Learning Dynamic Belief Graphs for Theory-of-mind Reasoning
arXiv:2603.20170v1 Announce Type: new Abstract: Theory of Mind (ToM) reasoning with Large Language Models (LLMs) requires inferring how people's implicit, evolving beliefs shape what they seek and how they act under uncertainty -- especially in high-stakes settings such as disaster response, emergency medicine, and human-in-the-loop autonomy. Prior approaches either prompt LLMs directly or use latent-state models that treat beliefs as static and independent, often producing incoherent mental models over time and weak reasoning in dynamic contexts. We introduce a structured cognitive trajectory model for LLM-based ToM that represents mental state as a dynamic belief graph, jointly inferring latent beliefs, learning their time-varying dependencies, and linking belief evolution to information seeking and decisions. Our model contributes (i) a novel projection from textualized probabilistic statements to consistent probabilistic graphical model updates, (ii) an energy-based factor graph r
arXiv:2603.20170v1 Announce Type: new Abstract: Theory of Mind (ToM) reasoning with Large Language Models (LLMs) requires inferring how people's implicit, evolving beliefs shape what they seek and how they act under uncertainty -- especially in high-stakes settings such as disaster response, emergency medicine, and human-in-the-loop autonomy. Prior approaches either prompt LLMs directly or use latent-state models that treat beliefs as static and independent, often producing incoherent mental models over time and weak reasoning in dynamic contexts. We introduce a structured cognitive trajectory model for LLM-based ToM that represents mental state as a dynamic belief graph, jointly inferring latent beliefs, learning their time-varying dependencies, and linking belief evolution to information seeking and decisions. Our model contributes (i) a novel projection from textualized probabilistic statements to consistent probabilistic graphical model updates, (ii) an energy-based factor graph representation of belief interdependencies, and (iii) an ELBO-based objective that captures belief accumulation and delayed decisions. Across multiple real-world disaster evacuation datasets, our model significantly improves action prediction and recovers interpretable belief trajectories consistent with human reasoning, providing a principled module for augmenting LLMs with ToM in high-uncertainty environment. https://anonymous.4open.science/r/ICML_submission-6373/
Executive Summary
This article presents a novel approach to Theory of Mind (ToM) reasoning with Large Language Models (LLMs) by introducing a structured cognitive trajectory model that represents mental state as a dynamic belief graph. The model jointly infers latent beliefs, learns their time-varying dependencies, and links belief evolution to information seeking and decisions. The approach significantly improves action prediction and recovers interpretable belief trajectories consistent with human reasoning in high-uncertainty environments. The model's contributions include a novel projection from textualized probabilistic statements to graphical model updates, an energy-based factor graph representation of belief interdependencies, and an ELBO-based objective that captures belief accumulation and delayed decisions. The model has the potential to augment LLMs with ToM in high-stakes settings, such as disaster response and emergency medicine.
Key Points
- ▸ Introduction of a novel structured cognitive trajectory model for ToM reasoning with LLMs
- ▸ Representation of mental state as a dynamic belief graph
- ▸ Joint inference of latent beliefs, time-varying dependencies, and belief evolution
Merits
Strength in Addressing Dynamic ToM
The model effectively addresses the challenges of dynamic ToM by representing mental state as a dynamic belief graph, allowing for the joint inference of latent beliefs, time-varying dependencies, and belief evolution.
Improved Action Prediction and Belief Trajectory Recovery
The model significantly improves action prediction and recovers interpretable belief trajectories consistent with human reasoning in high-uncertainty environments.
Demerits
Potential Overfitting to Specific Datasets
The model's performance may be sensitive to the specific datasets used for training and validation, potentially leading to overfitting and reduced generalizability.
Limited Evaluation in Real-World Settings
The model's evaluation is limited to simulated disaster evacuation datasets, and further evaluation in real-world settings is necessary to fully assess its effectiveness.
Expert Commentary
The article presents a novel and innovative approach to ToM reasoning with LLMs, addressing the challenges of dynamic ToM through the use of a structured cognitive trajectory model. The model's contributions, including the novel projection from textualized probabilistic statements to graphical model updates, energy-based factor graph representation of belief interdependencies, and ELBO-based objective, demonstrate a deep understanding of the underlying challenges and requirements of ToM reasoning. While the model's potential is significant, further evaluation and testing in real-world settings are necessary to fully assess its effectiveness and generalizability. Additionally, the model's limitations, including potential overfitting to specific datasets, must be carefully considered and addressed in future work.
Recommendations
- ✓ Further evaluation and testing of the model in real-world settings, such as disaster response and emergency medicine.
- ✓ Investigation of the model's potential for overfitting to specific datasets and development of strategies to mitigate this risk.
Sources
Original: arXiv - cs.AI
