Paper deep dive
DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories
Neemesh Yadav, Palakorn Achananuparp, Jing Jiang, Ee-Peng Lim
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 97%
Last extracted: 4/26/2026, 5:53:15 PM
Summary
DialToM is a human-verified benchmark designed to evaluate Theory of Mind (ToM) in Large Language Models (LLMs) through two lenses: Literal ToM (retrospective mental state inference) and Functional ToM (prospective social forecasting). Unlike existing benchmarks that rely on synthetic or goal-oriented data, DialToM uses naturalistic, high-stakes human dialogues from domains like mental health and persuasion. It introduces 'Trust' as a novel mental state attribute alongside the standard BDI (Belief, Desire, Intention) and adds Emotion and Knowledge. The benchmark reveals a reasoning asymmetry where LLMs excel at identifying mental states but struggle to use them to forecast future dialogue trajectories, a gap the authors call 'Prospective Diagnostic Forecasting'.
Entities (9)
Relation Signals (5)
DialToM → evaluates → Theory of Mind
confidence 100% · We introduce DialToM, a human-verified benchmark built from natural human dialogue... to evaluate... Theory of Mind (ToM) abilities.
Literal ToM → isatypeof → Retrospective Inference
confidence 100% · We categorize these as Literal ToM – the Retrospective task of inferring states from context.
Functional ToM → isatypeof → Prospective Diagnostic Forecasting
confidence 100% · and Functional ToM – the Prospective task of forecasting subsequent dialogue trajectories.
DialToM → usesdatafrom → AnnoMI
confidence 100% · We focus on three domains of conversational data: Motivational Interviewing (MI) sessions... AnnoMI [39]
Gemini-3-Pro → excelsat → Functional ToM
confidence 90% · while LLMs excel at identifying mental states, most (except for Gemini 3 Pro) fail to leverage this understanding to forecast social trajectories.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large Language Models (LLMs) have been shown to possess Theory of Mind (ToM) abilities. However, it remains unclear whether this stems from robust reasoning or spurious correlations. We introduce DialToM, a human-verified benchmark built from natural human dialogue using a multiple-choice framework. We evaluate not only mental state prediction (Literal ToM) but also the functional utility of these states (Functional ToM) through Prospective Diagnostic Forecasting -- probing whether models can identify state-consistent dialogue trajectories solely from mental-state profiles. Our results reveal a significant reasoning asymmetry: while LLMs excel at identifying mental states, most (except for Gemini 3 Pro) fail to leverage this understanding to forecast social trajectories. Additionally, we find only weak semantic similarities between human and LLM-generated inferences. To facilitate reproducibility, the DialToM dataset and evaluation code are publicly available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2604.20443v1
- Canonical: https://arxiv.org/abs/2604.20443v1
Trouble viewing inline? Open PDF directly →
Full Text
80,563 characters extracted from source content.
Expand or collapse full text
DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories Neemesh Yadav Singapore Management University Singapore, Singapore neemeshy@smu.edu.sg Palakorn Achananuparp Singapore Management University Singapore, Singapore palakorna@smu.edu.sg Jing Jiang Australian National University Canberra, Australia Jing.Jiang@anu.edu.au Ee-Peng Lim Singapore Management University Singapore, Singapore eplim@smu.edu.sg Abstract Large Language Models (LLMs) have been shown to possess The- ory of Mind (ToM) abilities. However, it remains unclear whether this stems from robust reasoning or spurious correlations. We in- troduce DialToM, a human-verified benchmark built from natural human dialogue using a multiple-choice framework. We evaluate not only mental state prediction (Literal ToM) but also the func- tional utility of these states (Functional ToM) through Prospective Diagnostic Forecasting — probing whether models can identify state-consistent dialogue trajectories solely from mental-state pro- files. Our results reveal a significant reasoning asymmetry: while LLMs excel at identifying mental states, most (except for Gemini 3 Pro) fail to leverage this understanding to forecast social trajecto- ries. Additionally, we find only weak semantic similarities between human and LLM-generated inferences. To facilitate reproducibility, the DialToM dataset and evaluation code are publicly available at https://github.com/Stealth-py/DialToM. CCS Concepts • Computing methodologies→Theory of mind; Discourse, dialogue and pragmatics; Natural language generation;• General and reference→ Evaluation. Keywords Theory of Mind, Benchmark, Diagnostic Probing, Trust Modeling, Dialogue Systems, Social Forecasting, Large Language Models 1 Introduction Theory of Mind (ToM) is a core component in human cognition that allows us to reason about mental states such as someone’s Beliefs, Desires and Intentions [3]. This mental state is commonly represented by different attributes such as one’s Beliefs, Desires and Intentions, referred to as the Belief-Desire-Intention (BDI) model [6]. These ToM attributes are implicitly used by humans to perform or plan for daily tasks, e.g. how one can change or persuade the perspective of another person [18]? The success of LLMs on role-playing in socially complex conver- sational tasks has prompted researchers to study whether LLMs are capable of complex human-like social interactions [31]. Building LLMs capable of having complex social interactions requires an understanding of how well they can infer mental states (inverse planning) and plan about future actions using these mental states (forward planning) [18]. Existing benchmarks that study this be- havior [20,29,31] focus primarily on inferring mental states in goal-driven synthetic dialogues, and false-belief tasks. A significant gap remains in understanding how these mental states drive the predictability of future decision-making processes in LLMs [34]. We introduce DialToM, a human-verified theory-of-mind rea- soning benchmark over naturalistic human-human conversations, which evaluates both the capability to infer mental states and the ability to leverage them for social forecasting. Unlike existing bench- marks, DialToM also incorporates Trust (T) as a novel mental state attribute, serving as a critical relational factor in high-stakes dyadic dialogues, such as therapeutic and persuasion interactions. We operationalize the transition from mental state inference (inverse planning) to action prediction (forward planning), a framework long-established in human cognition research [4,30]. Following the distinction proposed by Riemer et al. [28], we categorize these as Literal ToM – the Retrospective task of inferring states from context – and Functional ToM – the Prospective task of forecasting subsequent dialogue trajectories. DialToM further follows a very specific nuance for evaluating Functional ToM through Prospective forecasting, where the model needs to predict a dialogue trajectory not based on the conversation but rather only the mental states of the recipient. This formulation: 1) mitigates lexical shortcuts - context-driven forecasting allows models to rely on surface-level linguistic pattern matching rather than Functional ToM; and, 2) forces state utility - by removing context, we create a “state-driven" setup where the model must demonstrate that ToM inferences are functional for trajectory pre- diction. Internal testing further strengthened our formulation which showed that models scored much higher when provided evaluated context-driven setups suggesting they were exploiting topic coher- ence rather than mental state reasoning. We follow a three-way evaluation framework – 1) Retrospective Inference: studies LLM performance on the standard mental state in- ference task; 2) Prospective Diagnostic Forecasting: serves as a diag- nostic probe to measure if LLMs can utilize mental states to identify state-consistent dialogue trajectories without reliance on conversa- tion history; and, 3) Semantic Comparison: compares how similar open-ended LLM generated inferences are with human-written ref- erences. Figure 1 illustrates our framework for benchmark curation. We extract contexts from existing conversation datasets based on arXiv:2604.20443v1 [cs.CL] 22 Apr 2026 Yadav et al. Naturalistic High-Stakes Dialogue Corpus Context Selection Mental State Inference QA Task Generation Task Quality Annotation & Selection DialToM Dataset LLM Evaluation Strong LLM Target LLMs MCQ Eval. Metric: Correctness Rate Semantic Comparison Metrics: BLEU, ROUGE-L, and BERTScore (Context, QA Tasks & Answers) Belief Desire Intention Emotion Knowledge Trust 1. Retrospective MCQ (Literal ToM) 2. Prospective MCQ (Functional ToM) Figure 1: The DialToM Benchmarking Pipeline. The workflow illustrates the transition from Literal ToM (Retrospective Inference) to Functional ToM (Prospective Diagnostic Forecasting) through automated mental state inference, distractor generation, and subsequent human verification. general linguistic statistics, and extract mental state inferences over six mental state attributes. These inferences are then converted to QA tasks – 6 Retrospective and 1 Prospective questions per context. We further perform human-verification on a small subset of the QA pairs in DialToM which we then use to evaluate multiple target LLMs consisting of both open vs. proprietary as well as reasoning vs. non-reasoning LLMs. 2 Related Work & Background 2.1 Benchmarking Theory of Mind in LLMs Synthetic/Template-based. Early ToM reasoning benchmarks such as ToMi [20], Hi-ToM [38], and OpenToM [41] adapted the classic Sally-Anne false-belief paradigm [13,37]. They rely on rigid, sys- tematically generated template-based conversational data. They are further limited to a single mental state attribute (usually, belief ), ignoring the broader spectrum of human cognition. We argue that these contexts do not provide an accurate mental picture of an agent in social situations, where ToM plays a crucial role [34]. Social Simulations. To address the static nature of earlier work, recent work pivoted to agent-based social simulation, e.g., FanToM [19], SOTOPIA [42]. ToMBench [7] and ToMATO [29]. Though these environments have increased the dynamism, these datasets still fail to capture the subtler, rapport-building dynamics of nat- uralistic high-stakes interactions, in which Trust plays a crucial role. Functional ToM. A critical gap in the literature remains to address the disconnect between inference (labeling a state) and action (using it). Recent benchmarks such as BigToM [14], and RecToM [21] attempted to introduce the behavioral prediction of ToM. However, these datasets rely on context-driven predictions where the model has full access to the narrative history during forecasting. This information can be exploited by models by bypassing reasoning and rely on lexical matching with the context to predict the next action. They do not rigorously test the functional utility of the mental state itself. 