MetaMind is a novel multi-agent framework designed to endow Large Language Models (LLMs) with more human-like social reasoning capabilities. Traditional LLMs often struggle with the ambiguity and indirectness inherent in real-world human communication, failing to grasp unspoken intents, implied emotions, or culturally sensitive cues. MetaMind addresses this gap by explicitly modeling the multi-stage cognitive processes humans use to reason about unobservable mental states and navigate socially complex environments.

Inspired by psychological theories of metacognition and Theory of Mind (ToM), MetaMind employs a staged and collaborative system of three specialized agents and user modeling algorithm to interpret, refine, and respond to social interactions with greater nuance, empathy, and cultural sensitivity. Notably, metacognition is an awareness and understanding of one's own thought processes, mirrors by planning on declarative knowledge, monitoring procedural knowledge, and evaluating conditional knowledge to adapte its understanding in light of social norms and interactive constraints.

“What is meant often goes far beyond what is said, and that is what makes conversation possible.” - H. P. Grice

Key Features & Core Concepts

MetaMind's architecture is built upon a layered reasoning process involving three distinct agents:

1. Theory-of-Mind (ToM) Agent: This agent initiates the reasoning process by generating multiple hypotheses about the user's latent mental state (beliefs, desires, intentions, emotions, thoughts) based on contextual and social cues. It aims to move beyond literal interpretations to infer what the speaker might be truly trying to convey.

   • Hypothesis Generation: Produces a set of commonsense-based candidate mental state interpretations with explanations and type labels.
   • Mental-State Reasoning: Cross-referencing candidate mental state interpretations with social memory, identifying ToM markers, and diversifying hypotheses.

2. Domain Agent: This agent refines the hypotheses generated by the ToM Agent by incorporating socially grounded constraints. It assesses the appropriateness of interpretations against broader norms like cultural expectations, ethical guidelines, and situational context, ensuring socially responsible and domain-aware reasoning.

   • Hypothesis Refinement and Selection: Revises hypotheses based on domain rules (cultural, ethical, role-based) and selects the most appropriate one using a well-defined scoring mechanism that balances contextual plausibility and information gain.

3. Response Agent: This agent generates a contextually appropriate output based on the refined optimal hypothesis and the user's social memory (e.g., emotional patterns, prior preferences). It also includes a self-validation mechanism to ensure the response aligns with the inferred intent and maintains social and semantic quality.

   • Generation and Validation: Produces a natural language response conditioned on the selected hypothesis and social memory, then validates it for empathy and coherence. The validation mechanism will also update social memory, allowing MetaMind to predict in advance during the interaction and know the consequences of its actions.

Social Memory is a dynamic, structured knowledge base that evolves across interactions to capture long-term user patterns, social norms, and feedback-based adjustments. It is designed around three core principles: (1) Grounding in context: Memory is initialized based on the intuitive understanding of situation and roles involved in the interaction. (2) Updating through user modeling: Updating memory via validated long-term interpretations of user mental states. (3) Improving through feedback: It incorporates signals from failures or corrections to better guide future responses.

📊 Datasets and Metrics

MetaMind's performance is evaluated on a diverse set of challenging benchmarks that cover various aspects of social reasoning:

1. ToMBench:

   • Description: A multiple-choice benchmark designed to test Theory-of-Mind (ToM) reasoning across six key dimensions: Emotion, Desire, Intention, Knowledge, Belief, and Natural Language Communication.
   • Alignment: Primarily evaluates the capabilities of the ToM Agent (Stage 1) in generating structured mental state hypotheses.
   • Metric: Accuracy.

2. Social Cognition Tasks:

   • Description: A suite of tasks from ToMBench that probe context-sensitive reasoning under social, cultural, and ethical norms. This includes eight real-world tasks such as Faux Pas Recognition Test (FRT), Scalar Implicature Test (SIT), and the Ambiguous Story Task (AST).
   • Alignment: Assesses the Domain Agent's (Stage 2) ability to refine interpretations based on domain-specific constraints.
   • Metric: Accuracy.

3. Social Tasks in Sandbox Simulation (STSS):

   • Description: A benchmark testing goal-oriented social interaction in open-ended, interactive scenarios across six domains: Conversation, Public Activity, Appointment, Inviting Companions, Online Activity, and Asking for Help.
   • Alignment: Tests the Response Agent's (Stage 3) capacity to generate contextually appropriate and validated responses.
   • Metric: Average Score.

For more detailed experiments (SOTOPIA, SocialIQA), Human Study, and Qualitative Study, please refer to the paper.

🧪 Experimental Results

MetaMind demonstrates significant improvements over baseline LLMs and existing methods across all evaluated benchmarks. Here are some highlights:

Performance on Theory-of-Mind Reasoning (ToMBench)

MetaMind consistently boosts ToM reasoning capabilities across various LLMs, including both proprietary models like GPT-4, open-source alternatives, and SOTA reasoning models. For instance, MetaMind-enhanced GPT-4 achieves an average accuracy of 81.0% on ToMBench, a notable improvement from the base GPT-4's 74.8%. It is worth noting that with the support of MetaMind, we have narrowed the gap between all models and the average human performance, and even inspired some advanced models to surpass human performance in one or more tasks, which is a major breakthrough in AI for social tasks.

