https://arxiv.org/html/2609.14858v1
Instead of simply asking an AI agent to solve a problem over and over, Dream-RSI lets the system learn from its previous attempts and improve the strategy it uses to search for better solutions.
It does this by saving its history of discoveries and using that history as a kind of simulator. The AI can then cheaply test different approaches against what it has already learned before spending resources trying them in the real environment.
explore -> learn from the exploration -> improve the exploration strategy -> explore again
The researchers tested this on algorithm design, mathematical optimization, and GPU kernel engineering. They report competitive or improved results while substantially reducing the cost of discovery in several settings.
If AI can improve how it conducts research, it can potentially discover better algorithms, write better kernels, optimize infrastructure, reduce experimentation costs, and make the next round of AI research faster.