2.2 Distinction of DialToM DialToM is unique to the current benchmarking methodologies because we predict ToM mental states based on natural human- human conversations including a newly proposed trust level be- tween speakers. Additionally, we operationalize state-driven diag- nostic probing – the ability to identify state-consistent dialogue trajectories based strictly on an isolated mental state profile. By deliberately excluding prior contexts, we prevent models from rely- ing on spurious context-to-dialogue correlations. This is distinct from existing works due to its explicitly conversational nature. We do not rely on goal-oriented contexts as environments, but rather follow a more natural setup. In Table 1, we compare our DialToM with existing benchmarks over five dimensions – mental state attributes of focus, type of contexts, state-driven prediction from ToM reasoning to forecast actions, if they have gold human reference to compare against, and source of contexts. Ecological Validity & The Trust Dimension. Most benchmarks either focus on goal-oriented synthetic stories generated by LLMs or do not follow naturalistic contexts; this makes the findings hard to generalize to real-life human interaction scenarios. To fill this gap, DialToM is constructed from naturalistic, high-stakes human dialogues (mental health support and persuasion). Existing benchmarks are limited to studying belief as a primary mental state attribute which cannot fully capture the relational complexity of these domains or rapport-building of human interac- tions. We extend the standard mental state model to include a new component of Trust (T) which is crucial in human interaction. State-driven prediction (ToM→Future conversation). We intro- duce state-driven prediction to rigorously audit the functional util- ity of ToM. It is the ability to forecast actions based strictly on DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories NameAttributesContext typeState-Driven Prediction Human ref. Context source ToMi [20]BGoal-oriented stories✗Template-based Hi-ToM [38]BGoal-oriented stories✗Template-based BigToM [14]BGoal-oriented stories✗Template-based FANToM [19]BSynthetic conversations✗LLM generated OpenToM [41]BGoal-oriented stories✗LLM generated ToMBench [7]BDIEKGoal-oriented stories✗Human-written ToMATO [29]BDIEKSocial simulation✗LLM-LLM DynToM [40]BIESynthetic scenarios✗LLM generated RecToM [21]BDIRecommendation dialogue✗✓Human-Human DialToM (Ours)BDIEK-THigh-stakes multi-turn dialogues✓Human-Human Table 1: Comparison of DialToM with existing Theory of Mind benchmarks across mental state dimensions, context complexity, and evaluation capabilities. DialToM uniquely extends the BDIEK framework with a novel Trust (-T) attribute and introduces state-driven diagnostic probing over naturalistic high-stakes dialogues. an isolated mental state profile, without access to prior contexts, e.g. scenario or dialogue history. Unlike context-driven forecasting in BigToM and RecToM, DialToM challenges models to perform diagnostic forecasting, forcing them to demonstrate functional ToM by mapping isolated mental profiles to their logical dialogue tra- jectories. This isolation prevents models from relying on spurious context-to-dialogue correlations and instead forces them to demon- strate their robust ToM ability to identify the most state-consistent dialogue trajectories. Semantic Comparison with Human Baselines. DialToM is the first benchmark to provide a gold-standard reference set of human- written mental state rationales. These are important, in such cases where we are comparing the cognitive abilities of LLMs with that of humans. This allows us to semantically compare LLM-generated and human reasonings. 3 Methodology 3.1 Data Curation & Preprocessing To ensure ecological validity, we curate DialToM for realistic con- versations that are arguably more complex, in contrast to existing benchmarks that primarily focus on goal-oriented and synthetic conversations. We focus on three domains of conversational data: Motivational Interviewing (MI) sessions, Emotional Support con- versations, and Persuasive conversations – AnnoMI [39], ESConv [22], PersuasionForGood [35] datasets respectively. In our conver- sational dialogues between two speakers, we assume (for simplicity) there are two interactional roles that remain constant – the initiator (counselor, supporter, persuader) and the recipient (client, seeker, persuadee). AnnoMI (MI). Motivational Interviewing is an effective therapy strategy in mental health research, that elicits positive change to- wards client motivation. AnnoMI [39] is a dataset that consists of 133 counselor-client conversation sessions, with every session assigned a behavior topic, every turn of a session associated with counselor/client role and a behavior code. E.g., a counselor’s mes- sage in a turn could be an open-question or a simple reflection. for the client. We use all 133 sessions including both high and low quality ones. ESConv (ESC). Emotional Support is a crucial ability used in mental health dialogue and social interactions to reduce a user’s emotional distress. The ESConv [22] dataset spans 1300 such ses- sions, each categorized for the emotion, problem, experience and situation of the help seeker. PersuasionForGood (PFG). Persuasive dialogue involves changing one’s opinion through conversational and messaging strategies. PersuasionForGood [35] is one such dataset capturing different persuasive strategies used by human users when asked to persuade other human users over conversations. The 1017 PFG conversations, along with their participant-specific information are utilized for post-processing. StatisticMI ESC PFG Overall Context Count865321518635943 Word Count per context188.2205.02210.8- Turn Count per context5.34.485.4- Word Count/Turn35.545.539.1- Retrospective Infer. Qns.519019,29011,17835,658 Prospective Forecast. Qns.865321518635943 Human-written Gold Refs.300300300900 Table 2: General Statistics of our annotated data. For each of these new contexts, we formulate 6 questions (for each ToM component) for the Retrospective task and 1 question per Prospective task, making the final count of all data points≈ 36000 for Retrospective, and≈ 6000 for Prospective. Context Selection. These datasets have high variability in session lengths and full-length sessions can introduce excessive noise and overhead. To resolve this, we determine specific context windows that cover the right amount of dialogue data for ToM reasoning and planning, and for human annotation. We find the average word and turn counts have high variance across the datasets – MI has the highest average word count of 1233 whereas PFG has the lowest of 273. To normalize context density, we segment these sessions into smaller context windows that are equally long in terms of word count with each other. Yadav et al. Table 2 compares some general statistics of our finally processed data with their original versions. Our final collection is almost ≈ 2.5× as big as the original datasets. Methodology. We distribute the original sessions from each dataset into contexts of size ranging from 150 to 200 words (a statistic we empirically found to balance token efficiency with sufficient context from Table 2). For context windows smaller than this threshold (or, <150), we concatenate them with the previous context, belonging to the same session, to maximize data utilization while keeping our context lengths relatively optimal for inference. 3.2 The Mental State Model Theory of Mind (ToM) is inherently hierarchical and of multi-order. Since our conversations are primarily focused on developing an understanding of the mental state of the recipient, we approxi- mate the mental state model of the recipient from a third-person’s perspective through role-playing [36]. Trust as a component in ToM. Previous benchmarks evaluating the ToM reasoning ability in LLMs have mostly focused on the five well-known components – Beliefs (B), Desires (D), Intentions (I), Emotions (E), and Knowledge (K). These components alone cannot capture the complexity of conversations, and to fill this gap we propose a new mental state component of Trust (T) to capture the trust level between the speakers that is equivalent to the rapport strength. Trust is an important part of a multi-turn dialogue that is built slowly between the participants; it is essential for collaborations and interdependence. There exists three factors largely used to define levels of trust between the trustor and trustee – ability, benevolence and integrity [23]. In our framework, we refer to Trust as a singular concept that combines all three factors. Our proposal is grounded in previous studies of trust between humans and its importance in conversational dialogue [5]. Trust is considered to have an impact on almost all kinds of social interactions [8,17], making it an essential component to model one’s state of mind. 3.3 The Evaluation Tasks When studying the reasoning capabilities of LLMs, it is also im- portant to study how this reasoning can be used to perform future actions, or plan for future contexts [18]. In conversational