Nick Byrd, Ph.D.#Meta researchers find more support for #BoundedReflectivism: http://dx.doi.org/10.1111/meta.12534
A "slow" transformer model (reasoning chain + solution) outperforms a "fast" (solution only) model, but a dual-minded model that automatically switches between fast and slow modes performs as well as the slow mode, but with fewer reasoning steps — i.e., reflective reasoning was beneficial, but sometimes not worth it: https://doi.org/10.48550/arXiv.2410.09918
