DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories

arXiv:2604.20443 · cs.CL, cs.AI, cs.LG · Submitted 2026-08-23 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories".

Jane: The paper was written by Neemesh Yadav, Palakorn Achananuparp, Jing Jiang and Ee-Peng Lim from Singapore Management University and Australian National University.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title and Authors: Tom: The authors, Neemesh Yadav, Palakorn Achananuparp, Jing Jiang, and Ee-Peng Lim have developed this framework to test the "Theory of Mind" in AI.

Jane: Theory of Mind is essentially our ability to understand what other people are thinking or feeling—why they say what they say.

Lu: These researchers are challenging us by forcing the AI to move beyond just understanding that into actual action, which is a huge conceptual leap for me.

Meng: It's a strong team, and I wonder if their background in naturalistic human-human dialogues means the benchmark is more grounded than other synthetic ones we’ve seen.

Lalam: The focus on dialogue trajectories suggests they want to see how social intelligence translates into actual conversational flow, not just static labels.

Tom: It's a huge step forward from just trying to make the AI guess feelings; it requires predicting what comes next based on those feelings.

Jane: This framework is designed to test if the AI can actually use its understanding of human motivation to drive the conversation forward in a way that makes sense.

Lu: By focusing on "forecasting state-driven trajectories," they are setting up a very specific, measurable task for us, which is incredibly useful for research.

Meng: I'm interested in how the authors are using real conversations—counseling and persuasion—to make sure this test doesn's applicable to the real world.

Lalam: Using human interaction makes sense because that’s where our goal is to improve culture, right? We need AI that understands human intent, not just people talking around a computer screen.

Summary and Abstract Findings: Tom: The abstract really sets the stage by highlighting a critical gap in current AI capabilities.

Jane: They found that LLMs are really good at "Literal ToM"—meaning they can identify what someone is feeling or believing based on the context.

Lu: But they struggle with "Functional ToM," which is using those feelings to predict how the conversation will actually proceed.

Meng: It’s a frustrating finding for many engineers because, even if an AI knows exactly what you mean, it can't always figure out the next logical step in the interaction.

Lalam: If we can solve this gap, Lalam thinks that is huge for how AI interacts with people emotionally.

Tom: The paper suggests that current models are quite good at identifying mental states, which is impressive in itself.

Jane: But "DialToM" shows us that identifying the state isn't enough to predict the next natural move.

Lu: It’s a systemic failure to apply the knowledge, and that's what they found when testing across fourteen different state-of-the-art models.

Meng: That finding of near or below chance on the functional task is very stark; it shows we have fundamental limitations in how we model social dynamics yet have high expectations for LLMs.

Lalam: If AI can't transition from just understanding to actually acting based on human intent, it won't be able to serve as a helpful or empathetic partner in the future interactions we want.

Improvements and Methodology: Tom: The core of the paper is how they designed this test, which is really clever.

Jane: They use a "bifurcated" approach, meaning they look at the task in two different ways to get a clear picture.

Lu: They separate it into a "Retrospective Inference" task and that context-free "Prospective Diagnostic Probe."

Meng: The Prospective Diagnostic Probe is the key here; they strip away all the history, forcing us to rely only on a single isolated mental state profile.

Lalam: That isolation is where we see if AI can truly reason, because it can't use surface-level conversational patterns to cheat.

Tom: It's a way to prevent models from using "lexical shortcuts," which is exactly what they want to avoid.

Jane: By removing the conversation history, they are making the test incredibly rigorous, which is why their results are so reliable and hard to dispute.

Lu: This method provides a clear way to measure genuine functional reasoning without any ambiguity from relying on past conversational context.

Meng: I see how this could be used in practical applications, like having an AI assistant predict how a negotiation might flow based purely on the current emotional state of the other party.

Lalam: It’s about building a trust in the AI's reasoning that transcends simply seeing patterns and starts to seeing causality.

Conclusion and Wrap-up: Tom: So, we've seen how "DialToM" establishes a stark gap between what AI can infer versus what it can functionally apply.

Jane: The results are quite clear; the models are excellent at reading the room but struggle when they try to act on that information.

Lu: The fact that a domain expert achieved one hundred percent accuracy really shows that the task is sound and not just some random difficulty problem.

Meng: I think this finding, especially with Gemini three Pro scoring around eighty-three percent on the functional test, gives us a concrete target for building more advanced systems.

Lalam: It offers hope that we can design AI that not only understands human feeling but also to better support and understand human relationships in the future.

Tom: And we'll wrap up today's discussion of "DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories."

Jane: It really highlights that true understanding is more than just seeing what's on the surface, Tom.

Lu: It shows the complexity of social intelligence and how much we still have to learn from human interaction.

Meng: We need to start thinking about practical ways to make this gap smaller, especially with such capable models existing right now.

Lalam: I hope that by using "DialToM", we can help build a world where AI truly understands our human experiences and makes us feel more connected.