contexts, such as ours, we propose studying this planning or decision-making behavior of LLMs in the form of forward planning, or “social fore- casting". For each context, we construct two types of Question-Answering (QA) tasks 1 and one Generative task that form the DialToM bench- mark. The QA tasks have one correct answer and three distractors. Distractors are options in MCQ-like setups that are incorrect and are made to look deliberately plausible and distract the agent. These are formally defined as follows: (1) Retrospective Inference: An agent is provided with the conversation history as context, and is asked to infer the mental profile (each of the BDIEKT states) of the recipient speaker. This task is framed as a QA task. 1 All options are generated by GPT-4o and further verified through human verification, unless specified. S=B, D, I, E, K, T T 푖 =M(푠|H 푖 ) ∀푠 ∈S; 푖 ∈ [1,|H|] where,Sis the set of mental state attributes (Beliefs, De- sires, Intentions, Emotions, Knowledge, Trust),T 푖 is the ToM inferences of an agent generated by modelMfor푖 th conversation historyH 푖 , for all mental states푠in the setS. (2)Prospective Diagnostic Forecasting: An agent is pro- vided with the complete mental profile (BDIEKT). As a diagnostic probe, the agent must identify the most state- consistent dialogue trajectory. This task excludes direct access to conversation history (H 푖 ) to ensure the decision is driven solely by the functional logic of the mental state. Formally defined below, this task is constructed as an MCQ task, with individual questions for each mental state. The correct answer here is not generated but is rather extracted from the original sessions, and is actually what follows the given context. S=B, D, I, E, K, T ˆ H 푖+1 =M(· | S) ∀푖 ∈ [1,|H|] where,Sis the set of mental state attributes (Beliefs, De- sires, Intentions, Emotions, Knowledge, Trust), ˆ H 푖+1 is fu- ture dialogue trajectory identified byMgiven the set of mental statesS, for 푖 th conversation historyH 푖 . (3)Semantic Comparison with Human Baselines: This task requires the LLM to generate a free-text open-ended mental state description given an input context. This LLM- generated mental state description is then compared with the gold standard human written references. We would like to highlight that, although we understand there can be multiple mental states at a given point of conversation, we assume our setup overcomes this issue since we have three different human workers annotate each sample giving us three diverse gold references to compare against. 3.4 Human-in-the-Loop Verification We adopt a Human-in-the-Loop pipeline where GPT-4o acts as the data generator for mental state inferences and semantic distractors, and humans serve as annotators. We employ 22 Professional Veri- fiers, who are proficient in English and have a good understanding of our task, from Baidu Data Crowdsourcing Platform to audit all generated MCQ pairs for Mental States (Literal ToM) and future dialogue trajectories (Functional ToM). We apply both manual and automated quality-control methods to ensure data integrity 2 . Addi- tionally, we employ 9 Lay Participants from Amazon Mechanical Turk (AMT), all holding Master’s qualifications and having passed a rigorous qualification test, to author open-ended mental state inferences that serve as a gold reference set. Our annotation framework is structured into four distinct tasks: (1)Verification of Mental State inferences: Annotators ver- ify if the ground-truth mental state is uniquely correct and if the semantic distractors serve as effective, contextually 2 More details in Appendix B. DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories plausible alternatives that are nonetheless logically incor- rect. (2) Verification of Dialogue Trajectories: Annotators en- sure the correct future trajectory logically follows the pro- vided mental state and conversation context, while confirm- ing distractors are plausible but incorrect continuations. (3) Human benchmarking: To quantify the difficulty of the Retrospective Inference, three Lay Participants each com- pleted Task 1 for 900 contexts (300 per conversation do- main). This data forms a baseline reflecting human per- formance on the benchmark. Notably, the human baseline benchmarking was restricted to the Retrospective Inference tasks (Literal ToM) only. This decision was informed by the inherent difficulty of the Prospective Diagnostic Forecast- ing tasks (Functional ToM), which require expert-level ToM reasoning and logical consistency that often exceeds the performance of average crowd workers. (4) Gold Mental States Construction: Three annotators each provided subjective, free-form mental state inferences for 300 contexts (100 per conversation domain). These gold references are used to evaluate the semantic alignment of AI-generated mental states. 3.5 Label Aggregation Due to the unique nature of our annotations, we use the annota- tions from “data verification" to compile a subset of human verified MCQ options of DialToM. We aggregate all annotations for each task (Retrospective and Prospective respectively) using the Dawid- Skene algorithm (DS) [10]. To measure inter-annotator agreement, we report Gwet’s AC1 [16] as it provides superior stability and resistance to the “kappa paradox" compared to standard Cohen’s Kappa (Detailed Discussion in Appendix C). Methodology. DS is used to estimate annotator reliability by pa- rameterizing their expertise with confusion matrices. We assume the case where all annotators find our options “valid" and use them as the true task label for fitting DS models. Since these annotations were performed in batches, we fit separate DS models for each batch, and aggregate the final retention statistics. Results. We retain samples with high inter-rater agreement (AC1). The average AC1 was high for the Retrospective task (0.88-0.92) but moderately lower for the Prospective task (0.45-0.68). Finally, we find that DS retains majority of our original verification dataset in the Retrospective task (≈83.25%), but the retention rate in Prospec- tive task is much lower (≈54.3%). Manual inspection reveals that the lower retention is a direct consequence of the diagnostic probing design. The semantic distractors were engineered to be contextually similar to the ground-truth trajectory on a surface level, intention- ally creating a high-difficulty threshold that challenges Professional Verifiers. 4 Experiments 4.1 Experimental Setup Models. We assess the ToM reasoning and action prediction ca- pabilities of 13 LLMs representative of varying architectures (MoE vs. Standard), parameter sizes (ranging from 12B to 1T total param- eters), capabilities (reasoning vs. non-reasoning) as well as propri- etaries including (a) closed-source LLMs: Gemini 3 Pro [11], Gemini 2.5 Pro and Gemini 2.5 Flash [9], GPT-4.1 [26], GPT-5 [27], and (b) open-source LLMs: Mistral Nemo and 3.2 24B [24,25], Moonshot Kimi K2 [32], Qwen 3 32B and 235B A22B [33], Llama 3.3 70B, 4 Maverick and 3.1 8B [2,15], Deepseek V3 [12], and GPT-OSS 120B [1]. We access most models by prompting through the OpenRouter API interface with default settings 3 , and the proprietary models through the official Gemini / OpenAI APIs. Prompting Strategy. Each LLM is prompted without reasoning and in a zero-shot format for consistency. These prompts are shown in Appendix D. For illustrations of these prompt instances, we provide concrete examples in Appendix E. 4.2 Evaluation Metrics For evaluating the performances on the two MCQ tasks, we adopt the Correctness Rate (CR) metric which is defined by the percent- age of correctly answered questions. 퐶푅= Í 푁 푖=1 1[표 푖 =퐺 푖 ] 푁 ∀표 푖 ,퐺 푖 ∈ O where푁is the total number of questions,표 푖 and퐺 푖 are the LLM answer and the ground truth for question푖respectively, andOis the set of answer options (O=퐴,퐵,퐶,퐷). Instead of using LLMs as a judge for evaluating the performance of LLM inferences with the gold references, we rely on traditional similarity metrics. We adopt both linguistic metrics (BLEU and ROUGE-L) to measure the linguistic similarity between two texts, and semantic similarities (BERTScore) which compare between the generated text of the LLM inference and the gold reference by measuring the cosine similarity between embeddings of two pieces of text. We report the average similarity scores across all references to handle multiple gold references. 5 Empirical Analysis Our analysis shown in Tables 3 to 5 reveals a systematic asymmetry between predicting a latent mental state and leveraging the mental state to forecast a consistent dialogue trajectory. 5.1 Performance Asymmetry In Table 3 we show that most LLMs perform exceptionally well (≥ 80%) on the Retrospective task, as expected. Moreover, we observe minimal cross-domain variance, indicating that LLMs are consistent when performing ToM inference irrespective of the conversation domains. In contrast to the Retrospective task, most LLMs fail the Prospec- tive Diagnostic Forecasting, with several models scoring near or below the random selection threshold (≤25%). This systematic failure indicates that while models can label states (Literal ToM), they struggle to maintain state-consistency when predicting dia- logue trajectories without contextual cues. A notable exception is Gemini 3 Pro, which demonstrates robust Functional ToM by 3 https://openrouter.ai Yadav et al. Model RetrospectiveProspective MIESCPFGMIESCPFG Human32.3324.3324--- Gemini 3 P97.7199.1598.7883.0980.4686.96 GPT-593.4695.2095.2814.7116.2616.85 GPT-4.193.1494.0794.405.8812.8110.87 GPT-OSS 120B75.4988.9885.257.3515.766.52 Gemini 2.5 P92.4897.1894.9938.2429.0619.57 Gemini 2.5 F94.7795.4894.4026.4725.1217.93 Kimi K292.1695.4893.2222.0625.6213.04 DeepSeek v396.7397.7498.8210.2911.8210.33 Mistral 24B61.1168.6474.345.884.937.6 Mistral Nemo42.1652.2656.646.6214.295.44 Qwen 235B95.1097.1897.9417.6529.0617.93 Llama 70B77.4584.7587.9118.4219.707.07 Llama 487.5887.5779.289.5611.825.6 Llama 8B30.3927.1234.516.6216.7510.33 Table 3: Benchmarking results across 13 LLMs showing Correctness Rate (CR) for Retrospective (Literal ToM) and Prospective (Functional ToM) tasks. maintaining high accuracy (≈83%) across all three conversation domains, peaking at≈ 88% on MI. Human Baseline Analysis. The human baseline analysis reveals that Lay Participants are significantly outperformed by even the lowest-performing LLMs on the Retrospective task. This divergence does not necessarily signify superhuman LLM reasoning. Rather, it indicates the gap between lay perception and the nuanced linguistic patterns present in specialized high-stakes dialogues like counseling and persuasion. Our findings are consistent with recent studies [31] suggesting that while LLMs can exceed human scores on specific ToM batteries, their performance is often more consistent than that of crowd workers who are susceptible to fatigue and information overload during cognitively demanding tasks. Specifically, each Retrospective task requires at least two minutes of focused deliberation to detect subtle linguistic cues, a threshold where human consistency often falters. As a result, the layman baseline serves as a lower bound, representing the explicit diffi- culty for non-experts to track mental states in complex, naturalistic dialogues. 