DeepMind's Dream-RSI Lets AI Agents Dream Up Their Own Improvements
Dream-RSI: Recursive Self-Improvement through Evolving Worlds
Recursive self-improvement hinges on exploration, but fixed strategies can't adapt and online policy optimization is costly. Dream-RSI adds a lightweight orchestration layer that leaves the coding agent untouched: it treats accumulated discovery history as a replay simulator, performing off-policy "dreaming" to refine exploration policies without expensive online evaluations. The improved policy then drives further discovery, expanding the simulator pool in a self-improving loop. Tested on algorithm engineering, mathematical optimization, and GPU kernel engineering, it matches or beats baselines while cutting discovery cost.
Our key insight is that accumulated discovery history can serve as a replay simulator over the realized search space.
- eggbrain
Perhaps I'm not understanding it correctly, but here's my take on what the paper is doing.
Imagine you have a problem you want to solve (let's say, identify an OCR'd handwritten character, e.g. the MNIST Dataset). You tell 3 agents "Hey, each of you take a stab at getting really good at recognizing characters from this dataset. You can take 10 refinement steps to continue to improve ". You can't give each agent unlimited steps of course, because you have a finite amount of compute.
So each agent goes off, and by the end, Agent 1 got to 90% accuracy, Agent 2 got to 80% accuracy, and Agent 3 got to 89% accuracy. Agent 1 wins, of course.
But then you look at the refinement steps, and after 2 steps, Agent 1 was _already at_ 90% accuracy. So the agent spent the next 8 steps basically not moving at all. Agent 3 on the other hand, perhaps was continuously climbing in accuracy at every refinement step, but hit step 10 and had to stop.
Now because you recorded every step from every agent, you know what you'd do differently next time -- you'd not allocate as many steps to Agent 1, and give Agent 3 more steps, because perhaps that might result in Agent 3 coming up with a better answer.
From my understanding, that's what they built in the form of a "search" controller -- a way to evaluate automatically and reapply how you could allocate resources more effectively, when applied to a new problem.
But I guess my misunderstanding is how applicable the search controller is when applied to new […]
- benbenben111
FYI; the paper is clearly a reference to Danijar Hafner's 'Dreamer' line of work, which was published in 2019, and which Danijar has continued to iterate on. https://arxiv.org/abs/1912.01603
The TalkRL podcasts on this line of work are reasonable accessible and quite interesting.
https://www.talkrl.com/episodes/danijar-hafner
https://www.talkrl.com/episodes/danijar-hafner-on-dreamer-v4...
- rybosworld
Unless I'm misunderstanding, calling this RSI seems misleading?
This looks like an optimization of current training methods, and a good one, but not "RSI" in the sense of a system that can perpetually improve itself forever.
- deadbunny
Hapsburgs: Recursive self improvement by inbreeding
- bob1029
There is no way this could be reasonably framed as RSI.
This iterative, online optimization of an exploration policy is not recursively intelligent in any way. It simply reallocates the available computational resources to more promising (hopefully) parts of the search space as system conditions change over time.