Ablation Study: The Importance of Each Stage

An ablation study was conducted on the STSS benchmark to validate the contribution of each stage in the MetaMind framework. Removing any stage leads to a noticeable degradation in performance, underscoring the importance of the layered cognitive architecture.

Table: Ablation study on STSS (Average Score %)

Configuration	Conv.	Pub. Act.	Appo.	Inv. Com.	Online Act.	Help	Avg.
MetaMind	80.8	81.9	65.0	67.1	75.1	73.0	73.9
w/o Stage 1 (ToM)	78.1	78.4	59.0	60.3	72.1	62.3	68.3
w/o Stage 2 (Domain)	79.2	79.3	61.7	62.2	73.7	67.0	70.5
w/o Stage 3 (Response)	58.7	67.2	54.2	43.2	61.9	61.7	57.8
w/o SocialMemory	70.5	72.3	57.0	58.0	64.8	61.2	63.9

• Stage 1 (Mental-State Reasoning): Removing structured mental-state reasoning leads to an average drop, particularly in high-ambiguity tasks.
• Stage 2 (Norm-Aware Refinement): Removing domain-specific rule refinement results in a substantial performance drop, especially in tasks requiring pragmatic interpretation or adherence to social norms.
• Stage 3 (Response via Validation): Bypassing response validation significantly impacts overall performance, highlighting the criticality of this final step for high-quality, appropriate responses.

Parameter Sensitivity and Further Details

Sensitivity analyses for key hyperparameters (k for hypothesis count, λ for Domain Agent trade-off, β for Response Agent trade-off) confirm the robustness of MetaMind's performance across a range of parameter settings.

Below are visualizations illustrating the parameter sensitivity for the number of hypotheses (k) in the ToM Agent. These surfaces show how performance (average accuracy on ToMBench) varies with changes in λ (Domain Agent) and β (Response Agent) for different values of k, helping in understanding the interplay between different parameters and selecting optimal values for k, λ, and β to achieve the best performance.

Parameter Sensitivity for k=6	Parameter Sensitivity for k=7	Parameter Sensitivity for k=8

For more detailed experimental setups, results on (additional) benchmarks, and qualitative case studies, please refer to the full research paper.

Running the Project

Follow these steps to get MetaMind up and running on your local machine. At this point, we are developing the project website including MetaMind playground and more.

1. Clone the Repository

First, clone the MetaMind repository to your local system using Git:

git clone <repository-url> # Replace <repository-url> with the actual URL of your Git repository
cd MetaMind

2. Setup and Configuration

It's recommended to set up a virtual environment for MetaMind.

python -m venv venv
# On Windows
.\venv\Scripts\activate
# On macOS/Linux
source venv/bin/activate

Install the required dependencies:

pip install -r requirements.txt

⚙️ Configuration

Before running the project, you need to configure the settings in the config.py file. This file contains crucial parameters for the LLM API, agent behaviors, and social memory.

Important: You MUST replace the placeholder values in config.py with your actual credentials and desired settings. Specifically, pay attention to the api_key and base_url for the LLM.

Here's an overview of the key configurable parameters in config.py:

Category	Parameter	Description	Default Value
LLM API Settings	api_key	Your API key for the LLM service (e.g., OpenAI).	"your-openai-key"
	base_url	The base URL for the LLM API (if using a custom or self-hosted endpoint).	"your-openai-url"
	model_name	The specific LLM model to use (e.g., "gpt-4o").	"gpt-4o"
	temperature	Controls the randomness of the LLM's output.	0.7
	max_tokens	The maximum number of tokens the LLM can generate in a single response.	1000
ToM Agent	hypothesis_count	Number of mental state hypotheses to generate.	7
	target_diversity	Desired diversity score among generated hypotheses.	0.4
	evidence_threshold	Minimum evidence level required for a hypothesis.	"medium-high"
Domain Agent	lambda	Weight for balancing prior and conditional probabilities in hypothesis scoring.	0.7
	epsilon	Small constant to avoid log(0) errors.	1e-10
Response Agent	beta	Trade-off weight between empathy and coherence in response generation.	0.8
	utility_threshold	Minimum utility score for an acceptable response before revision.	0.9
	max_revisions	Maximum number of revisions for a response if below utility threshold.	3
Social Memory	memory_decay_rate	Rate at which the importance of memory items decays over time.	0.05
	max_memory_items	Maximum number of items to store in the social memory.	100

Make sure to review and update these settings in config.py according to your needs and API access.

3. Running the Terminal Version

To run the terminal-based version of MetaMind, execute the main.py script:

python main.py

This will typically start an interactive command-line interface where you can interact with the MetaMind agents.

4. Running the Web Version

To run the web-based interface for MetaMind, execute the app.py script:

python app.py

This will start a local web server. Open your web browser and navigate to the URL provided in the terminal output (commonly http://127.0.0.1:5000 or similar).

Citation

We now have a paper you can cite:

@article{metamind,
  title={MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems},
  author={Zhang, Xuanming and Chen, Yuxuan and Yeh, Min-Hsuan and Li, Yixuan},
  journal={arXiv preprint arXiv:2505.18943},
  year={2025}
}