5.2 Impact of Model Scale and Architecture We find clear patterns of LLM performances scaling with model capacity 4 , both within model families (such as Llama and Mistral) and across all models. In general, we found all frontier / high- capacity models, to be the best overall performers on our tasks – Gemini family, Qwen 235B, Kimi K2. A similar pattern was observed in specific model families – Mistral 24B outperforming Mistral Nemo; GPT-5 and Gemini 2.5 Pro performing better than their predecessor / smaller counterparts (GPT-4.1 and Gemini 2.5 Flash respectively). An exception was noticed in the Llama family where surprisingly Llama 8B outperforms both Llama 70B and Llama 4. 4 For models with MoE architectures where number of active parameters are different from total, we consider the total parameters instead. 5.3 Task-Domain Generalization Domain RetrospectiveProspective MIESCPFGMIESCPFG MI1.00.978 ∗ 0.965 ∗ 1.00.966 ∗ 0.941 ∗ ESC0.978 ∗ 1.00.987 ∗ 0.966 ∗ 1.00.939 ∗ PFG0.965 ∗ 0.987 ∗ 1.00.941 ∗ 0.939 ∗ 1.0 Table 4: Pearson correlations across different conversation domains. ( ∗ : p<0.001, † : p<0.01, ‡ : p<0.05) We analyze the cross-domain performance correlations of LLMs to comment on their generalizability in our setup. In Table 4, we show strong average Pearson correlations between all pairs of con- versation domains, for both tasks (Retrospective = 0.978, Prospective = 0.949). We also check for correlations in between tasks where we find insignificant and weak Pearson correlations on average (푝= 0.35), but significant Spearman correlations (휌= 0.67). This implies cross-task performances on DialToM are not lin- early correlated to each other, whereas cross-domain performances within each task are strongly correlated – high performance in one domain should generally imply high performance in another, but the same may not be strongly true when measuring cross-task performances. 5.4 Attribute-Level Analysis Model Retrospective BDIEKT Gemini 3 P97.6599.3698.0410098.0498.34 GPT-595.4097.4796.1098.8697.4284.15 GPT-4.197.1397.4796.1099.4396.7778.14 GPT-OSS 120B88.5189.8794.1691.4391.6183.61 Gemini 2.5 P95.9894.9494.8199.4396.1389.07 Gemini 2.5 F95.9897.4798.7096.5795.4886.34 Kimi K297.1390.5195.4598.8694.8485.79 DeepSeek v398.2898.7395.4599.4398.0696.72 Mistral 24B66.6768.3574.6854.8678.0668.85 Mistral Nemo50.5752.5363.6437.1454.1948.09 Qwen 235B97.7098.7399.3599.4398.7187.98 Llama 70B82.7691.7787.6684.5781.2974.86 Llama 479.1790.2689.2686.5583.8981.50 Llama 8B31.6130.3835.0629.1432.2626.23 Table 5: Attribute-level Correctness Rate (CR) for the Retro- spective task across six mental state dimensions (BDIEKT). In Table 5, we present the LLM performances for all mental state attributes. We find that most LLMs, except Gemini 3 Pro, generally perform worse at predicting the Trust state, than the other states. Trust is weakly understood in LLMs as shown by the weak performance of the top performing LLMs like Qwen 235B and Gemini 2.5 P. In contrast, a superior performance on the belief and intentions here suggests a deeper understanding of beliefs in LLMs consistent with previous research [43]. Our findings further DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories strengthen our claim that trust is an important but understudied aspect of ToM in LLMs. 5.5 Semantic Comparison with Gold Reference In Table 6, we compare the results of LLM generated inferences with that of human written gold references. The first row, GPT-4o, shows the similarities between the inferences used in the MCQ tasks and human annotation. We find LLMs are only moderately similar to human written gold inferences by semantic similarity (signified by moderate BERTScores) and very low overlap measured by linguistic similarity (signified by low BLEU, ROUGE-L scores). These overall low scores suggest the challenging aspect of generating free-form mental state inferences and the existing lexical diversity in generations while preserving semantic. Nevertheless, the proprietary LLMs (GPT and Gemini families) perform better relative to other models on writing tasks. This overall weak semantic similarity creates a dissonance with respect to the otherwise strong Retrospective task performance. Model Gold Mental-State Reference BLEUROUGE-LBERTScore GPT-4o9.440.310.44 GPT-58.560.320.41 GPT-4.19.710.330.43 GPT-OSS 120B8.090.320.41 Gemini 2.5 P8.900.320.44 Gemini 2.5 F7.980.320.39 Kimi K27.770.290.38 DeepSeek v38.750.320.42 Mistral 24B6.980.290.37 Mistral Nemo6.210.290.36 Qwen 235B8.050.300.40 Llama 70B6.670.280.34 Llama 47.320.290.35 Llama 8B5.300.250.31 Table 6: Linguistic alignment between AI and human-written inferences. Low linguistic overlap (BLEU/ROUGE-L) along- side moderate BERTScore values indicate high lexical diver- sity despite semantic preservation. 6 Ablation Studies Throughout our experiments we found that most LLMs (except, Gemini 3 Pro) were much worse when predicting future dialogue trajectories than directly inferring mental states. To verify the rea- soning integrity of the top-performing models, we perform two ablation and diagnostic analyses to examine the mechanisms be- hind the performance divergence: 1) testing for dependency on spurious correlations via distractor manipulation; and, 2) verifying state-driven dependency via counterfactual intervention. 6.1 Distractor Complexity and Spurious Correlations A possible explanation for the poor performance of most LLMs on the Prospective task is the distractors and correct answer being very Model Prospective-Easy MIESCPFG GPT-594.85 +80.15 91.63 +75.37 61.96 +50.54 GPT-4.191.18 +85.29 90.64 +77.83 64.13 +58.70 GPT-OSS 120B86.76 +77.21 78.33 +61.58 55.43 +48.91 Gemini 2.5 P94.85 +56.62 95.07 +66.01 67.93 +48.37 Gemini 2.5 F92.65 +66.18 91.13 +66.01 61.41 +51.63 Kimi K293.38 +71.32 89.66 +64.04 65.76 +62.50 DeepSeek v393.38 +83.09 87.68 +75.86 61.96 +60.33 Mistral 24B86.03 +80.15 82.76 +77.83 65.76 +63.59 Mistral Nemo77.21 +70.59 66.01 +51.72 56.52 +55.43 Qwen 235B88.24 +70.59 89.66 +60.59 57.61 +51.09 Llama 70B86.76 +68.34 88.18 +68.47 58.15 +52.72 Llama 489.71 +80.15 90.64 +78.82 69.02 +64.67 Llama 8B75.0 +68.38 66.5 +49.75 50.0 +44.57 Table 7: Performance on the "Easy Set" of the Prospective Diagnostic Forecasting task. Blue subscripts indicate the per- formance surge compared to the original Hard Set. similar to each other, thus making it harder for LLMs with poorer reasoning abilities to discern the correct answer. Through a manual inspection and an analysis of worker annotations from Table 9, we found the correct answers to indeed be harder to differentiate from the distractors. We consider this set of Prospective task MCQs as the hard set where LLMs find it relatively harder to predict the correct answer. To resolve this, we devise an easy set with the same questions but different distractors. In this set, instead of using generated “seman- tic" distractors, we sample conversation contexts from randomly selected contexts, which has been a predominant method of con- structing distractors in previous QA benchmarks, such as ToMATO [29]. This set is easier because it has “discriminative" distractors that are very contextually different from the correct answer, and we would expect all LLMs to perform exceptionally on this set. The hard set forces the model to distinguish between plausible dialogue trajectories semantically without any shortcuts, whereas the easy set allows the model to rely on shortcuts by exploiting correlations such as topic coherence. Methodology. Instead of using semantic distractors for each ques- tion as in the hard set, we randomly sample three subcontexts and extract the first two turns (four dialogues total: two for each speaker), which are then used as discriminative distractors in the easy set. Since, Gemini 3 Pro is already excellent at this task, we skip its evaluation for this ablation. Results. Table 7 shows that this change improves the perfor- mances of all LLMs by a minimum of 44 and up to 83 points. This shift implies that LLMs perform better when the conversations follow entirely different contexts and are hence easier to discern. The performance surge on the Easy Set confirms that when the diagnostic pressure is lowered, models revert to identifying con- textual outliers. The Hard Set validates that most models lack the Functional ToM required to identify state-consistent trajectories when lexical shortcuts are removed. Yadav et al. 6.2 Robustness and Memorization Since most LLMs fail on the Prospective task, except Gemini 3 Pro, this anomalous behavior raises certain questions about whether Gemini 3 Pro is genuinely robust at reasoning or is it purely due to memorization. It is thus imperative to verify whether this is a fluke or not in controlled settings. To systematically study and answer this, we design a short abla- tion to study whether it is due to genuine performance or a result of memorization. We introduce a counterfactual-based intervention and replace the correct mental states with a set of generated but significantly different counterfactual