Deep reinforcement learning is currently the dominant paradigm for pursuing AGI. Thinking about whether this will remain the case has some important implications, as there is major controversy over whether this particular framework for building AI -- opaque, agenty, neural network-like systems trained by gradient descent on an objective -- can be safely aligned with human values at all (c.f. Paul Christiano on Prosaic AI alignment, a reply from Eliezer Yudkowsky, and Eric Drexler on the Comprehensive AI Services Model).
The final metric will be determined using the arXiv advanced search feature for papers in the computer science OR statistics categories (including cross-listed papers) with terms "RL" OR "reinforcement learning" in the title (using quotes to search for exact matches), between January 1 and December 31 2025.
Previous values of this metric since 1991, as well as counts for the phrase "reinforcement learning" in abstracts, can be found in this spreadsheet.
Most recent values:
2019 (thus far): 2.1%