks. This was despite the fact that the Feature Attribution method slightly outperformed the Attribute-Threshold-Attribute strategy for the Wine Quality dataset. The reason for this choice was due to a shorter reasoning time with the latter method, emphasizing the importance of considering cognitive load while improving performance of metrics. We also discovered that different strategies have different advantages. While Attribute-Threshold-Attribute is fast and effective when the XAI is shown, it is susceptible to poor memory recall. On the other hand, Feature Attribution is robust to memory recall, though by its incremental learning nature, has to take a few instances to be effective. Comparing the strategies and XAI Schemas without knowledge of the task requirements is not fruitful, and CoXAM provides an evaluation basis for future developers to identify the best reasoning strategies based on their requirements. 8.2. Implications of Explainable User Interpretation with CoXAM Using the CoXAM cognitive model to simulate user interpretation of XAI, AI and HCI researchers, designers and developers can investigate the effectiveness of user reasoning or debug the causes of user misinterpretations. Rather than focusing solely on ”explainable AI,” CoXAM prioritizes the ”explainable human” within the XAI interaction loop. This approach is not intended to replace human-subjects in a validation or evaluation study, but to facilitate the XAI method iteration. Specifically, CoXAM assists in hypothesis formation to anticipate experimental results, allows designers to pre-empt interpretation failures to avoid wasteful user studies, and provides a framework to analyze why an XAI method might fail. Ultimately, subsequent user studies remain essential to validate the insights generated by the CoXAM model. Suppose that CoXAM modeling finds that an XAI method would lead to poor understanding. There are three approaches that could be used by various stakeholders to improve the XAI: tutorial, UI design, optimization. The HCI researcher can prepare a tutorial for the user study to prime the participants by training them employ the identified effective reasoning strategies instead of fallacious ones; e.g., using the Feature Attribution reasoning strategy rather than Approximate Calculation as we identified for the Wine Quality task. The XAI UI designer can use CoXAM to systematically filter the attributes included in an explanation by analyzing the trade-off between information gain and cognitive load; by simulating how user performance marginally improves or diminishes as the number of attributes increases, designers can identify the optimal attribute count and calibrate specific cognitive parameters to maximize interpretability without exceeding the target user’s capacity. The XAI developer can optimize the XAI method to maximize user decision performance by formally regularizing for desiderata (e.g., cognitive load (Abdul et al., 2020)), or training an XAI model via backpropagation from a pre-trained neural network proxy based on CoXAM-labeled decisions like in (Li and Yin, 2024). 8.3. Scope and Generalization of CoXAM In this work, we had focused on binary classification over tabular datasets with six attributes, using forward and counterfactual simulation tasks to examine the effective use of XAI. While this design provided controlled insights, future work can examine implementing other types of explanation to scale our approach for XAI cognitive modeling. We examine generalizing CoXAM to different global XAI Schemas, explanation modalities, and decision tasks. Table 3. Scope of CoXAM for user interpretation of global XAI Schemas on interpretable features. XAI Schema Representative XAI Methods CoXAM Coverage Rule-based Decision Tree (Breiman et al., 2017), Rule List (Angelino et al., 2018b), Rule Set (Lakkaraju et al., 2016a) Covered: Topology-agnostic by converting to equivalent DNF. Future work: Account for differences in parsing topology. Weight-based (Linear Additive) Linear Regression (Bo et al., 2024), Logistic Regression (Poursabzi-Sangdeh et al., 2021) Covered. Additive (Non-Linear) Partial Dependence Plot (Krause et al., 2016), Generalized Additive Model (GAM) (Hastie and Tibshirani, 1990), Piecewise Linear (Friedman, 1991; Bo et al., 2024) Future work: encode interpretation of nonlinear trends as local linear based on prototype examples (Kalish et al., 2004). Symbolic Regression (Non-Linear) Domain-specific equations (Alaa and Van der Schaar, 2019) Future work: encode interpretation with rational model of function learning (Lucas et al., 2015). Example-based Prototypes (Narayanan and Bergen, 2024), Criticisms (Kim et al., 2016) Future work: encode examples in memory, retrievable based on computational rationality, to infer labels based on similarity (Nosofsky, 1986). 8.3.1. Global XAI Schemas on Interpretable Features Focusing on global explanations with interpretable features, Table 3 shows which XAI Schemas are currently supported in CoXAM and how it could be extended to other schemas. Although we had evaluated on decision trees for rule-based XAI, other methods like Rule Lists (Angelino et al., 2018b) or Decision Sets (Lakkaraju et al., 2016b) with different topologies can also be converted to a unified representation in disjunctive normal form (DNF) to be ingested into CoXAM. However, this neglects the relative difficulties or cognitive short-cuts in interpreting these structures, which future work needs to elicit and model. While CoXAM implements linear weight-based explanations, this is a simplification of explainable additive models, where each attribute contributes independently but nonlinearly to the AI decision. Such models include explanations with partial dependence plots (Hastie and Tibshirani, 1990), generalized additive models (GAM) (Hastie and Tibshirani, 1990), and piecewise linear models (Friedman, 1991; Bo et al., 2024). Future work should explore how users encode nonlinear information, perhaps, piecewise with local linear segments around prototype examples (Kalish et al., 2004). For domains with specific symbolic formulas or equations, or users with knowledge of basis functions (e.g., exponential, logarithm, sinusoidal), future work can model how users learn and interpret explicit functions (Lucas et al., 2015). Finally, using analogous reasoning, people often learn key examples and make inferences based on them. To model user interpretation of example-based explanations, such as prototypes (Narayanan and Bergen, 2024) and criticisms (Kim et al., 2016), future work can model example retrieval based on computational rationality and infer labels based on similarity with the Generalized Context Model (Nosofsky, 1986). 