I came across this paper introducing a Station environment, where different AI models work together on maths problems. The agents can write papers for other agents to read, and these stay around after the agents themselves are replaced.
There is a collection of their papers online. Some seem to build on previous papers, with citations between them. The study claims this led to new mathematical discoveries.
An agent could spend its whole lifetime on an idea without finishing it, yet leave enough for another agent to continue. Progress would then depend partly on what previous agents had worked out, as well as on the ability of the current models. The same models could have a better starting point later in the run simply because there was more useful work to build on.
That makes me wonder how much the agents actually learn from these papers. Can they pick up a new method developed by another agent, or do they mostly fall back on what they already knew from training?
This seems connected to in-context learning. The models are not being retrained as the literature grows; what changes is the work available for them to read. Could that eventually let the same models tackle problems they struggled with at the start?
I came across this paper introducing a Station environment, where different AI models work together on maths problems. The agents can write papers for other agents to read, and these stay around after the agents themselves are replaced.
There is a collection of their papers online. Some seem to build on previous papers, with citations between them. The study claims this led to new mathematical discoveries.
An agent could spend its whole lifetime on an idea without finishing it, yet leave enough for another agent to continue. Progress would then depend partly on what previous agents had worked out, as well as on the ability of the current models. The same models could have a better starting point later in the run simply because there was more useful work to build on.
That makes me wonder how much the agents actually learn from these papers. Can they pick up a new method developed by another agent, or do they mostly fall back on what they already knew from training?
This seems connected to in-context learning. The models are not being retrained as the literature grows; what changes is the work available for them to read. Could that eventually let the same models tackle problems they struggled with at the start?
Paper: https://arxiv.org/abs/2608.23691
The agents’ papers: https://dualverse-ai.github.io/station_data_v2/#/finite_kakeya/archive