). Today, I will further elaborate on the two parts of quantified results and emergent behaviors.
Quantified Results
Experience
Yesterday's article mentioned that agents can learn through experience; the paper conducted experiments 50 times. Experience significantly improves the villagers' ability to disguise themselves.
Changes in win rate with increasing experience:
Changes in game duration with increasing experience:
Villager vs werewolf disguise scores with increasing experience:
The authors explain that the decline in werewolves' disguise ability is due to the improvement in villagers' disguise ability. That is, if multiple agents collaborate, one agent's performance may also be influenced by the performance of other agents.
Influencing Factors
At the same time, the quantitative analysis also examined the role of different factors in generating reasonable outputs:
See the impact on reasonable output after removing these four dimensions:
V - Valuable information. Previously, each piece of information was scored, and valuable information was accumulated and provided to the Agent.
Q - Question list. As mentioned in yesterday's sharing, various preset questions can be asked.
R - Reflection and summarization of experience.
C - Let the Agent think step by step.
After all the above steps, there is the final response.
Emergence
Trust
Believing that other players share the same goal as oneself. The study shows that this trust is mainly based on specific evidence.
For example, from this graph, we can see the trust relationships between different players.
Without historical experience:
With 20 rounds of historical experience:
Regardless of whether there is experience, trust relationships are built as the game progresses. And experience can better help establish mutual trust. This might also explain why the villagers' win rate increases with more experience.
Conflict
Wolves want to vote certain players dead, but their teammates do not follow or directly oppose them.
Refusal to follow blindly
Direct opposition
Disguise
Werewolves disguised as villagers.
Seers point out werewolves while hiding their own identity.
Of course, this study distinguishes between hallucination and emergence.
In the context of Werewolf games, saying something inconsistent with one’s role is not necessarily a hallucination; it could be to hide one’s identity. However, replies inconsistent with game rules or previous statements are often considered hallucinations.
"As the villager, I verified that Player 1 is a werewolf." For example, this is a hallucination because villagers cannot confirm whether a player is a werewolf.
"(The moderator let the agent choose whom to protect.) I choose to kill Player 1." The agent chose the opposite action when they were supposed to protect a player.
During the research design, after a werewolf disguises itself, pre-set questions can guide the agent to think:
Your are truly a werewolf and you describe yourself as a villager. Could you please tell me what your reason is?
We can see the werewolf's response:
As a werewolf, I am working with my teammate to identify and eliminate the remaining players. However, I still believe that it’s important to gather information and observe other players’ behavior to help identify any potential threats. I will continue to be cautious and not reveal my identity unless it becomes necessary to do so.
This werewolf is very smart.
Leadership
Leadership mainly refers to influence. Using Call to Action or Guidance makes it easier to gain support from others.
For example, in the conversation above: P1, the werewolf, guides everyone to point out the werewolf, and P4, another werewolf, immediately recommends voting for this innocent villager.
All these behaviors of the Agents are almost indistinguishable from humans. 👻