8.3.2. Multimodal Explanations Although predicting and explaining on structured data with symbolic reasoning is still paramount (Rudin, 2019; Shwartz-Ziv and Armon, 2022; Fang et al., 2024), many explanations are presented as visualizations, and recent machine learning models, such as deep learning and large language models (LLM), are facilitating new modalities in AI and multimodal explanations. We discuss how CoXAM and cognitive modeling, can account for visual explanations, explanations of unstructured image data and LLM-generated text explanations. Many XAI methods present their explanations as visualizations. For example, linear LIME as tornado plot (Ribeiro et al., 2016), SHAP attributions as scatter plot (Lundberg and Lee, 2017), linear weights as lollipop charts (Abdul et al., 2020), or in tables (Bo et al., 2024; Poursabzi-Sangdeh et al., 2021). In this work, we had focused on user reasoning of the semantic information in the explanations, rules and weights, which we had parsed before ingesting into CoXAM as memory chunks. However, the visual representations can affect the intuitiveness and readability of the explanations. Following past works to model human visual attention and encoding (Anderson et al., 1997), and computational rationality for chart reading (Shi et al., 2025), future work can model how users read popular XAI visualizations to more precisely account for varying visual representations. Other than tabular data, deep learning has been successful in reasoning on image data, and this necessitates image-based XAI. Popular explanation techniques include pixel-based saliency maps (Selvaraju, 2017) and image segmentation (Kirillov et al., 2023). These explanations require users to take additional steps to interpret them (Lim et al., 2025), such as identifying relevant objects or concepts from the image and matching them to salient regions. Indeed, concept-based explanations (Koh et al., 2020; Kim et al., 2018) support these steps by extracting concepts and communicating their prevalence or importance to the user. Nonetheless, this clarifies that image-based explanations are two-staged: i) inferring concepts from perception, i) symbolic reasoning on the interpretable concepts. CoXAM already handles the latter, albeit without accounting for uncertain concepts. Future work could be extended with a concept inference module, though this also requires encoding or making assumptions about the user’s prior visual knowledge. Recent developments in large vision-language models (VLMs) have led to highly usable, chat-based interactions for decision systems. For example, instead of inputting a feature vector to a model, a user could supply an image and prompt “Here’s a photo of a mushroom, GPT, tell me whether it is edible.”, and receive a natural language explanation. To interpret this, users construct mental models by extracting symbolic relations from the language (Johnson-Laird, 1983). For future work, CoXAM could simulate this by integrating ACT-R-style production parsing (Lewis and Vasishth, 2005), or leveraging LLMs as symbolic parsers to map unstructured text onto structured attributes and rules (Tian et al., 2024). However, because LLM-based decision systems still have critical limitations of hallucination (Ji et al., 2023), sycophancy (Fanous et al., 2025), and human-like biases (Deng et al., 2025), they are not yet reliable for high-stakes settings. Nevertheless, given their rapid development and increasing adoption, this is a crucial extension for CoXAM. 8.3.3. Decision Tasks We evaluated CoXAM user understanding based on forward and counterfactual simulation tasks, where the user estimates what the AI would predict, and how to change its prediction, respectively. Another way to demonstrate understanding is to correctly attribute the important factors for an AI decision. This is attribution simulation (Waa et al., 2021). Since the schema of XAI can take various forms, users can further demonstrate their understanding by reconstructing or recalling the explanation for decisions, such as in the form of decision trees (Lim et al., 2009), non-linear line charts (Abdul et al., 2020), saliency maps (Boggust et al., 2022), etc. Future work could also examine human-AI collaboration tasks, where performance is defined by joint decision accuracy (Doshi-Velez and Kim, 2017). This would require modeling users with two mental models, one of the AI’s decision process and one of the ground-truth labels, as well as incorporating additional cognitive variables such as trust in the AI (Wang et al., 2022). Finally, self-reported measures such as perceived usefulness, cognitive load, and ease-of-use are common in XAI user studies (Rong et al., 2023). Prior work has shown that such measures are often uncorrelated with objective assessments of understanding, and we therefore leave their modeling to future work. 8.4. LLMs as Cognitive Models With the advent of Large Language Models (LLMs) and other foundation models, there has been a line of work attempting to use them to model human behavior (Park et al., 2023; Hwang et al., 2025; Park et al., 2024). While this field looks extremely promising, we argue that on its own, an LLM is still not at a phase to replace more traditional modeling approaches. LLMs can replicate external human behavior, but they do not explicitly model components of the human brain and are not explicitly controllable, in contrast to cognitive models where behavior is defined through explicit functions. Fitting an LLM to a human through persona-based prompting (Sun et al., 2025; Yeo et al., 2025a) with chain-of-thought reasoning (Wei et al., 2022) may appear to rationalize (Ehsan et al., 2019) like humans, but these are proxy explanations and not mechanistically traceable (Rudin, 2019). We instead suggest, as in Chartist (Shi et al., 2025), that LLMs could be used in conjunction with cognitive modeling to supplement the controller module in our framework with prior knowledge and flexible reasoning strategies while not operating independently as a black box model. 8.5. Risk of Overreliance on XAI Cognitive Modeling While CoXAM facilitates systematic XAI evaluation, over-relying on it without considering its underlying assumptions poses several risks. We discuss implications due to neglecting: real user validation, target user representation, other explanation goals, and human-AI collaborative decisions. First, using CoXAM is ultimately for hypothesis formation of how users may (mis)interpret XAI methods; to verify these results, evaluations with real users must be conducted. Next, the reasoning strategies implemented in the cognitive models should be informed from formative user studies of representative target users, otherwise, the decision outcomes may be spurious and deviate from how the relevant users would perform. Furthermore, evaluating XAI benefits solely with the two decision tasks we have implemented (forward and counterfactual simulation) omits other goals and desiderata for explanation, such as trust, satisfaction and robustness (Wang et al., 2019; Langer et al., 2021; Thomas and Uminsky, 2022); neglecting these risk unintended consequences, i.e., Goodhart’s law (Strathern, 1997). Finally, users can also use AI collaboratively rather than only for automated decisions. Although human-aligned XAI could lead to convergent thinking where both AI and human share the same decision flaws, our work on CoXAM works toward human-aligned cognitive modeling for human-augmenting XAI. Instead of increasing the illusion of explanation for increased subjective satisfaction or perceived understanding (Skirzynski et al., ; Vasconcelos et al., 2023), iterating XAI methods with CoXAM aims to improve the objective performance of understanding correctness. The goal is to improve actual understanding rather than make understanding easier. Rather than aligning the XAI which could be induced with human biases (Li and Yin, 2024), CoXAM helps to account for fallacious reasoning such that XAI can be designed to mitigate them. Thus, CoXAM could improve human-AI collaborative