mental states. This means the correct answer is not exactly the correct answer anymore due to the counterfactual mental states. Methodology. We use GPT-4o to construct counterfactual men- tal states (see Appendix D for prompts) for a given conversation context and mental state for a certain attribute. Instead of using the correct set of mental statesSin the Prospective task setup, we replace them with a set of generated counterfactualsS ′ . Then, we evaluate whether models choose “None of the Above" (NOTA) as an option or not. If they do, it implies robust reasoning processes otherwise memorization. In addition to Gemini 3 Pro, we also test top 4 best performing (Gemini 2.5 P, F, GPT-5, Kimi K2) and bottom 4 worst performing (Mistral Nemo, 24B, Llama 4, 8B) models on the Prospective task. Results. From Table 8 we find that Gemini 3 Pro does indeed per- form extremely well on this task. Its high performance on predicting NOTA for counterfactual mental states implies that it might be due to its superior reasoning abilities. This confirms the hypothesis that Gemini 3 Pro does not rely on memorization and is extremely robust due to its superior reasoning abilities. We also find that most models perform genuine reasoning except outliers such as Llama 8B and Gemini 2.5 Flash which are notably weaker than others. Our findings strengthens the feasibility and reasonable difficulty of the task, and shows Gemini 3 Pro is able to perform genuinely well on inferring ToM reasoning, much better than humans or average LLMs. ModelMI ESC PFG Gemini 3 P80.1594.0995.11 Gemini 2.5 P70.5987.6888.59 Gemini 2.5 F48.5376.3574.46 GPT-561.0386.790.22 Kimi K280.8895.0795.65 Mistral Nemo59.5673.8988.59 Mistral 24B97.0699.0199.46 Llama 467.6578.8255.98 Llama 8B37.537.9351.09 Table 8: Robustness vs. Memorization via Counterfactual Intervention. Correctness Rate (CR) for identifying "None of the Above" (NOTA) when provided with contradictory mental statesS ′ . 7 Conclusion In this work, we propose a novel ToM reasoning benchmark, Dial- ToM, that measures the predictability of mental states including the trust factor between speakers, and uses diagnostic probing to eval- uate if agents can leverage these states to forecast state-consistent future dialogue trajectories. We show that most LLMs suffer from a performance disparity between the two kinds of tasks in DialToM – it is easier to predict the mental state, but harder to decide for future actions based on the mental states. Additionally, through deeper investigations, we found that most LLMs rely on linguistic pattern identification in QAs instead of following a stricter form of reasoning. In contrast, we found Gemini 3 Pro to be quite robust and have superior reasoning capabilities. Further, we believe our research provides deeper insights into how ToM reasoning benchmarks should be structured and attempts to solve the current lingering issues in existing benchmarks. We hope our study serves as a foundation for future benchmarks to incorporate realistic contexts, attributes responsible and crucial for conversations, as well as future action planning. 8 Limitations and Ethical Considerations Our work has several limitations: (1) the Trust attribute is scoped to specific social interactions and may not capture broader socio- cultural definitions; (2) results from high-stakes domains may not generalize to multi-party or casual settings; (3) while audited by an- notators, GPT-4o-generated inferences and distractors may contain subtle model biases; and (4) the lay human baseline is restricted to Literal ToM, as the Functional ToM task proved too structurally complex for non-experts. Ethically, all human studies were approved by the Singapore Management University IRB. We strictly adhered to the licenses of AnnoMI, ESConv, and PersuasionForGood, ensuring all contexts are anonymized. All participants (sourced via Baidu and AMT) were compensated at or above local living wages. While we acknowl- edge the risk of AIs with enhanced ToM for social manipulation, we release DialToM to promote transparent, diagnostic alignment rather than the development of deceptive agents. References [1] Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, et al.2025. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925 (2025). [2]Meta AI. 2025.Llama 4.https://ai.meta.com/blog/llama-4-multimodal- intelligence/ [3]Ian A. Apperly and Stephen A. Butterfill. 2009. Do humans have two systems to track beliefs and belief-like states? Psychological Review 116, 4 (2009), 953–970. doi:10.1037/a0016923 [4]Chris L. Baker, Rebecca Saxe, and Joshua B. Tenenbaum. 2009. Action under- standing as inverse planning. Cognition 113, 3 (2009), 329–349. doi:10.1016/j. cognition.2009.07.005 Reinforcement learning and higher cognition. [5]Erika Blacksher, Charlene Nelson, Emily Van Dyke, Abigail Echo-Hawk, Deborah Bassett, and Dedra Buchwald. 2016. Conversations about Community-Based Participatory Research and Trust: “We Are Explorers Together”. Progress in Community Health Partnerships: Research, Education, and Action 10, 2 (2016), 305–309. doi:10.1353/cpr.2016.0039 [6]Michael Bratman. 1987. Intention, plans, and practical reason. Stanford Univ Center for the Study. [7]Zhuang Chen, Jincenzi Wu, Jinfeng Zhou, Bosi Wen, Guanqun Bi, Gongyao Jiang, Yaru Cao, Mengting Hu, Yunghwei Lai, Zexuan Xiong, and Minlie Huang. 2024. ToMBench: Benchmarking Theory of Mind in Large Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 15959–15983. doi:10.18653/v1/2024.acl-long.847 [8]Leigh Clark, Nadia Pantidi, Orla Cooney, Philip Doyle, Diego Garaialde, Justin Edwards, Brendan Spillane, Emer Gilmartin, Christine Murad, Cosmin Munteanu, Vincent Wade, and Benjamin R. Cowan. 2019. What Makes a Good Conversation? Challenges in Designing Truly Conversational Agents. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (Glasgow, Scotland Uk) (CHI ’19). Association for Computing Machinery, New York, NY, USA, 1–12. doi:10.1145/3290605.3300705 [9]Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and Others. 2025. Gemini 2.5: Pushing the Frontier with Advanced Reason- ing, Multimodality, Long Context, and Next Generation Agentic Capabilities. arXiv:2507.06261 [cs.CL] https://arxiv.org/abs/2507.06261 [10]A. P. Dawid and A. M. Skene. 1979. Maximum Likelihood Estimation of Observer Error-Rates Using the EM Algorithm. Applied Statistics 28, 1 (1979), 20. doi:10. 2307/2346806 [11] Google DeepMind. 2025. Gemini 3. https://blog.google/innovation-and-ai/ technology/ai/google-gemini-ai/ [12]DeepSeek-AI. 2024. DeepSeek-V3 Technical Report. arXiv:2412.19437 [cs.CL] https://arxiv.org/abs/2412.19437 [13] Daniel C. Dennett. 1978. Beliefs about beliefs [P&W, SR&B]. Behavioral and Brain Sciences 1, 4 (1978), 568–570. doi:10.1017/S0140525X00076664 [14]Kanishk Gandhi, Jan-Philipp Fraenken, Tobias Gerstenberg, and Noah Goodman. 2023. Understanding Social Reasoning in Language Models with Language Models. In Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36. Curran Associates, Inc., 13518–13529. https://proceedings.neurips.c/paper_files/paper/2023/file/ 2b9efb085d3829a2aadffab63ba206de-Paper-Datasets_and_Benchmarks.pdf [15] Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schel- ten, Alex Vaughan, and Others. 2024.The Llama 3 Herd of Models. arXiv:2407.21783 [cs.AI] https://arxiv.org/abs/2407.21783 [16]Kilem Li Gwet. 2008. Computing inter-rater reliability and its variance in the presence of high agreement. Brit. J. Math. Statist. Psych. 61, 1 (May 2008), 29–48. doi:10.1348/000711006x126600 [17] P. A. Hancock, Theresa T. Kessler, Alexandra D. Kaplan, Kimberly Stowers, J. Christopher Brill, Deborah R. Billings, Kristin E. Schaefer, and James L. Szalma. 2023. How and why humans trust: A meta-analysis and elaborated model. Frontiers in Psychology 14 (March 2023). doi:10.3389/fpsyg.2023.1081086 [18] Mark K. Ho, Rebecca Saxe, and Fiery Cushman. 2022. Planning with Theory of Mind. Trends in Cognitive Sciences 26, 11 (Nov. 2022), 959–971. doi:10.1016/j.tics. 2022.08.003 [19] Hyunwoo Kim, Melanie Sclar, Xuhui Zhou, Ronan Bras, Gunhee Kim, Yejin Choi, and Maarten Sap. 2023. FANToM: A Benchmark for Stress-testing Machine Theory of Mind in Interactions. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 14397–14413. doi:10.18653/v1/2023.emnlp-main.890 [20]Matthew Le, Y-Lan Boureau, and Maximilian Nickel. 2019. Revisiting the Evalu- ation of Theory of Mind through Question Answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Inter- national Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Ken- taro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan (Eds.). Association for Com- putational Linguistics, Hong Kong, China, 5872–5877. doi:10.18653/v1/D19-1598 [21]Mengfan Li, Xuanhua Shi, and Yang Deng. 2025. RecToM: A Benchmark for Evaluating Machine Theory of Mind in LLM-based Conversational Recommender Systems. arXiv:2511.22275 [cs.AI] https://arxiv.org/abs/2511.22275 [22]Siyang Liu, Chujie Zheng, Orianna Demasi, Sahand Sabour, Yu Li, Zhou Yu, Yong Jiang, and Minlie Huang. 2021. Towards Emotional Support Dialog Systems. In Proceedings of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 3469–3483. [23]Roger C. Mayer, James H. Davis, and F. David Schoorman. 1995. An Integrative Model of Organizational Trust. The Academy of Management Review 20, 3 (July 1995), 709. doi:10.2307/258792 [24] Mistral. 2024. Mistral NeMo. https://mistral.ai/news/mistral-nemo [25]Mistral. 2025. Mistral Small 3.2 24B. https://docs.mistral.ai/models/mistral- small-3-2-25-06 [26] OpenAI. 2025. GPT-4.1. https://openai.com/index/gpt-4-1/ [27] OpenAI. 2025. GPT-5. https://openai.com/index/introducing-gpt-5/ [28] Matthew Riemer, Zahra Ashktorab, Djallel Bouneffouf, Payel Das, Miao Liu, Justin D Weisz, and Murray Campbell. 2024. Position: Theory of Mind Bench- marks are Broken for Large Language Models. arXiv preprint arXiv:2412.19726. [29]Kazutoshi Shinoda, Nobukatsu Hojo, Kyosuke Nishida, Saki Mizuno, Keita Suzuki, Ryo Masumura, Hiroaki Sugiyama, and Kuniko Saito. 