search

agi-templar / Stable-Alignment

Multi-agent Social Simulation + Efficient, Effective, and Stable alternative of RLHF. Code for the paper "Training Socially Aligned Language Models in Simulated Human Society".
https://arxiv.org/pdf/2305.16960.pdf
Other
336 stars 18 forks source link

Is that the only diffenerce among agents in 10*10 grid society is the memory system? #4

Open Benstime opened 1 year ago

Benstime commented 1 year ago

If every agent models is same model, they acts the same behavior, why used 10×10 grids society, not 1×1 grid.

agi-templar commented 1 year ago

The agents are different in: 1. They are initialized with different memories, which are existing opinions on some questions; 2. Their memories will be updated differently during the procedure; 3. Every time the agent answers a question, they will first retrieve their previous opinions on similar questions from the current memory, and then answer it.

So as you can see, their world views are controlled to be different at the beginning and will be updated towards different paths. 100 (or more) agents can better model a society where people have diverse initial opinions, and we would love to see with relatively large quantities of AI agents, what would happen if they are allowed to do thorough discussion.

If you replace the grid with a 1x1, it will degenerate to simple distillation settings, which we find is not ideal for alignment since we want diverse and step-by-step improved answers. 1x1 setting will only have one view, and you can hardly cover all possible good answers to a controversial social question.