decisions as long as users verify XAI (Buçinca et al., 2021), though this may be cognitively demanding since they need to maintain contrasting mental models of themselves and the AI. 8.6. Limitations of Ecological Validity The CoXAM controller is pre-trained on four datasets, which we assume provide inductive biases about strategy selection that participants may have acquired from real-world experience. However, this constitutes a limited corpus and may embed priors specific to those datasets. A more comprehensive training regime using a broader and more diverse dataset collection would yield a more faithful representation of human experience. Furthermore, our experimental design fixed the counterfactual simulation session after the forward simulation session. This ordering ensured that participants had an adequate forward-model understanding before attempting counterfactual reasoning, but it may have reduced ecological validity by introducing potential order effects. Notably, we observed that memory limitations did not influence counterfactual performance under this setup. Finally, since we focused on semantic understanding, we chose assessable explanation schemas (decision trees and linear regression), and low-dimensional domains with understandable contexts (e.g, mushrooms). This facilitated evaluations with lay participants, and is in line with user studies of XAI (e.g., (Poursabzi-Sangdeh et al., 2021; Abdul et al., 2020; Lai et al., 2022; Lim et al., 2009; Buçinca et al., 2021; Bo et al., 2024; Ma et al., 2023; Lakkaraju et al., 2016a; Rong et al., 2023)). Future work can investigate modeling user understanding for more specialized, complex, high-dimensional domains (e.g., healthcare (Lim et al., 2025; Panigutti et al., 2023; van Leersum and Maathuis, 2025), and high-dimensional visual-analytic explanations (Lyu et al., 2024; Kahng et al., 2017, 2017; Boggust et al., 2022)). 9. Conclusion We have presented CoXAM, a cognitive XAI-Adaptive Model, as a framework for debugging XAI interpretability. We conducted a formative elicitation user study which revealed insights into 7 user reasoning strategies for interpreting Rules, Weights and Hybrid explanations, which we then evaluated in a summative validation study. We implemented the XAI-adaptive CoXAM cognitive model based on the computational rationality to choose among reasoning strategies based on the trade-off between utility and reasoning time. CoXAM demonstrated its superior alignment with human decision-making compared to baseline machine learning proxy models. CoXAM successfully replicated and provided cognitive explanations for key findings: that counterfactual tasks are inherently harder, that decision tree rules are harder to recall than weight-based linear factors, and that the optimal XAI explanation type depends heavily on the application data context. Ultimately, CoXAM provides an interpretable cognitive basis for accelerating the debugging and benchmarking of disparate XAI techniques. GenAI Usage Disclosure The authors used generative AI tools (ChatGPT, Gemini) to assist with language refinement and clarity. The authors reviewed and edited all AI-generated content and take full responsibility for the final version of the manuscript. Acknowledgements.We thank Emily Chen, Li Zhaobin and Kevin Chang for their insightful inputs during the exploratory stage of this work. This research is supported by the National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative (Award No. DTC-RGC-09). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore and Infocomm Media Development Authority. References A. Abdul, C. Von Der Weth, M. Kankanhalli, and B. Y. Lim (2020) COGAM: measuring and moderating cognitive load in machine learning model explanations. In Proceedings of the 2020 CHI conference on human factors in computing systems, p. 1–14. Cited by: §8.2, §8.3.2, §8.3.3, §8.6. A. M. Alaa and M. Van der Schaar (2019) Demystifying black-box models with symbolic metamodels. Advances in neural information processing systems 32. Cited by: Table 3. J. R. Anderson, D. Bothell, M. D. Byrne, S. Douglass, C. Lebiere, and Y. Qin (2004) An integrated theory of the mind.. Psychological review 111 (4), p. 1036. Cited by: §5.1. J. R. Anderson, M. Matessa, and C. Lebiere (1997) ACT-r: a theory of higher level cognition and its relation to visual attention. Human–Computer Interaction 12 (4), p. 439–462. External Links: Document Cited by: §8.3.2. E. Angelino, N. Larus-Stone, D. Alabi, M. Seltzer, and C. Rudin (2018a) Learning certifiably optimal rule lists for categorical data. JMLR. Cited by: §3.2. E. Angelino, N. Larus-Stone, D. Alabi, M. Seltzer, and C. Rudin (2018b) Learning certifiably optimal rule lists for categorical data. Journal of Machine Learning Research 18 (234), p. 1–78. Cited by: §8.3.1, Table 3. S. Arora, D. Pruthi, N. Sadeh, W. W. Cohen, Z. C. Lipton, and G. Neubig (2022) Explain, edit, and understand: rethinking user study design for evaluating model explanations. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 36, p. 5277–5285. Cited by: §2.1. M. H. Ashcraft (1992) Cognitive arithmetic: a review of data and theory. Cognition 44 (1-2), p. 75–106. Cited by: §5.4.2. T. Baron, M. D. Levine, and Y. Yeshurun (1994) Exploring with a foveated robot eye system. In Proceedings of the 12th IAPR International Conference on Pattern Recognition, Vol. 2-Conference B: Computer Vision & Image Processing.(Cat. No. 94CH3440-5), p. 377–380. Cited by: §2.3. R. Bhattacharyya, B. Wulfe, D. J. Phillips, A. Kuefler, J. Morton, R. Senanayake, and M. J. Kochenderfer (2022) Modeling human driving behavior through generative adversarial imitation learning. IEEE Transactions on Intelligent Transportation Systems 24 (3), p. 2874–2887. Cited by: §2.3. C. M. Bishop and N. M. Nasrabadi (2006) Pattern recognition and machine learning. Vol. 4, Springer. Cited by: §A.1.1, §5.2.2. J. Y. Bo, P. Hao, and B. Y. Lim (2024) Incremental xai: memorable understanding of ai with incremental explanations. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, p. 1–17. Cited by: §3.2, §3.2, §8.3.1, §8.3.2, §8.6, Table 3, Table 3. R. Bogacz, E. Brown, J. Moehlis, P. Holmes, and J. D. Cohen (2006) The physics of optimal decision making: a formal analysis of models of performance in two-alternative forced-choice tasks.. Psychological review 113 (4), p. 700. Cited by: §5.2. A. Boggust, B. Hoover, A. Satyanarayan, and H. Strobelt (2022) Shared interest: measuring human-ai alignment to identify recurring patterns in model behavior. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, p. 1–17. Cited by: §8.3.3, §8.6. V. Borisov, T. Leemann, K. Seßler, J. Haug, et al. (2022) Deep neural networks and tabular data: a survey. IEEE Transactions on Knowledge and Data Engineering. Cited by: §1, §3. L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone (1984) Classification and regression trees. Wadsworth. Cited by: §1. L. Breiman, J. Friedman, R. A. Olshen, and C. J. Stone (2017) Classification and regression trees. Chapman and Hall/CRC. Cited by: Table 3. J. S. Bruner, J. J. Goodnow, and G. A. Austin (1956) A study of thinking. Wiley. Cited by: §1, §3. Z. Buçinca, M. B. Malaya, and K. Z. Gajos (2021) To trust or to think: cognitive forcing functions can reduce overreliance on ai in ai-assisted decision-making. Proceedings of the ACM on Human-computer Interaction 5 (CSCW1), p. 1–21. Cited by: §8.5, §8.6. L. Cao and ZhiMin (2019) An overview of deep reinforcement learning. In Proceedings of the 2019 4th International Conference on Automation, Control and Robotics Engineering, p. 1–9. Cited by: §2.3. R. Caruana, Y. Lou, J. Gehrke, P. Koch, M. Sturm, and N. Elhadad (2015) Intelligible models for healthcare: predicting pneumonia risk and hospital 30-day readmission. In Proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, p. 1721–1730. Cited by: §1, §1, §3. J. Caterson, O. Ambler, N. Cereceda-Monteoliva, M. Horner, A. Jones, and A. T. Poacher (2024) Application of generative language models to orthopaedic practice. BMJ open 14 (3), p. e076484. Cited by: §1. K. Charmaz (2014) Constructing grounded theory (introducing qualitative methods series). Constr. grounded theory. Cited by: §4.2. V. Chen, N. Johnson, N. Topin, G. Plumb, and A. Talwalkar (2022) Use-case-grounded simulations for explanation evaluation. Advances in neural information processing systems 35, p. 1764–1775. Cited by: §2.2, §7.1. C. Choi, S. Malla, A. Patil, and J. H. Choi (2021) DROGON: a trajectory prediction model based on intention-conditioned behavior reasoning. In Conference on Robot Learning, p. 49–63. Cited by: §2.3. A. T. Corbett and J. R. Anderson (1994) Knowledge tracing: modeling the acquisition of procedural knowledge. User modeling and user-adapted interaction 4 (4), p. 253–278. Cited by: §1. J. Czerlinski, G. Gigerenzer, and D. G. Goldstein (1999) How good are simple heuristics?. In Simple heuristics that make us smart, p. 97–118. Cited by: §2.1. R. M. Dawes and B. Corrigan (1974) Linear models in decision making.. Psychological bulletin 81 (2), p. 95. Cited by: §2.1. R. M. DAWES (1979) The robust beauty of improper linear models in decision making. AMERICAN PSYCHOLOGIST, p. 571. Cited by: §2.1. C. Deng, Y. Duan, X. Jin, H. Chang, Y. Tian, H. Liu, Y. Wang, K. Gao, H. P. Zou, Y. Jin, et al. (2025) Deconstructing the ethics of large language models from long-standing issues to new-emerging dilemmas: a survey. AI and Ethics 5 (5), p. 4745–4771. Cited by: §8.3.2. A. D. Dobson, E. De Lange, A. Keane, H. Ibbett, and E. Milner-Gulland (2019) Integrating models of human behaviour between the individual and population levels to inform conservation interventions. Philosophical Transactions of the Royal Society B 374 (1781), p. 20180053. Cited by: §2.3. F. Doshi-Velez and B. Kim (2017) Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608. Cited by: §3.3, §8.3.3, footnote 1. U. Ehsan, P. Tambwekar, L. Chan, B. Harrison, and M. O. Riedl (2019) Automated rationale generation: a technique for explainable ai and its effects on human perceptions. In Proceedings of the 24th international conference on intelligent user interfaces, p. 263–274. Cited by: §8.4. H. J. Einhorn and R. M. Hogarth (1975) Unit weighting schemes for decision making. Organizational behavior and human performance 13 (2), p. 171–192. Cited by: §2.1. X. Fang, W. Xu, F. A. Tan, J. Zhang, Z. Hu, Y. Qi, S. Nickleach, D. Socolinsky, S. Sengamedu, and C. Faloutsos (2024) Large language models (llms) on tabular data: prediction, generation, and understanding–a survey. arXiv preprint arXiv:2402.17944. Cited by: §8.3.2. A. Fanous, J. Goldberg, A. A. Agarwal, et al. (2025) SycEval: evaluating llm sycophancy. arXiv preprint arXiv:2502.08177. Cited by: §8.3.2. M. Fiori, G. Civitarese, and C. Bettini (2024) Using large language models to compare explainable models for smart home human activity recognition. In Companion of the 2024 on ACM International Joint Conference on Pervasive and Ubiquitous Computing, p. 881–884. Cited by: §2.2. J. H. Friedman (1991) Multivariate adaptive regression splines. The annals of statistics 19 (1), p. 1–67. Cited by: §8.3.1, Table 3. I. Fründ, N. V. Haenel, and F. A. Wichmann (2011) Inference for psychometric functions in the presence of nonstationary behavior. Journal of vision 11 (6), p. 16–16. Cited by: §5.4.1. A. Fuchs, A. Passarella, and M. Conti (2023) Modeling, replicating, and predicting human behavior: a survey. ACM Transactions on Autonomous and Adaptive Systems 18 (2), p. 1–47. Cited by: §2.3. G. Gigerenzer and D. G. Goldstein (1996a) Reasoning the fast and frugal way: models of bounded rationality.. Psychological review 103 (4), p. 650. Cited by: §2.1. G. Gigerenzer and D. G. Goldstein (1996b) Reasoning the fast and frugal way: models of bounded rationality. Psychological Review 103 (4), p. 650–669. Cited by: §1, §3. B. Glaser and A. Strauss (2017) Discovery of grounded theory: strategies for qualitative research. Routledge. Cited by: §4.2. N. D. Goodman, J. B. Tenenbaum, J. Feldman, and T. L. Griffiths (2008) A rational analysis of rule-based concept learning.. Cognitive Science 32 (1), p. 108–154. Cited by: §1. L. Grinsztajn, E. Oyallon, and G. Varoquaux (2022) Why do tree-based models still outperform deep learning on typical tabular data?. Advances in neural information processing systems 35, p. 507–520. Cited by: §1, §3. P. Hase and M. Bansal (2020) Evaluating explainable ai: which algorithmic explanations help users predict model behavior?. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, p. 5540–5552. Cited by: §1, §2.1. T. Hastie and R. Tibshirani (1990) Generalized additive models. Chapman and Hall. Cited by: §8.3.1, Table 3. S. Hilgard, N. Rosenfeld, M. R. Banaji, J. Cao, and D. Parkes (2021) Learning representations by humans, for humans. In International conference on machine learning, p. 4227–4238. Cited by: §2.2. A. H. Hwang, M. S. Bernstein, S. S. Sundar, R. Zhang, M. Horta Ribeiro, Y. Lu, S. Chang, T. Wu, A. Yang, D. Williams, et al. (2025) Human subjects research in the age of generative ai: opportunities and challenges of applying llm-simulated data to hci studies. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, p. 1–7. Cited by: §8.4. Z. Ji, N. Lee, R. Frieske, et al. (2023) Survey of hallucination in natural language generation. ACM Computing Surveys. Cited by: §8.3.2. B. E. John, K. Prevas, D. D. Salvucci, and K. Koedinger (2004) Predictive human performance modeling made easy. In Proceedings of the SIGCHI conference on Human factors in computing systems, p. 455–462. Cited by: §2.3. P. N. Johnson-Laird (1983) Mental models: towards a cognitive science of language, inference, and consciousness. Harvard University Press. Cited by: §8.3.2. J. Jokinen, A. Acharya, M. Uzair, X. Jiang, and A. Oulasvirta (2021) Touchscreen typing as optimal supervisory control. In Proceedings of the 2021 CHI conference on human factors in computing systems, p. 1–14. Cited by: §1. D. Kahneman (2011) Thinking, fast and slow. macmillan. Cited by: §5.2.2. M. Kahng, P. Y. Andrews, A. Kalro, and D. H. Chau (2017) A cti v is: visual exploration of industry-scale deep neural network models. IEEE transactions on visualization and computer graphics 24 (1), p. 88–97. Cited by: §8.6. M. L. Kalish, S. Lewandowsky, and J. K. Kruschke (2004) Population of linear experts: knowledge partitioning and function learning.. Psychological review 111 (4), p. 1072. Cited by: §8.3.1, Table 3. N. Karelaia and R. M. Hogarth (2008) Determinants of linear judgment: a meta-analysis of lens model studies.. Psychological bulletin 134 (3), p. 404. Cited by: §5.2.2. W. G. Kennedy (2011) Modelling human behaviour in agent-based models. In Agent-based models of geographical systems, p. 167–179. Cited by: §2.3. B. Kim, R. Khanna, and O. Koyejo (2016) Examples are not enough, learn to criticize! criticism for interpretability. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §8.3.1, Table 3. B. Kim, M. Wattenberg, J. Gilmer, C. Cai, J. Wexler, F. Viegas, et al. (2018) Interpretability beyond feature attribution: quantitative testing with concept activation vectors (tcav). In International conference on machine learning, p. 2668–2677. Cited by: §8.3.2. A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W. Lo, P. Dollár, and R. Girshick (2023) Segment anything. arXiv:2304.02643. Cited by: §8.3.2. K. R. Koedinger, J. R. Anderson, W. H. Hadley, and M. A. Mark (1997) Intelligent tutoring goes to school in the big city. International journal of artificial intelligence in education 8, p. 30–43. Cited by: §1. P. W. Koh, T. Nguyen, Y. S. Tang, S. Mussmann, E. Pierson, B. Kim, and P. Liang (2020) Concept bottleneck models. In International conference on machine learning, p. 5338–5348. Cited by: §8.3.2. S. Kolekar, S. Gite, B. Pradhan, and K. Kotecha (2021) Behavior prediction of traffic actors for intelligent vehicle using artificial intelligence techniques: a review. IEEE Access 9, p. 135034–135058. Cited by: §2.3. M. Kozielski, M. Sikora, and Ł. Wawrowski (2025) Towards consistency of rule-based explainer and black box model—fusion of rule induction and xai-based feature importance. Knowledge-Based Systems 311, p. 113092. Cited by: §3.2. J. Krause, A. Perer, and K. Ng (2016) Interacting with predictions: visual inspection of black-box machine learning models. In Proceedings of the 2016 CHI conference on human factors in computing systems, p. 5686–5697. Cited by: Table 3. I. Lage, A. Ross, S. J. Gershman, B. Kim, and F. Doshi-Velez (2018) Human-in-the-loop interpretability prior. Advances in neural information processing systems 31. Cited by: §2.2. V. Lai, S. Carton, R. Bhatnagar, Q. V. Liao, Y. Zhang, and C. Tan (2022) Human-ai collaboration via conditional delegation: a case study of content moderation. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, p. 1–18. Cited by: §2.2, §8.6. H. Lakkaraju, S. H. Bach, and J. Leskovec (2016a) Interpretable decision sets: a joint framework for description and prediction. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, p. 1675–1684. Cited by: §1, §8.6, Table 3. H. Lakkaraju, S. H. Bach, and J. Leskovec (2016b) Interpretable decision sets: a joint framework for description and prediction. KDD. Cited by: §8.3.1. M. Langer, D. Oster, T. Speith, H. Hermanns, L. Kästner, E. Schmidt, A. Sesing, and K. Baum (2021) What do we want from explainable artificial intelligence (xai)?–a stakeholder perspective on xai and a conceptual model guiding interdisciplinary xai research. Artificial intelligence 296, p. 103473. Cited by: §8.5. L. Le Mero, D. Yi, M. Dianati, and A. Mouzakitis (2022) A survey on imitation learning techniques for end-to-end autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems 23 (9), p. 14128–14147. Cited by: §2.3. R. L. Lewis and S. Vasishth (2005) An activation-based model of sentence processing as skilled memory retrieval. Cognitive Science 29 (3), p. 375–419. Cited by: §8.3.2. Z. Li and M. Yin (2024) Utilizing human behavior modeling to manipulate explanations in ai-assisted decision making: the good, the bad, and the scary. Advances in Neural Information Processing Systems 37, p. 5025–5047. Cited by: §8.2, §8.5. F. Lieder and T. L. Griffiths (2017) Strategy selection as rational metareasoning. Psychological review 124 (6), p. 762–794. Cited by: §2.3, §5.4.3. F. Lieder and T. L. Griffiths (2020) Resource-rational analysis: understanding human cognition as the optimal use of limited computational resources. Behavioral and brain sciences 43, p. e1. Cited by: §2.3. B. Y. Lim, J. P. Cahaly, C. Y. Sng, and A. Chew (2025) Diagrammatization and abduction to improve ai interpretability with domain-aligned explanations for medical diagnosis. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, p. 1–25. Cited by: §8.3.2, §8.6. B. Y. Lim, A. K. Dey, and D. Avrahami (2009) Why and why not explanations improve the intelligibility of context-aware intelligent systems. In Proceedings of the SIGCHI conference on human factors in computing systems, p. 2119–2128. Cited by: §8.3.3, §8.6. C. G. Lucas, T. L. Griffiths, J. J. Williams, and M. L. Kalish (2015) A rational model of function learning. Psychonomic bulletin & review 22 (5), p. 1193–1215. Cited by: §8.3.1, Table 3. S. M. Lundberg and S. Lee (2017) A unified approach to interpreting model predictions. Advances in neural information processing systems 30. Cited by: §1, §8.3.2. Y. Lyu, H. Lu, M. K. Lee, G. Schmitt, and B. Y. Lim (2024) IF-city: intelligible fair city planning to measure, explain and mitigate inequality. IEEE Transactions on Visualization and Computer Graphics 30 (7), p. 3749–3766. Cited by: §8.6. S. Ma, Y. Lei, X. Wang, C. Zheng, C. Shi, M. Yin, and X. Ma (2023) Who should i trust: ai or myself? leveraging human and ai correctness likelihood to promote appropriate trust in ai-assisted decision-making. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, p. 1–19. Cited by: §2.2, §2.2, §8.6. D. J. MacKay (1992) A practical bayesian framework for backpropagation networks. Neural computation 4 (3), p. 448–472. Cited by: §A.1.1, §5.2.2. W. T. Maddox and F. G. Ashby (2004) Discriminating between rule-based and information-integration categorization processes. Journal of Experimental Psychology: Learning, Memory, and Cognition 30 (2), p. 316–333. Cited by: §1, §3. P. E. Meehl (1954) Clinical versus statistical prediction. Vol. 1, University of Minnesota Press Minneapolis. Cited by: §2.1. J. Memelink and B. Hommel (2013) Intentional weighting: a basic principle in cognitive control. Psychological Research 77 (3), p. 249–259. Cited by: §1. M. Mersha, K. Lam, J. Wood, A. K. Alshami, and J. Kalita (2024) Explainable artificial intelligence: a survey of needs, techniques, applications, and future direction. Neurocomputing 599, p. 128111. Cited by: §1. H. Mozannar, A. Satyanarayan, and D. Sontag (2022) Teaching humans when to defer to a classifier via exemplars. In Proceedings of the aaai conference on artificial intelligence, Vol. 36, p. 5323–5331. Cited by: §2.2. K. P. Murphy (2012) Machine learning: a probabilistic perspective. MIT press. Cited by: §A.1.1, §5.2.2. A. Narayanan and K. Bergen (2024) Prototype-based methods in explainable ai and emerging opportunities in the geosciences. In ICML 2024 AI for Science Workshop, Cited by: §8.3.1, Table 3. D. J. Navarro and I. G. Fuss (2009) Fast and accurate calculations for first-passage times in wiener diffusion models. Journal of mathematical psychology 53 (4), p. 222–230. Cited by: §5.2. R. M. Nosofsky (1986) Attention, similarity, and the identification–categorization relationship.. Journal of experimental psychology: General 115 (1), p. 39. Cited by: §8.3.1, Table 3. A. Oulasvirta, J. P. Jokinen, and A. Howes (2022) Computational rationality as a theory of interaction. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, p. 1–14. Cited by: §1, §2.3, §2.3. C. Panigutti, A. Beretta, D. Fadda, F. Giannotti, D. Pedreschi, A. Perotti, and S. Rinzivillo (2023) Co-design of human-centered, explainable ai for clinical decision support. ACM Transactions on Interactive Intelligent Systems 13 (4), p. 1–35. Cited by: §8.6. J. S. Park, J. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein (2023) Generative agents: interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, p. 1–22. Cited by: §8.4. J. S. Park, C. Q. Zou, A. Shaw, B. M. Hill, C. Cai, M. R. Morris, R. Willer, P. Liang, and M. S. Bernstein (2024) Generative agent simulations of 1,000 people. arXiv preprint arXiv:2411.10109. Cited by: §8.4. C. Paulo, C. A., A. F., M. T., and R. J. (2009) Wine Quality. Note: UCI Machine Learning RepositoryDOI: https://doi.org/10.24432/C56S3T Cited by: §3.4. C. Piech, J. Huang, A. Nguyen, M. Phulsuksombati, M. Sahami, and L. Guibas (2015) Learning program embeddings to propagate feedback on student code. In International conference on machine Learning, p. 1093–1102. Cited by: §1. F. Poursabzi-Sangdeh, D. G. Goldstein, J. M. Hofman, J. W. Wortman Vaughan, and H. Wallach (2021) Manipulating and measuring model interpretability. In Proceedings of the 2021 CHI conference on human factors in computing systems, p. 1–52. Cited by: §1, §3.2, §8.3.2, §8.6, Table 3. N. Prins (2012) The psychometric function: the lapse rate revisited. Journal of Vision 12 (6), p. 25–25. Cited by: §5.4.1. N. Rabinowitz, F. Perbet, F. Song, C. Zhang, S. A. Eslami, and M. Botvinick (2018) Machine theory of mind. In International conference on machine learning, p. 4218–4227. Cited by: §2.3. R. Ratcliff and G. McKoon (2008) The diffusion decision model: theory and data for two-choice decision tasks. Neural computation 20 (4), p. 873–922. Cited by: §5.4.1. R. Ratcliff, P. L. Smith, S. D. Brown, and G. McKoon (2016) Diffusion decision model: current issues and history. Trends in cognitive sciences 20 (4), p. 260–281. Cited by: §1. R. Ratcliff and P. L. Smith (2004) A comparison of sequential sampling models for two-choice reaction time.. Psychological review 111 (2), p. 333. Cited by: §5.2. R. Ratcliff (1978) A theory of memory retrieval. Psychological Review 85 (2), p. 59–108. External Links: Document Cited by: §5.2. R. Ratcliff (1993) Methods for dealing with reaction time outliers.. Psychological bulletin 114 (3), p. 510. Cited by: §1. K. Rayner (1998) Eye movements in reading and information processing: 20 years of research.. Psychological bulletin 124 (3), p. 372. Cited by: §5.4.2. D. Reichman, J. C. Peterson, and T. L. Griffiths (2024) Machine learning for modeling human decisions.. Decision 11 (4), p. 619. Cited by: §2.2. M. T. Ribeiro, S. Singh, and C. Guestrin (2016) ” Why should i trust you?” explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, p. 1135–1144. Cited by: §1, §8.3.2. M. T. Ribeiro, S. Singh, and C. Guestrin (2018) Anchors: high-precision model-agnostic explanations. In Proceedings of the AAAI conference on artificial intelligence, Vol. 32. Cited by: §1, §2.1. K. Rivers (2017) Automated data-driven hint generation for learning programming. Ph.D. Thesis, Carnegie Mellon University. Cited by: §1. Y. Rong, T. Leemann, T. Nguyen, L. Fiedler, P. Qian, V. Unhelkar, T. Seidel, G. Kasneci, and E. Kasneci (2023) Towards human-centered explainable ai: a survey of user studies for model explanations. IEEE transactions on pattern analysis and machine intelligence 46 (4), p. 2104–2122. Cited by: §8.3.3, §8.6. C. Rudin (2019) Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence 1 (5), p. 206–215. Cited by: §8.3.2, §8.4. J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov (2017) Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. Cited by: §5.4.4. G. Schwalbe and B. Finzel (2024) A comprehensive taxonomy for explainable artificial intelligence: a systematic survey of surveys on methods and concepts. Data Mining and Knowledge Discovery 38 (5), p. 3043–3101. Cited by: §1. R. R. e. al. Selvaraju (2017) Grad-cam: visual explanations from deep networks via gradient-based localization. In ICCV, Cited by: §8.3.2. R. N. Shepard, C. I. Hovland, and H. M. Jenkins (1961) Learning and memorization of classifications. Psychological Monographs 75 (13), p. 1–42. Cited by: §1, §3. D. Shi, Y. Wang, Y. Bai, A. Bulling, and A. Oulasvirta (2025) Chartist: task-driven eye movement control for chart reading. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, p. 1–14. Cited by: §8.3.2, §8.4. D. Shi, Y. Zhu, J. P. Jokinen, A. Acharya, A. Putkonen, S. Zhai, and A. Oulasvirta (2024) CRTypist: simulating touchscreen typing behavior via computational rationality. In Proceedings of the CHI Conference on Human Factors in Computing Systems, p. 1–17. Cited by: §1, §2.3. R. Shwartz-Ziv and A. Armon (2022) Tabular data: deep learning is not all you need. Information Fusion 81, p. 84–90. Cited by: §8.3.2. H. A. Simon (1957) Models of man: social and rational; mathematical essays on rational human behavior in society setting. New York: Wiley. Cited by: §2.3. [122] J. Skirzynski, E. Glassman, and B. Ustun On interpretability and overreliance. In Interpretable AI: Past, Present and Future, Cited by: §8.5. D. Slack, S. Friedler, C. Roy, and C. Scheidegger (2019) Assessing the local interpretability of machine learning models.. In NeurIPS Workshop on Human-Centric Machine Learning (HCML), Cited by: §2.1. J. Snoek, H. Larochelle, and R. P. Adams (2012) Practical bayesian optimization of machine learning algorithms. Advances in neural information processing systems 25. Cited by: §7.1. M. Strathern (1997) ‘Improving ratings’: audit in the british university system. European review 5 (3), p. 305–321. Cited by: §8.5. A. Strauss, J. Corbin, et al. (1990) Basics of qualitative research. Vol. 15, sage Newbury Park, CA. Cited by: §4.2. L. Sun, T. Qin, A. Hu, J. Zhang, S. Lin, J. Chen, M. Ali, and M. Prpa (2025) Persona-l has entered the chat: leveraging llms and ability-based framework for personas of people with complex needs. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, p. 1–31. Cited by: §8.4. R. S. Sutton (2018) Reinforcement learning: an introduction. A Bradford Book. Cited by: §2.3. R. L. Thomas and D. Uminsky (2022) Reliance on metrics is a fundamental challenge for AI. Patterns 3 (5), p. 100476. External Links: Document Cited by: §8.5. Y. Tian, F. Xia, and Y. Song (2024) Large language models are no longer shallow parsers. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 7131–7142. Cited by: §8.3.2. M. Vaccaro, A. Almaatouq, and T. Malone (2024) When combinations of humans and ai are useful: a systematic review and meta-analysis. Nature Human Behaviour 8 (12), p. 2293–2303. Cited by: §2.2. C. M. van Leersum and C. Maathuis (2025) Human centred explainable ai decision-making in healthcare. Journal of Responsible Technology 21, p. 100108. Cited by: §8.6. H. Vasconcelos, M. Jörke, M. Grunde-McLaughlin, T. Gerstenberg, M. S. Bernstein, and R. Krishna (2023) Explanations can reduce overreliance on ai systems during decision-making. Proceedings of the ACM on Human-Computer Interaction 7 (CSCW1), p. 1–38. Cited by: §8.5. M. Virgolin, A. De Lorenzo, E. Medvet, and F. Randone (2020) Learning a formula of interpretability to learn interpretable formulas. In Parallel Problem Solving from Nature–PPSN XVI: 16th International Conference, PPSN 2020, Leiden, The Netherlands, September 5-9, 2020, Proceedings, Part I 16, p. 79–93. Cited by: §2.2. J. v. d. Waa, E. Nieuwburg, A. Cremers, and M. Neerincx (2021) Evaluating xai: a comparison of rule-based and example-based explanations. Artificial Intelligence 291, p. 103404. External Links: Document Cited by: §8.3.3. D. Wagner, D. Heider, and G. Hattab (2021) Secondary mushroom. Note: UCI Machine Learning