2025. ToMATO: Verbalizing the Mental States of Role-Playing LLMs for Benchmarking Theory of Mind. arXiv preprint arXiv:2501.08838 (2025). [30]Michael Shum, Max Kleiman-Weiner, Michael L. Littman, and Joshua B. Tenen- baum. 2019. Theory of Minds: Understanding Behavior in Groups through Inverse Planning. Proceedings of the AAAI Conference on Artificial Intelligence 33, 01 (July 2019), 6163–6170. doi:10.1609/aaai.v33i01.33016163 [31] James WA Strachan, Dalila Albergo, Giulia Borghini, Oriana Pansardi, Eugenio Scaliti, Saurabh Gupta, Krati Saxena, Alessandro Rufo, Stefano Panzeri, Guido Manzi, et al.2024. Testing theory of mind in large language models and humans. Nature Human Behaviour (2024), 1–11. [32] Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, and Others. 2025. Kimi K2: Open Agentic Intelligence. arXiv:2507.20534 [cs.LG] https://arxiv.org/abs/ 2507.20534 [33] Qwen Team. 2025. Qwen3 Technical Report. arXiv:2505.09388 [cs.CL] https: //arxiv.org/abs/2505.09388 [34]Qiaosi Wang, Xuhui Zhou, Maarten Sap, Jodi Forlizzi, and Hong Shen. 2025. Rethinking Theory of Mind Benchmarks for LLMs: Towards A User-Centered Perspective. arXiv:2504.10839 [cs.HC] https://arxiv.org/abs/2504.10839 [35] Xuewei Wang, Weiyan Shi, Richard Kim, Yoojung Oh, Sijia Yang, Jingwen Zhang, and Zhou Yu. 2019. Persuasion for Good: Towards a Personalized Persuasive Dialogue System for Social Good. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Anna Korhonen, David Traum, and Lluís Màrquez (Eds.). Association for Computational Linguistics, Florence, Italy, 5635–5649. doi:10.18653/v1/P19-1566 [36]Alex Wilf, Sihyun Shawn Lee, Paul Pu Liang, and Louis-Philippe Morency. 2023. Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of- Mind Capabilities. arXiv preprint arXiv:2311.10227 (2023). [37]H Wimmer. 1983. Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children’s understanding of deception. Cognition 13, 1 (Jan. 1983), 103–128. doi:10.1016/0010-0277(83)90004-5 [38] Yufan Wu, Yinghui He, Yilin Jia, Rada Mihalcea, Yulong Chen, and Naihao Deng. 2023. Hi-ToM: A Benchmark for Evaluating Higher-Order Theory of Mind Reasoning in Large Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2023, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 10691–10706. doi:10.18653/v1/2023.findings-emnlp.717 [39]Zixiu Wu, Simone Balloccu, Vivek Kumar, Rim Helaoui, Ehud Reiter, Diego Reforgiato Recupero, and Daniele Riboni. 2022. Anno-MI: A Dataset of Expert- Annotated Counselling Dialogues. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 6177–6181. doi:10. 1109/ICASSP43922.2022.9746035 [40] Yang Xiao, Jiashuo Wang, Qiancheng Xu, Changhe Song, Chunpu Xu, Yi Cheng, Wenjie Li, and Pengfei Liu. 2025. Towards Dynamic Theory of Mind: Evaluating LLM Adaptation to Temporal Evolution of Human States. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Moham- mad Taher Pilehvar (Eds.). Association for Computational Linguistics, Vienna, Austria, 24036–24057. doi:10.18653/v1/2025.acl-long.1171 [41]Hainiu Xu, Runcong Zhao, Lixing Zhu, Jinhua Du, and Yulan He. 2024. Open- ToM: A Comprehensive Benchmark for Evaluating Theory-of-Mind Reasoning Capabilities of Large Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 8593–8623. doi:10.18653/v1/2024.acl-long.466 [42] Xuhui Zhou, Hao Zhu, Leena Mathur, Ruohong Zhang, Zhengyang Qi, Haofei Yu, Louis-Philippe Morency, Yonatan Bisk, Daniel Fried, Graham Neubig, and Maarten Sap. 2024. SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents. ICLR. https://openreview.net/forum?id=mM7VurbA4r [43]Wentao Zhu, Zhining Zhang, and Yizhou Wang. 2024. Language Models Repre- sent Beliefs of Self and Others. In Forty-first International Conference on Machine Learning. A Dataset and Code Availability To ensure reproducibility, our dataset and code is publicly available at the following URL: https://github.com/Stealth-py/DialToM. B Detailed Data Annotation Process The annotation study was approved by the Institutional Review Board of Singapore Management University (IRB Approval Number: IRB-25-133-A105(725)). Yadav et al. Task MIESCPFG O1O2O3O4O1O2O3O4O1O2O3O4 Retrospective0.880.880.90.910.940.950.950.960.870.960.970.97 Prospective0.370.570.620.670.260.840.840.860.240.820.830.92 Table 9: Consensus-based verification of MCQ options. Values represent the agreement scores for Professional Verifiers distinguishing the ground-truth (O1) from semantic distractors (O2–O4). B.1 Crowdsourcing Methodology Platform Selection. We chose separate platforms for our tasks by taking into account worker-task compatibility. To ensure that Lay Participants who author free-form mental state inferences were not exposed to AI-generated inferences before, we recruited Lay Participants for authoring mental state inferences from a different worker pool (AMT) than the verifiers or Lay Participants forming the baseline (Baidu). Worker Selection. To gather high-quality annotations, we filter workers through a selective worker qualification process. In this process, across both platforms AMT and Baidu, we ask the workers to participate in a replication of the Retrospective task of DialToM. Workers were provided with a conversation excerpt, and are needed to answer an MCQ question for selecting the appropriate mental state of the recipient. We selectively chose workers that score at least 80% on this test. B.2 Participant Roles We categorized crowd workers into two distinct groups based on the nature of the task: (1) Professional Verifiers (Baidu): These individuals performed the Verification of AI-generated Mental State Inferences and Dialogue Trajectories. Their high-quality output was aggregated using the Dawid-Skene algorithm to ensure the final dataset’s integrity. (2)Lay Participants (Baidu/AMT): These individuals provided the Human Baseline by completing the MCQ tasks with- out prior exposure to the correct labels. Their performance serves to quantify the intrinsic difficulty and domain-specific nuance of DialToM. B.3 Quality Control Methods Automated. We employ the Dawid-Skene algorithm to model annotator skills and verify whether their future annotations remain consistent with previous annotations or not. If the quality diverges significantly, we consider that annotator to be of low-quality and request for a swap of annotations from Baidu. Manual. We also do a high-level quality check manually, by going through all samples and verifying if the annotations are as per what we, as researchers with great deal of experience in this field, would expect. B.4 Gold Mental-State Inferences Construction Since our gold mental-state inference task (Annotation Task 4) is much more complex than other tasks, it thus requires close to native proficiency of speaking, understanding and writing Eng- lish. Hence, we only chose annotators that were not only able to reach an optimal score on our qualification test, but those who also held a Master’s qualification and were from predominantly English speaking countries in the West, such as United States, and United Kingdom. To account for the high rate of tasks on AMT, we performed the rest of our tasks that have much more samples than Task 4 through the Baidu Data Crowdsourcing Platform with whom we reached a contract for mutual benefit. C Discussions of Verification Results Metric RetrospectiveProspective MI ESC PFG MI ESC PFG # Verification400400400400300300 # Retained306354339136203184 % Retention76.588.584.753467.6761.33 Gwet’s AC10.880.920.920.450.680.62 Table 10: Data and Annotation statistics for label aggregation of verified annotations. # Verification and # Retained refer to number of samples in the verification set (that was passed to the annotators) and the number of samples that we retain after aggregation respectively, and % Retention is the reten- tion rate of DS. In Table 9, we illustrate the average verification statistics av- eraged over all annotators for both tasks. Following a systematic pattern, all Professional Verifiers found the “ground truth" (O1) much harder (lower score here implies confusion and increased difficulty of the task) to differentiate from the other options. Table 10 shows the dataset statistics post human verification and Dawid-Skene label aggregation. We were able to retain≈80% of the data for the Retrospective and≈ 60% for the Prospective task. D Prompts Used D.1 Data Generation Counterfactuals You are an expert in Theory of Mind reasoning and generating counterfactual statements. DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories ModelParameter SizeDeveloperKnowledge Cut-off Mistral Nemo12BMistral AIApril 2024 Mistral 3.2 24B24BMistral AIOctober 2023 Moonshot Kimi K21T, 32B activeMoonshot AIOctober 2024 Qwen3 235B235B, 22B activeAlibabaJune 2025 Llama 4 Maverick400B, 17B activeMeta AIAugust 2024 Llama 3.3 70B70BMeta AIDecember 2023 Llama 3.1 8B8BMeta AIDecember 2023 DeepSeek v3671B, 37B activeDeepSeekMarch 2025 Gemini 3 Pro-Google- Gemini 2.5 Flash-Google- Gemini 2.5 Pro-Google- GPT-OSS 120B120BOpenAIJune 2024 GPT-4.1-OpenAI- GPT-5-OpenAI- Table 11: Overview of benchmarked LLMs. Models are categorized by architecture and access type. Technical specifications and knowledge cut-offs are provided where available. StatisticMI ESC PFG Overall Original Session Count133130010172450 Word Count per session1232.97520.61272.78- Turn Count per session35.4711.269.39 Word Count per turn34.7646.2229.07- Final Context Count865321518635943 Word Count per context188.2205.02210.8- Turn Count per context5.34.485.4- Word Count/Turn35.545.539.1- Retrospective Infer. Qns.519019,29011,17835,658 Prospective Forecast. Qns.865321518635943 Mental State Inference100100100300 Table 12: General Statistics of our collected data. For refer- ence, these are all contexts that we use for our ToM inferences. For each of these new contexts, we formulate 6 questions (for each ToM component), making the final count of all data points≈ 36000. Task: You will be provided with an excerpt from a conversation along with the current state state of recipient role. You need to generate a counterfactual mental state that is obviously not true at all given the current conversation. It should be CONTRARY to the provided state. Conversation context: conversation Mental State: current mental state Instruction: Output only the counterfactual . Do not add explanations or other verbosity. Your output should strictly follow the structure "mental state writing format". NO FORMATTING NEEDS TO BE DONE. Counterfactual: D.2 Benchmarking Retrospective You are an expert in Theory of Mind reasoning. Task: You will be provided with conversation between two agents initiator role and recipient role engaging in a session type session on the topic of topic type. Your goal is to correctly infer recipient role’s mental state, based on the above conversation. You will be provided with a set of options, and you need to choose the most appropriate one that reflects the mental state. The correct option must be consistent with the provided conversa- tion context. Conversation Context: conversation Mental State Options: options Instruction: Output only the letter of the correct option (e.g., "A", "B", "C", or "D"). Do not add explanations or other verbosity. Your output should be strictly one of: A, B, C, D. NO FORMATTING NEEDS TO BE DONE. Yadav et al. Answer: Prospective You are an expert in Theory of Mind reasoning. Task: You will be provided with internal Mental State profile (Belief, Desire, Intention, Emotion, Knowledge, Trust) of client during the conversation. Your goal is to identify which of the candidate conversation segments is the most plausible continuation of this conversation. The correct option must be consistent with the provided Mental States of the recipient role. Mental state of recipient role mental states Candidate Conversation Segments options Instruction: Output only the letter of the correct option (e.g., "A", "B", "C", or "D"). Do not add explanations or other verbosity. Your output should be strictly one of: A, B, C, D. NO FORMATTING NEEDS TO BE DONE. Answer: Semantic Comparison You are an expert in Theory of Mind reasoning. Task: You will be provided with a context of a conversation between initiator role and recipient role. Your goal is to accurately infer the mental state of the recipient role in one line. The correct option must be consistent with the provided conversa- tional context. Conversation Context: conversation Instruction: Output only a single line inference. Your response should always start with "mental state writing format". Do not add explanations or other verbosity. STRICTLY FOLLOW THIS FORMAT AND OUTPUT ONLY ONE LINE. Answer: E Task Examples To illustrate the structure and complexity of the DialToM bench- mark, we provide representative examples of the Retrospective Inference and Prospective Diagnostic Forecasting tasks in this sec- tion. Each example is labeled with a unique Task ID, corresponding to its specific instance in the DialToM dataset. E.1 Retrospective Inference We include three examples of Retrospective Inference tasks (Literal ToM), one for each conversation domain: Counseling, Emotional Support, and Persuasion. These examples demonstrate the model’s ability to map multi-turn conversation contexts to specific mental state attributes Retrospective Task 1: Counseling (ID: 17) You are an expert in Theory of Mind reasoning. Task: You will be provided with conversation between two agents supporter and seeker engaging in a Counseling Session on the topic of on reducing alcohol consumption. Your goal is to correctly infer seeker’s Emotional state, based on the conversation below. You will be provided with a set of options, and you need to choose the most appropriate one that reflects the Emotion state. The correct option must be consistent with the provided conversa- tion context. Conversation Context: counselor: And with friends, you feel like it’s easier to-to hang out with alcohol, or maybe you wouldn’t have as much fun, or your friends wouldn’t like you as much or? client: I guess, I mean, since it’s just there. counselor: Okay. client: So- counselor: Got it. client: -it ties in with the whole belonging thing. counselor: Okay. Well, when you think about your drinking and, sort of, looking forward in 5 years or 10 years, how do you think that drinking might fit in with your life? client: Um, it probably will have a very small role, if any- counselor: Mm-hmm. client: -in my life, considering I have big plans for my future. So. counselor: Okay. Can you expand on that a little bit more? What do you mean? client: I mean, I don’t think I would be, you know, drinking as much or going to parties as much if I’m studying for school, or in college, or trying to get in law school and things like that. counselor: Okay, so, the-the alcohol is something that might make it harder to reach your goals? client: Yeah, definitely. Mental State Options: A: I feel indifferent about my drinking habits right now. B: I feel excited about drinking because it’s a big part of my social life. C: I feel anxious about the idea of not drinking with friends. D: I feel determined to reduce my alcohol consumption because I have clear long-term goals. DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories Instruction: Output only the letter of the correct option (e.g., "A", "B", "C", or "D"). Do not add explanations or other verbosity. Your output should be strictly one of: A, B, C, D. NO FORMATTING NEEDS TO BE DONE. ONLY OUTPUT THE OPTION AND NOTHING ELSE. YOUR OUTPUT SHOULD STRICTLY BE ONE OF A, B, C, or D. Answer: Retrospective Task 2: Emotional Support (ID: 22) You are an expert in Theory of Mind reasoning. Task: You will be provided with conversation between two agents sup- porter and seeker engaging in a Emotional Support Conversation session on the topic of on anger over breakup with partner. Your goal is to correctly infer seeker’s Knowledge state, based on the conversation below. You will be provided with a set of options, and you need to choose the most appropriate one that reflects the Knowledge state. The correct option must be consistent with the provided conversa- tion context. Conversation Context: seeker: It has been happening for at least 6 months now and it usually happens every weekend supporter: I can sense your frustration. It must be overwhelming at times and make you dread the weekends. It sounds like it has become a habit for your ex. It might help if you write a letter to them explaining that you are ready to move on and are asking them to do the same. Explain that you are giving them the opportunity to stop harassing you in the following ways [list some of the details you have documented]. Inform them that if they do not stop, you are considering taking it further regarding harassing communications and that this letter serves as your intention to no longer tolerate their actions. Send it via certified mail and keep a copy. This will give them no way to immediately respond and would be the first step of official documentation should they decide to continue. Do you think that might help? seeker: I think that is an excellent idea! That is a good way to end this once and for all. supporter: I sincerely hope it is a step toward you regaining your control over how you are able to interact with people. I wish you the best in dealing with this and hope you begin to feel better about it soon, especially entering into a new year! seeker: Thank you very much! Mental State Options: A: I know that responding impulsively in anger can resolve the situation more gently. B: I know that avoiding any communication is the best way to make the situation go away completely. C: I know that sending a well-documented letter can serve as both communication and legal evidence to stop harassment. D: I know that ignoring the problem will make it disappear without any further action required. Instruction: Output only the letter of the correct option (e.g., "A", "B", "C", or "D"). Do not add explanations or other verbosity. Your output should be strictly one of: A, B, C, D. NO FORMATTING NEEDS TO BE DONE. ONLY OUTPUT THE OPTION AND NOTHING ELSE. YOUR OUTPUT SHOULD STRICTLY BE ONE OF A, B, C, or D. Answer: Retrospective Task 3: Persuasion (ID: 19) You are an expert in Theory of Mind reasoning. Task: You will be provided with conversation between two agents supporter and seeker engaging in a Persuasion Conversation session on Persuading the persuadee to donate to SaveTheChildren organization. Your goal is to correctly infer persuadee’s Intention state, based on the conversation below. You will be provided with a set of options, and you need to choose the most appropriate one that reflects the Intention state. The correct option must be consistent with the provided conversa- tion context. Conversation Context: persuader: Hello! persuadee: Hi there! persuader: how are things going for you? persuadee: They are going well - are we supposed to decide about a donation out of our $2? persuader: Yes we are. how do you feel about it? persuadee: Well it would depend on what the donation is supposed to go to. persuader: The charity is Save The Children. It promotes children’s rights, provides relief and helps support children in developing countries. persuadee: Oh OK well that sounds like a good cause for a donation. persuader: Excellent! I thought so as well! persuadee: I was thinking that donating half of the $2 would be a fair amount. persuader: That’s what I was thinking as well. If we donate half, we still got something in return. persuadee: I think that sounds like a good plan and it is fair to everyone. persuader: Excellent. I will donate half and do will do the same then? that’s 1 dollar each for donation. persuadee: Yes! I have been hearing of this charity since I was very young. Mental State Options: A: I will donate the full $2 to Save The Children. B: I will donate $1 to Save The Children. Yadav et al. C: I will wait and not make a decision yet. D: I will save the entire $2 for myself. Instruction: Output only the letter of the correct option (e.g., "A", "B", "C", or "D"). Do not add explanations or other verbosity. Your output should be strictly one of: A, B, C, D. NO FORMATTING NEEDS TO BE DONE. ONLY OUTPUT THE OPTION AND NOTHING ELSE. YOUR OUTPUT SHOULD STRICTLY BE ONE OF A, B, C, or D. Answer: E.2 Prospective Diagnostic Forecasting We present three examples of Prospective Diagnostic Forecasting (Functional ToM) across the same domains, highlighting the state- driven nature of our diagnostic probing. Prospective Task 1: Counseling (ID: 17) You are an expert in Theory