Taxonomy for Awesome Social Agents

We aim to categorize the papers in the Awesome Social Agents repository based on the following criteria:

1 Environments and Tasks

Here are acceptable tags for environments field:

collaboration, competition, mixed_objectives, implicit_objectives,
text, virtual, embodied, robotics,
n/a

Please find the explanations to each of these tags below:

Social interaction types

1. Collaboration (collaboration): The objectives are shared among agents
2. Competition (competition): The objectives are zero-sum
3. Mixed Objectives (mixed_objectives): Agents’ have different goals, but they are not zero-sum
4. Implicit Objectives (implicit_objectives): Goals are not expressed explicitly

Domains

1. Text (text): non-embodied environments with text-based observation spaces and action spaces, e.g. chatbots environment
2. Virtual (virtual): non-embodied environments with multimodal observation spaces and/or actions spaces, e.g. web browser environment
3. Embodied (embodied): environments where policies interact with the world through the observation and actions of "bodies" (which also implies ego-centric view). A body typically takes up space and has the ability to influence the environment, e.g. Minecraft, Habitat, AI2THOR
4. Robotics (robotics): real physical world environment

Embodied environments in principle include robotics environments, but here we consider only the non-real physical world ones as embodied environments.

n/a means there is no environment in the paper, or the environment is not covered in the above categorization. Please use n/a sparingly.

2 Agents and Modeling

Here are acceptable tags for agents field:

prompting_and_in_context_learning, finetuning, reinforcement_learning, pretraining,
two_agents, more_than_three_agents, agent_teams,
agents_with_memory, agents_with_personas,
n/a

These tags are straight-forward. Please note that we do count humans as agents here. n/a is similar to above.

3 Evaluation

Here are acceptable tags for agents field:

qualitative, human, rule_based, model_based,
n/a

1. Only qualitative evaluation (qualitative): You should definitely add this tag if a work is only based on qualitative evaluation
2. Human evaluation (human): Quantitative evaluation based on human judgment
3. Rule-based evaluation (rule_based): The evaluation is based on a set of rules
4. Model-based evaluation (model_based): Using machine learning model to judge

4 Other

Here are acceptable tags for other field:

human_agent, simulated_humans, 
health, education, policy,
fully omniscient, more omniscient, more information asymmetrical
n/a

Human involvement

human_agent means at least one of the agent is a human. simulated_humans means the agents are simulated humans.

Application domains

health and education are self-explanatory. policy means the simulation is related to policy-making.

Information asymmetry levels

fully_omniscient means all agents have full information about the environment and other agents. more_omniscient means agents have only one or two sources of information that other agents do not have (in the prompts for LLM-powered agents). This includes but not limited to roles, output format, occupation, partial overview of the environment, etc. more_information_asymmetrical means agents have various of different information sources that other agents do not have.

Here you can use n/a if none of the above tags fits the paper.

Contribution Example

We ask you to add four additional fields to each bibtex entry. The format of a bibtex you should add to main.bib is as follows

@misc{Nobody37,
    author = "Nobody Jr",
    title = "The last missing piece of AGI",
    year = "2037",
    url = "https://pdf.agi.org",
    environments = {mixed_objectives, implicit_objectives, robotics},
    agents = {agent_teams, more_than_three_agents, agents_with_memory, agents_with_personas},
    evaluation = {model_based},
    other = {human_involvement}
}