Repositoryhttps://doi.org/10.24432/C5FP5Q Cited by: §3.4. D. Wang, Q. Yang, A. Abdul, and B. Y. Lim (2019) Designing theory-driven user-centric explainable ai. In Proceedings of the 2019 CHI conference on human factors in computing systems, p. 1–15. Cited by: §8.5. X. Wang, Z. Lu, and M. Yin (2022) Will you accept the ai recommendation? predicting human behavior in ai-assisted decision making. In Proceedings of the ACM web conference 2022, p. 1697–1708. Cited by: §2.2, §8.3.3. Y. Wang, A. R. Srinivasan, J. P. Jokinen, A. Oulasvirta, and G. Markkula (2025) Pedestrian crossing decisions can be explained by bounded optimal decision-making under noisy visual perception. Transportation Research Part C: Emerging Technologies 171, p. 104963. Cited by: §2.3. J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al. (2022) Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35, p. 24824–24837. Cited by: §8.4. F. A. Wichmann and N. J. Hill (2001) The psychometric function: i. fitting, sampling, and goodness of fit. Perception & psychophysics 63 (8), p. 1293–1313. Cited by: §5.4.1. T. V. Wiecki and M. J. Frank (2013) A computational model of inhibitory control in frontal cortex and basal ganglia.. Psychological review 120 (2), p. 329. Cited by: §5.2. C. Witton, J. B. Talcott, and G. B. Henning (2017) Psychophysical measurements in children: challenges, pitfalls, and considerations. PeerJ 5, p. e3231. Cited by: §5.4.1. J. M. Wolfe (2021) Guided search 6.0: an updated model of visual search. Psychonomic bulletin & review 28 (4), p. 1060–1092. Cited by: §5.4.2. G. Yeo, F. T. A. Ting, K. Jaidka, S. Furniturewala, W. Fanyou, W. Xu, V. Jain, A. Chadha, Y. Liu, and S. K. Ng (2025a) PHAnToM: persona-based prompting has an effect on theory-of-mind reasoning in large language models. In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 19, p. 2124–2142. Cited by: §8.4. W. J. Yeo, W. Van Der Heever, R. Mao, E. Cambria, R. Satapathy, and G. Mengaldo (2025b) A comprehensive review on financial explainable ai. Artificial Intelligence Review 58 (6), p. 1–49. Cited by: §1. X. Zeng, Y. H. Sun, F. Gao, L. Hua, S. Xu, and Z. Yuan (2024) Concurrent behavioral modeling and multimodal neuroimaging reveals how feedback affects the performance of decision making in internet gaming disorder. NeuroImage 297, p. 120726. Cited by: §2.3. Appendix A Appendix A.1. Technical modeling details In this section we detail the strategy update method for the feature attribution strategy as well as the hyperparameters of Reinforcement Learning controller in CoXAM. A.1.1. Update method for the Feature Attribution strategy After each trial, the Feature Attribution strategy updates its internal Gaussian beliefs over each coefficient using the observed AI prediction y^∈0,1 y∈\0,1\ and the model’s predicted probability p=P(y=y^|e)p=P(y= y|e). Each feature parameter is represented as a Gaussian distribution ωρ∼(μρ,σρ2) _ρ ( _ρ, _ρ^2), where ρ indexes the feature dimension. We perform an incremental Bayesian update following a Laplace approximation for logistic regression with a Gaussian prior (Bishop and Nasrabadi, 2006; Murphy, 2012; MacKay, 1992). For each feature ρ with input value xρx_ρ, we compute the curvature term χ=p(1−p),χ=p(1-p), and update the posterior precision, variance, and mean as: (26) λρ′ _ρ =1σρ2+χxρ2, = 1 _ρ^2+χ x_ρ^2, (27) σρ′⁣2 _ρ 2 =1λρ′, = 1 _ρ , (28) μρ′ _ρ =μρ+σρ′⁣2xρ(y^−p). = _ρ+ _ρ 2x_ρ( y-p). where λρ′ _ρ is the updated precision, σρ′⁣2 _ρ 2 is the updated variance and μρ′ _ρ is the updated mean. These updates refine each feature’s mean and variance based on the discrepancy between the predicted probability p and the observed AI output y y, while weighting the change by both attribute magnitude xρx_ρ and current uncertainty σρ2 _ρ^2. A.1.2. Reinforcement Learning Training We detail the hyperparameters for the PPO algorithm that was trained using stable baselines 3: Table 4. Hyperparameters used for counterfactual RL training (PPO). Setting Value Algorithm PPO Policy MlpPolicy Policy activation Tanh Policy network (actor) [64, 64] Value network (critic) [64, 64] Learning rate 3×10−43× 10^-4 Entropy coefficient (ent_coef) 0.01 Discount factor (γ) 0.8 Rollout steps per env (n_steps) 1024 Device cpu Verbose 0 Total timesteps 400,000 Number of vector envs (n_envsn\_envs) 8 (SubprocVecEnv) Eval frequency 1250 steps Env: instances per episode 40 Env: max attributes 6 A.2. Full summative study results Figure A1. Forward and Counterfactual Simulation Accuracy, measured with respect to AI outputs across all conditions, and dataset application scenarios. Shows four line plots comparing forward and counterfactual simulation accuracy for the Wine Quality and Mushrooms datasets. The top row shows forward accuracy and the bottom row shows counterfactual accuracy. The left column corresponds to Wine Quality and the right column to Mushrooms. Results are shown for three explanation types (Rules, Weights, and Hybrid), each under low and high conditions. Red circle markers indicate performance without XAI and blue plus markers indicate performance with XAI. Error bars show variability. Overall, forward accuracy is higher than counterfactual accuracy, and the effect of XAI varies across explanation types, conditions, and datasets. A.3. Modeling Results Figure A2. Normalized absolute change in attribute values across different XAI Schemas, datasets and with or without XAI trials. Figure A2 shows two line plots of normalized absolute change in attribute values for the Wine Quality dataset (left) and the Mushrooms dataset (right). The x-axis lists explanation types (Rules, Weights, Hybrid) and the y-axis shows normalized absolute change in percent. Red circle markers indicate trials without XAI and blue plus markers indicate trials with XAI, with error bars showing variability. Across both datasets, the Weights explanation shows the largest changes, followed by Hybrid, with Rules showing the smallest changes. A.3.1. Individual accuracies and timings of each strategy Table. 5 and Table. 6 show the accuracies and timings of the individual strategies for forward and counterfactual tasks respectively. Table 5. Forward simulation performance (mean ± SE) across reasoning strategies, XAI conditions, and datasets. Accuracy is reported in %, and time in seconds. Dataset Tested w/ XAI Strategy Accuracy (%) Time (s) Mushrooms w/ XAI Attribute-Threshold-Attribute Traversal 74.6 ± 1.1 9.2 ± 0.1 Approximate Calculation 64.6 ± 1.4 11.0 ± 0.2 Feature Attribution 49.4 ± 5.3 11.2 ± 0.6 w/o XAI Attribute-Threshold-Attribute Traversal 63.2 ± 1.2 8.9 ± 0.1 Approximate Calculation 60.9 ± 1.5 10.7 ± 0.2 Feature Attribution 56.1 ± 4.8 12.3 ± 0.6 Wine Quality w/ XAI Attribute-Threshold-Attribute Traversal 72.1 ± 1.3 8.7 ± 0.1 Approximate Calculation 48.6 ± 5.5 11.4 ± 1.2 Feature Attribution 75.5 ± 1.9 11.2 ± 0.2 w/o XAI Attribute-Threshold-Attribute Traversal 65.5 ± 1.3 8.3 ± 0.1 Approximate Calculation 49.1 ± 6.2 9.4 ± 1.2 Feature Attribution 73.5 ± 1.9 11.0 ± 0.2 Table 6. Counterfactual accuracy (mean ± SE, %) across reasoning strategies, whether XAI is shown, and datasets. Accuracy is reported against both the AI and XAI. Dataset Tested w/ XAI Strategy Accuracy vs. AI (%) Accuracy vs. XAI (%) Mushrooms w/ XAI Availability Heuristic 32.7 ± 2.5 43.0 ± 2.6 Inverse Calculation 48.9 ± 3.0 69.5 ± 2.7 Inverse Feature Attribution 44.9 ± 2.1 71.4 ± 1.9 Node Threshold Crossing 49.0 ± 1.3 69.4 ± 1.2 w/o XAI Availability Heuristic 30.7 ± 1.2 43.4 ± 1.2 Inverse Feature Attribution 45.8 ± 1.7 70.4 ± 1.6 Node Threshold Crossing 45.1 ± 2.0 68.3 ± 1.9 Wine Quality w/ XAI Availability Heuristic 20.9 ± 2.3 31.4 ± 2.6 Inverse Calculation 47.3 ± 3.7 57.0 ± 3.6 Inverse Feature Attribution 50.3 ± 2.5 55.8 ± 2.5 Node Threshold Crossing 33.2 ± 1.1 61.3 ± 1.2 w/o XAI Availability Heuristic 24.6 ± 1.0 34.7 ± 1.2 Inverse Feature Attribution 50.1 ± 2.3 55.3 ± 1.5 Node Threshold Crossing 29.3 ± 1.8 53.8 ± 2.0 A.4. Analysis of cognitive parameters Figure A3. Variations in CoXAM performance against cognitive parameters. a) Strategy accuracy in the forward simulation task against cognitive parameters and specific trial strategies. b) Attribute selection, X of the six attributes in Wine Quality dataset against opportunity cost, γ. This figure illustrates how forward simulation performance varies with cognitive parameters and task strategies. Panel (a) plots human and CoXAM forward accuracy as a function of three fitted cognitive parameters—opportunity cost ($γ$), diffusion noise ($ν$), and retrieval threshold ($κ$)—for each explanation type (Rules, Weights, Hybrid). Shaded regions indicate 95% confidence intervals. Accuracy declines as cognitive costs increase, particularly under higher γ and ν, showing greater sensitivity for strategies requiring more computation. Panel (b) shows the distribution of attribute selections (X) across the six Wine Quality attributes as a function of opportunity cost. Participants increasingly focused on fewer, more diagnostic attributes (e.g., Alcohol, Vinegar Taint) as $γ$ rose, reflecting selective attention under higher cognitive effort. Figure A4. Variation in strategy accuracy in counterfactual task with varying cognitive parameters. Error areas are 95% CI. This figure shows how counterfactual accuracy changes across reasoning strategies as cognitive parameters vary. Each row corresponds to one XAI Schema: Rules, Weights, and Hybrid. The three columns represent variations in opportunity cost gamma, margin epsilon, and retrieval threshold kappa. The four curves represent the strategies Inverse Calculation, Inverse Feature Attribution, Node Threshold Crossing, and Availability Heuristic. Shaded areas show 95 percent confidence intervals. Accuracy trends show that Inverse Feature Attribution and Node Threshold Crossing maintain stable performance under higher cognitive cost, while the Availability Heuristic declines more sharply with increasing gamma or kappa. A.5. Counterfactual Selection Probabilities of CoXAM Figure A5. Attribute Selection by dataset and XAI Schemas (CoXAM Counterfactual Simulation Results). A.6. Survey Flow: User Study and XAI UI Figure A6. Introduction to User Study. A study start page interface showing a title introducing a research study on AI understanding, a short description of the study purpose, a list of study steps, an estimated completion time, and a button to proceed. Figure A7. System UI of User Study and Tutorial. An application scenario screen explaining that an AI predicts one of two wine types based on chemical attributes. A table lists attributes such as alcohol, sulphates, sulfur dioxide, vinegar taint, pH, and chlorides, along with their descriptions and minimum and maximum values. Figure A8. UI of Data Instance (Attribute Name, Value and Visual Slider). A tutorial screen showing a single data instance. Each row lists an attribute name, its numeric value, and a horizontal bar indicating the relative magnitude of that value. An AI prediction label is shown to the right. Figure A9. Screening Question (Data and Prediction Task). A multiple-choice screening question asking participants to identify correct statements about how the data values, bars, and AI prediction should be interpreted. Figure A10. Tutorial for Weight-based XAI (top) and screening question (bottom). A tutorial screen explaining a weight-based AI explanation. Attributes and values are shown alongside fixed numeric factors, partial sums for each attribute, and a total contribution. The summed value leads to an explained prediction, which may or may not match the AI’s prediction. Also, below, a multiple-choice screening question asking participants to identify correct statements about how the weight-based explanation works, including whether factors are constant and how the final prediction is determined. Figure A11. Tutorial for Rule-based XAI (top) and screening question (bottom). A tutorial screen explaining a rule-based AI explanation using a decision tree. The tree checks attributes one at a time against thresholds, follows a single highlighted path, and ends at a predicted class. Also, below, a multiple-choice screening question asking participants to identify correct statements about how the rule-based explanation should be read and interpreted. Figure A12. Start page of 40-trial forward simulation session. A session introduction page explaining that participants will complete multiple trials where they learn how the AI predicts one of two classes. Instructions emphasize learning from feedback and answering thoughtfully. Figure A13. User predicts AI output without XAI. A task screen without explanation where participants view attribute values with bars and must estimate the AI’s prediction by selecting one of two options. Figure A14. Feedback on AI output without XAI. A feedback screen showing the correct AI prediction for the previous case, alongside the same attribute values and bars. Figure A15. User predict AI output with the Weight-based XAI. A task screen with a weight-based explanation where participants view attribute values, factors, and partial sums, and then estimate the AI’s prediction. Figure A16. Feedback on AI output with the Weight-based XAI. A feedback screen for the weight-based explanation showing the computed total contribution, the explained prediction, and the AI’s actual prediction. Figure A17. User predicts AI output with the Rule-based XAI. A task screen with a rule-based explanation where participants see attribute values, a decision tree, and must estimate the AI’s prediction. Figure A18. Feedback on AI output with the Rule-based XAI. A feedback screen for the rule-based explanation showing the AI’s prediction and the path taken through the decision tree. Figure A19. Start page of 40-trial counterfactual simulation session. A session introduction page explaining a counterfactual task where participants must change exactly one attribute by as little as possible to flip the AI’s prediction, with no feedback provided on success. Figure A20. Change an attribute value without XAI. A counterfactual task screen without explanation showing current attribute values, the AI’s prediction, and sliders allowing one attribute to be adjusted to change the prediction. Figure A21. Change an attribute value with the Weight-based XAI (No AI feedback). A counterfactual task screen with a weight-based explanation showing attribute values, factors, partial sums, and sliders for adjusting a single attribute, without feedback on whether the prediction changed. Figure A22. Change an attribute value with the Rule-based XAI (No AI feedback). A counterfactual task screen with a rule-based explanation showing attribute values, a decision tree, and sliders for adjusting a single attribute, without feedback on whether the prediction changed.

Instant research discovery