ML Newsnew | past | best | rss | submit about
Can Generative Multi-Agents Spontaneously Form a Society? (arxiv.org)
48 points by geuds on Sep 17, 2024 | hide | past | pdf | 5 comments on HN

In plain words: A new setup lets each AI agent spot and drop actions that would hurt social interactions, then pick friendlier ones. Unlike usual agents working alone, they formed cliques with ranks around a chosen leader and organized group activities on their own.

Abstract · Can Agents Spontaneously Form a Society? Introducing a Novel Architecture for Generative Multi-Agents to Elicit Social Emergence

Generative agents have demonstrated impressive capabilities in specific tasks, but most of these frameworks focus on independent tasks and lack attention to social interactions. We introduce a generative agent architecture called ITCMA-S, which includes a basic framework for individual agents and a framework called LTRHA that supports social interactions among multi-agents. This architecture enables agents to identify and filter out behaviors that are detrimental to social interactions, guiding them to choose more favorable actions. We designed a sandbox environment to simulate the natural evolution of social relationships among multiple identity-less agents for experimental evaluation. The results showed that ITCMA-S performed well on multiple evaluation indicators, demonstrating its ability to actively explore the environment, recognize new agents, and acquire new information through continuous actions and dialogue. Observations show that as agents establish connections with each other, they spontaneously form cliques with internal hierarchies around a selected leader and organize collective activities.

H. Zhang, J. Yin, M. Jiang, C. Su
arXiv:2409.06750 · cs.MA, cs.AI, cs.HC, cs.LG · submitted Sep 10, 2024 · updated Nov 19, 2024
abstract · pdf · html · 13 pages, 8 figures

add comment on HN

bbor on Sep 17, 2024 [–]
Fascinating look into how culture impacts research. This is done in China, and seemingly takes a 100% empirical mindset, deciding not to mention any of the theoretical frameworks that could back such a system, the most obvious of them being Marvin Minsky’s Society of Mind (highly recommend, the most hacker-style book on theoretical AI out there). If I had to guess, that’s 50% due to language barriers and 50% due to overall intellectual orientation.

Regardless, their results are very interesting! Dedicated doomers may remember that Stanford did a proof of concept for this almost exactly a year ago, but it was focused on “believability” of the society rather than the society’s collective ability to perform tasks: https://hai.stanford.edu/news/computational-agents-exhibit-b...

On a humorous side note: I was curious about their choice of a fun name for their simulated environment — “IrollanValley” - in an otherwise very empirical and dry paper, and was excited to learn something about Chinese folklore! But, uhh no lol, it’s just a reference to the video game Heroes of Might and Magic https://mightandmagic.fandom.com/wiki/Irollan

Thanks for posting! I’ve gotta hurry up and publish before this gets out of hand. In addition to the need for a Society of Mind discussion in any follow up paper, I personally hope to see a discussion of Terminator’s SkyNet, as well…

reply
krunck on Sep 17, 2024 [–]
This is amazing research. What I want to know is if they repeated the same experiment but trained the agents without language or, rather, on random data instead of language, would their gibbering eventually evolve into a structured language? After how many iterations?
reply
kridsdale3 on Sep 17, 2024 | parent [–]
In humans this happens. Isolated clans of non-socialized children (like the result of some disaster, or in a very neglected orphanage) have demonstrated spontaneous generation of novel language for collaboration amongst themselves.
reply
kridsdale3 on Sep 17, 2024 [–]
This immediately evoked in my memory David Chiang’s The Lifecycle of Software Objects. Here, though, there is no need for human caretakers to guide the socialization of the NPCs.
reply
byus on Sep 17, 2024 [–]
How is this not intentionally drawing a person?

If we write code that intentionally takes into consideration our own experience, why is it mesmerizing when it behaves as expected?

It isn’t spontaneously forming society if it’s weighted to do so in its initial states given the intentional implementation choices of the makers.

This is like watching a farmer farm and saying the carrots magically grew.

reply
About | RSS | RSS (all) | HN arXiv