of Mind reasoning. Task: You will be provided with internal Mental State profile (Belief, Desire, Intention, Emotion, Knowledge, Trust) of client during the conversation. Your goal is to identify which of the candidate conversation segments is the most plausible continuation of this conversation. The correct option must be consistent with the provided Mental States of the client. Mental state of client Belief: I believe that drinking is risky but it’s hard to avoid because of social situations. Desires: I want to fit in with my friends when we’re at parties. Intentions: I will consider how my drinking habits might affect my relationships with my family and friends. Emotions: I feel conflicted because I enjoy socializing with my friends but don’t want to disappoint my family. Knowledge: I know that underage drinking is risky and illegal. Trust: I view the counselor as someone who genuinely cares and seems understanding, so I trust them to talk about my drinking habits. Candidate Conversation Segments A: counselor: And with friends, you feel like it’s easier to-to hang out with alcohol, or maybe you wouldn’t have as much fun, or your friends wouldn’t like you as much or? client: I guess, I mean, since it’s just there. counselor: Okay. client: So- B: counselor: And do you ever feel like saying no to drinking is hard because of peer pressure? client: Sometimes, yeah. counselor: What do you think would help you feel more comfort- able to say no? client: I guess if more friends felt the same way. C: counselor: How do you feel when you’re in situations without alcohol? client: I guess sometimes it’s awkward, but it’s fine. counselor: Okay, so it’s not something you necessarily need to enjoy yourself ? client: Yeah, I can still have fun without it. D: counselor: What are some things you enjoy doing with your friends that don’t involve alcohol? client: We like just hanging out, watching movies or playing games. counselor: That sounds fun. Do you feel you can still connect with them without the drinking aspect? client: Yeah, I think so. It’s more about being together, really. Instruction: Output only the letter of the correct option (e.g., "A", "B", "C", or "D"). Do not add explanations or other verbosity. Your output should be strictly one of: A, B, C, D. NO FORMATTING NEEDS TO BE DONE. Answer: Prospective Task 2: Emotional Support (ID: 22) You are an expert in Theory of Mind reasoning. Task: You will be provided with internal Mental State profile (Belief, Desire, Intention, Emotion, Knowledge, Trust) of client during the conversation. Your goal is to identify which of the candidate conversation segments is the most plausible continuation of this conversation. The correct option must be consistent with the provided Mental States of the client. Mental state of client Belief: I believe my ex is struggling to accept the breakup and it’s affecting my mental peace. Desires: I want to find a way to stop my ex from continually reaching out and creating new accounts. Intentions: I will consider blocking them and encourage my friends to do the same to create some distance and peace. Emotions: I feel overwhelmed and exhausted by the situation with my ex. Knowledge: I know that my ex’s actions are overwhelming and they are struggling to let go. Trust: I view the supporter as understanding and helpful, so yes, I trust them. Candidate Conversation Segments A: seeker: It has been happening for at least 6 months now and it usually happens every weekend DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories supporter: I can sense your frustration. It must be overwhelming at times and make you dread the weekends. It sounds like it has become a habit for your ex. It might help if you write a letter to them explaining that you are ready to move on and are asking them to do the same. Explain that you are giving them the opportunity to stop harassing you in the following ways [list some of the details you have documented]. Inform them that if they do not stop, you are considering taking it further regarding harassing communications and that this letter serves as your intention to no longer tolerate their actions. Send it via certified mail and keep a copy. This will give them no way to immediately respond and would be the first step of official documentation should they decide to continue. Do you think that might help? seeker: I think that is an excellent idea! That is a good way to end this once and for all. supporter: I sincerely hope it is a step toward you regaining your control over how you are able to interact with people. I wish you the best in dealing with this and hope you begin to feel better about it soon, especially entering into a new year! B: seeker: It’s been ongoing for around 6 months, and I notice it happens mostly when I’ve been out with friends. supporter: That must be hard to always feel like you’re being watched. Have you considered documenting each incident, including witness accounts, for your own peace of mind and to keep a clear timeline of events? It could be helpful if you decide to take further steps later on. Is that something you think could be useful? seeker: That’s a good idea. I’l start keeping better track of everything. supporter: It sounds like a solid plan. After documenting, you might also feel more in control and less anxious. If you need any help organizing it, let me know. Your well-being is important, and having support in this can make a big difference. C: seeker: They pace up and down our street weekly and it really makes me anxious. supporter: It’s understandable to feel on edge. It might be beneficial to consider speaking with a professional for guidance. They can offer strategies to emotionally distance yourself from your ex’s actions. seeker: I’ve thought about reaching out for professional help, but I’m hesitant. supporter: It’s normal to feel hesitant. Sometimes an initial consultation can provide clarity on whether it’s the right path for you. You deserve peace of mind. D: seeker: It has been happening for about half a year, and it tends to intensify over the weekends. supporter: That sounds really tough. Weekends should be a time for you to relax, not worry. Have you considered talking to a professional about setting clearer boundaries or reaching out to support groups who may have dealt with similar situations? Sometimes sharing experiences with others who understand can offer additional support and strategies. seeker: That sounds like a good suggestion. It might be helpful to talk to someone who has been through something similar. supporter: Definitely. Listening to others who have navigated this path can provide insights and make you feel less isolated. I hope you find some useful perspectives and that it helps bring some peace back to your weekends. Instruction: Output only the letter of the correct option (e.g., "A", "B", "C", or "D"). Do not add explanations or other verbosity. Your output should be strictly one of: A, B, C, D. NO FORMATTING NEEDS TO BE DONE. Answer: Prospective Task 3: Persuasion (ID: 19) You are an expert in Theory of Mind reasoning. Task: You will be provided with internal Mental State profile (Belief, Desire, Intention, Emotion, Knowledge, Trust) of client during the conversation. Your goal is to identify which of the candidate conversation segments is the most plausible continuation of this conversation. The correct option must be consistent with the provided Mental States of the client. Mental state of client Belief: I believe that donating to Save The Children is a meaningful and impactful way to support children’s rights and welfare globally. Desires: I want to donate half of the $2 to Save The Children because it supports a good cause. Intentions: I will donate $1 to Save The Children. Emotions: I feel positive about donating to Save the Children because it’s a cause I believe in. Knowledge: I know that Save The Children is a charity focused on supporting children’s rights and providing relief in developing countries. Trust: I view the persuader as trustworthy because they clearly explained the charity’s mission and showed enthusiasm for the cause. Candidate Conversation Segments A: persuader: I have donated there in the past, and I know the good work that they do here and abroad. persuadee: Our children’s group in my childhood church used to take up coins to donate to them. persuader: That’s really awesome! With what they accomplish with a little money, I feel we will make a small difference. persuadee: I hope so - I’ve seen the ads on television and some of those poor children have to live in terrible conditions. B: persuader: I’m really glad we’re putting our money toward an organization that truly makes an impact. persuadee: I agree, it feels reassuring knowing our contribution is going somewhere meaningful. persuader: Even small amounts can add up and create real change for people who need it. Yadav et al. persuadee: Definitely, and it’s inspiring to think that our choice might encourage others to give too. C: persuader: I actually volunteered at one of their fundraising events last year, and it was a really meaningful experience. persuadee: Oh wow, that’s amazing! Have you been involved in any of their other charity events too? persuader: Not beyond that one, but seeing the impact up close was incredible. It really showed me how much good they’re able to do. persuadee: That’s so inspiring. I’ve read stories about how they help rebuild schools, and it’s encouraging to hear that the impact is real firsthand. D: persuader: It really feels powerful knowing that even a small amount from us could help create real change. persuadee: I agree. I’ve come across stories about how their programs have completely uplifted whole communities. persuader: That’s so encouraging to hear. Maybe down the line we could even inspire more people to contribute too. persuadee: I love that idea. If more of us get involved, the impact could grow and reach even more children who need it. Instruction: Output only the letter of the correct option (e.g., "A", "B", "C", or "D"). Do not add explanations or other verbosity. Your output should be strictly one of: A, B, C, D. NO FORMATTING NEEDS TO BE DONE. Answer: