Meta AI (FAIR) latest paper integrates system-1 and system-2 thinking into reasoning models.

Basically, it introduces the term "Dualformer" which integrates both system-1 (fast-thinking) and system-2 (slow-thinking) into the transformer to improve its reasoning capability. The high level idea is to train the model with "randomized trace", which randomly drop parts of the reasoning tokens. This approach improves model's inference speed, accuracy, and diversity. It also enables model to perform system-1 and system-2 thinking in a controllable fashion.

I think this paper is interesting because it integrates human level intelligence into the transformer, a model capable of perfoming system-1 and system-2 level thinking. 

The paper's link here:

https://arxiv.org/abs/2410.09918v1