Feed
Dream-RSI: Recursive Self-Improvement through Evolving Worlds
An agent learns to search better by replaying its own past search trees.
162×
fewer agent calls
Agent calls saved on Lasso
Faster Lasso solver found with a tiny fraction of the search budget
What changes
Wanders widely, burns calls, solution arrives late
Direct route, few calls, solution arrives sooner
How it works
5 stepsConvDiv task: 3D convolution on a 128×8×16×64×64 input, divide by 2, max pool, average pool, add bias, sum over channels. Minimize GPU runtime and stay numerically correct.
Online exploration
The current policy sets branching, parallelism and stopping. The unchanged base agent writes and scores candidates, building a tree.
- Base Discovery Agent
- Online Exploration
The example here
Round E0 on ConvDiv: the starting policy runs up to 10 workers and 110 attempts. Each writes a kernel, scored in 1/ms. Best is 0.427.
From the paper