Neemesh Yadav, Palakorn Achananuparp, Jing Jiang, Ee-Peng Lim

Singapore Management University · Australian National University

cs.CL, cs.AI, cs.LG

Submitted: 2026-08-23

Updated: 2026-08-25

Code: https://github.com/Stealth-py/DialToM

Importance score: 87/100

The gist: DialToM is introduced as a "Theory of Mind (ToM) benchmark built from naturalistic human–human dialogues using a multiple-choice evaluation framework" designed to address the "gap between explicit

Key concepts

Theory of Mind (ToM)
The ability to understand what other people are thinking or feeling, and figuring out why they say what they say. The benchmark uses this concept to test if AI can grasp human motivation beyond just surface-level conversation.
Literal ToM
This refers to the capability of current LLMs to identify a person's mental state—what they are feeling or believing—simply by analyzing the context provided in a conversation. The paper notes that models are quite good at this task.
Functional ToM
This is a more advanced test that requires AI to use its understanding of human feelings and motivation to predict how the conversation will actually proceed. The benchmark shows LLMs struggle with this predictive action.

Terminology

Summary

DialToM is introduced as a Theory of Mind (ToM) benchmark built from naturalistic human–human dialogues using a multiple-choice evaluation framework designed to address the gap between explicit mental-state inference and applied ToM in synthetic settings. This gap is defined by the distinction between Literal ToM and Functional ToM.

The study operationalizes this division by establishing a dual-stage diagnostic probe. The first stage is the Retrospective inference task (Literal ToM: inferring states from dialogue), where models are provided with conversational context to infer mental states. The second, and more critical, stage is the Prospective Diagnostic Probe (Functional ToM: forecasting actions from isolated mental states). In this latter setup, models must forecast state-consistent dialogue trajectories based solely on an isolated mental-state profile without dialogue context.

DialToM extends existing benchmarks by incorporating several critical features:

  1. Ecological Validity: Grounding the evaluation in high-stakes human dialogues (counseling and persuasion).

  2. Richer Attribute Model: Utilizing a six-attribute model—Belief, Desire, Intention, Emotion, Knowledge, and Trust (T)—to capture relational dynamics.

  3. ** Rigorous Design:** Enforcing the context-free nature of the Prospective task to prevent models from relying on spurious context-to-dialogue correlations or superficial lexical shortcuts.

The empirical analysis reveals a systematic reasoning asymmetry. The findings show that LLMs are proficient at the first stage, as they excel at inferring mental states (Literal ToM), but they significantly struggle with the second stage, failing to leverage those inferences for behavioral forecasting. This is quantified by a stark human-AI capability gap: a domain expert achieves 100% accuracy on this task.

The study further identifies a notable exception to this trend. Gemini 3 Pro achieved high performance, maintaining robust Functional ToM capabilities for context-free forecasting with approximately 83% accuracy on the Prospective task. This was further demonstrated through a teacher-student reasoning injection probe, which showed that injecting Gemini 3 Pro’s reasoning traces could boost the performance of weaker models by up to 76 points, providing an empirical proof of concept for functional ToM-grounding in LLMs.

In summary, DialToM provides a rigorous framework that isolates true functional ToM reasoning—the ability to translate latent mental states into action—from mere pattern matching or context exploitation.

Improvements for AI systems

The DialToM framework provides a critical methodological blueprint for moving beyond superficial performance metrics in Theory of Mind (ToM). The core improvement lies in transforming LLMs from models that infer states to models that functionally utilize those states.

Based on this research, I propose the following specific improvements and the resulting capabilities of an enhanced AI system:

The Improvement: We will re-engineer the predictive layer to strictly enforce a context-free, state-driven mapping. Instead of relying on standard sequence prediction where conversational history (H) is implicitly leveraged, the model must be trained and evaluated using an isolated mental state profile (S) to predict the next logical dialogue trajectory (A).

Model M: S to A

The Capability: The improved AI system will demonstrate Functional ToM. It will not merely label a belief (Literal ToM); it will actively use that belief, combined with other state attributes, to predict the most action-consistent subsequent utterance, even when no surrounding conversational context is provided.

The Improvement: We will integrate a sixth foundational attribute—Trust (T)—into the state profile S. This moves beyond simple BDI (Belief, Desire, Intention) to capture relational rapport and its influence on decision-making.

S = B, D, I, E, K, T

The Capability: The AI will exhibit Relational Reasoning. It will modulate its predicted dialogue trajectory based on the perceived trust level between two agents. For example, if the Trust attribute is high (T high), the model will predict a collaborative or persuasive trajectory; if T low, it will predict a defensive or evasive one, mirroring human social intelligence in high-stakes interactions (e.g., counseling).

The Improvement: We must implement a rigorous evaluation protocol that goes beyond simple accuracy by introducing adversarial Easy Set distractors alongside the standard Hard Set. This is an ablation designed to eliminate reliance on superficial lexical shortcuts or topical coherence.

Evaluation Metric: CR Hard CR Easy

The Capability: The improved AI will achieve True Generalization. Its performance will remain high even when contextual scaffolding is removed, ensuring that its successful prediction is rooted in genuine state-driven logic, not merely pattern matching against the most common conversational tropes.

The Improvement: We will implement a comparative diagnostic probe (Context Degradation Analysis) to quantify the utility of historical context. This involves systematically degrading the availability of H (the history) while holding S constant, measuring the corresponding drop in predictive accuracy.

Measure: Accuracy = f(Availability of H)

The Capability: The AI will demonstrate Contextual Scaffolding Awareness. It will precisely quantify how much contextual information is required for optimal performance, allowing developers to tune the model's operational constraints and ensures that the system is not failing due to a lack of memory, but a lack of functional reasoning.

In summary, the improved AI system will transition from being a context-matching language predictor to an autonomous state-to-action reasoner, possessing verifiable Functional ToM.

Sources

Related papers