Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Probabilistic Modeling of Intentions in Socially Intelligent LLM Agents

Feifan Xia, Yuyang Fang, Defang Li, Yantong Xie, Weikang Li, Yang Li, Deguo Xia, Jizhou Huang

TL;DR

The paper tackles the challenge of social intelligence in open-ended, multi-turn dialogue by modeling the partner’s hidden intentions as a probabilistic belief. It introduces the Stochastic Theory of Mind (SToM) framework, comprising an Intention Model (IM) that maintains B_t(theta), a Likelihood Model (LHM) that computes P(o_{t+1} | theta) to update beliefs, and a confidence-aware policy that conditions actions on uncertainty. Beliefs are continuously updated via Bayesian inference and serialized into prompts to guide the agent, all without parameter training. In SOTOPIA evaluations, SToM improves the Overall score by up to 9.0% on All and 4.1% on Hard compared to a strong base model, and even slightly surpasses an oracle with ground-truth intentions, highlighting the practical value and robustness of probabilistic intent modeling for socially intelligent LLM agents.

Abstract

We present a probabilistic intent modeling framework for large language model (LLM) agents in multi-turn social dialogue. The framework maintains a belief distribution over a partner's latent intentions, initialized from contextual priors and dynamically updated through likelihood estimation after each utterance. The evolving distribution provides additional contextual grounding for the policy, enabling adaptive dialogue strategies under uncertainty. Preliminary experiments in the SOTOPIA environment show consistent improvements: the proposed framework increases the Overall score by 9.0% on SOTOPIA-All and 4.1% on SOTOPIA-Hard compared with the Qwen2.5-7B baseline, and slightly surpasses an oracle agent that directly observes partner intentions. These early results suggest that probabilistic intent modeling can contribute to the development of socially intelligent LLM agents.

Probabilistic Modeling of Intentions in Socially Intelligent LLM Agents

TL;DR

The paper tackles the challenge of social intelligence in open-ended, multi-turn dialogue by modeling the partner’s hidden intentions as a probabilistic belief. It introduces the Stochastic Theory of Mind (SToM) framework, comprising an Intention Model (IM) that maintains B_t(theta), a Likelihood Model (LHM) that computes P(o_{t+1} | theta) to update beliefs, and a confidence-aware policy that conditions actions on uncertainty. Beliefs are continuously updated via Bayesian inference and serialized into prompts to guide the agent, all without parameter training. In SOTOPIA evaluations, SToM improves the Overall score by up to 9.0% on All and 4.1% on Hard compared to a strong base model, and even slightly surpasses an oracle with ground-truth intentions, highlighting the practical value and robustness of probabilistic intent modeling for socially intelligent LLM agents.

Abstract

We present a probabilistic intent modeling framework for large language model (LLM) agents in multi-turn social dialogue. The framework maintains a belief distribution over a partner's latent intentions, initialized from contextual priors and dynamically updated through likelihood estimation after each utterance. The evolving distribution provides additional contextual grounding for the policy, enabling adaptive dialogue strategies under uncertainty. Preliminary experiments in the SOTOPIA environment show consistent improvements: the proposed framework increases the Overall score by 9.0% on SOTOPIA-All and 4.1% on SOTOPIA-Hard compared with the Qwen2.5-7B baseline, and slightly surpasses an oracle agent that directly observes partner intentions. These early results suggest that probabilistic intent modeling can contribute to the development of socially intelligent LLM agents.
Paper Structure (23 sections, 4 equations, 1 figure, 1 table)

This paper contains 23 sections, 4 equations, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Problem Formulation
  5. Stochastic Theory of Mind (SToM) Framework
  6. Intention Model (IM)
  7. Initialization Phase.
  8. Update Phase.
  9. Likelihood Model (LHM)
  10. Confidence-Aware Policy
  11. High Confidence Regime ($C_t > \tau_{high}$).
  12. Low Confidence Regime ($C_t < \tau_{low}$).
  13. Medium Confidence Regime ($\tau_{low} \leq C_t \leq \tau_{high}$).
  14. Implementation Details
  15. Experimental Setup
  16. ...and 8 more sections

Figures (1)

  • Figure 1: Overview of the Stochastic Theory-of-Mind (SToM) framework. At each dialogue turn, the Intention Model (IM) produces a prior belief distribution over the partner’s possible intentions. After observing the partner’s utterance, the Likelihood Model (LHM) updates the distribution via Bayesian inference. The updated belief is serialized and appended to the policy model’s prompt, guiding subsequent actions. This process enables belief-driven adaptation without any parameter training.