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.
Does RSI actually mean anything specific anymore? RSI, AGI, at this point seem like buzzwords. Sure AGI has definition that are measurable, say "better than 95% of humans on 95% of intellectual tasks" but if we used that definition we already have AGI and almost no one thinks we have achieved AGI. We use AIs to train AIs which we use to train AIs, why is that not RSI? How much human intervention means that is not RSI?
The replay simulator from history for off-policy eval is clever - avoids expensive rollouts. Curious how they prevent the policy from overfitting to already-discovered branches and going stale as the search space expands?
I ended up building a simplified version of this as /self-improve in https://github.com/DanMcInerney/orchflows. History is the state ledger, memory and RSI just cite the history as evidence and can be rewritten. I feel like strong immutable state is the missing piece of the puzzle for most of these memory libraries.
I might be wrong but is this really a solution for RSI? I interpret it's as a way to reducing wasted tokens and compute on paths that don't yield better results. It's an optimization. It's a faster way to get to RSI though. what's wrong?
"Across algorithm engineering, mathematical optimization, and GPU kernel engineering, Dream-RSI achieves competitive or improved discovery quality while substantially reducing discovery cost in several settings."
Fairly certain all the labs are doing this (RSI) at this point. It's a question of how public their proclamations are about it and how they're positioning PR etc.
Even today's lighter weight models know how to write kernels and optimize them. I've had DeepSeek 4.1 Flash tune the crap out custom CUDA kernels on my own codebase and it was entirely competent at it. And cheap.
The innovation pieces will be in the harnesses to support this. Which I guess is partially what's going on here.
It's not really RSI if you are just using the AI as a tool to help make it better. It has to be doing it itself, no? Otherwise self-hosted compilers are RSI.
This is a solid and very interesting paper! The authors were kind enough to publish the complete prompt for it (appendix B.1 on page 18), so anyone can try their approach with any LLM and see the results.
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.
I'd say it's much closer to the concept of continual learning, but I'm only a few pages deep and haven't groqued it fully yet.
"Across algorithm engineering, mathematical optimization, and GPU kernel engineering, Dream-RSI achieves competitive or improved discovery quality while substantially reducing discovery cost in several settings."
This is like asking why a paper called “I discovered God. This is the specific proof.” is revolutionary. Just tedious levels of incredulity.
Even today's lighter weight models know how to write kernels and optimize them. I've had DeepSeek 4.1 Flash tune the crap out custom CUDA kernels on my own codebase and it was entirely competent at it. And cheap.
The innovation pieces will be in the harnesses to support this. Which I guess is partially what's going on here.