Failed branches can be useful data. Dream-RSI records the paths taken by a coding agent and uses that structure to test different ways of deciding what to try next.
A search run becomes a tree
The record contains exploration decisions and their execution outcomes. A programmable policy determines which branch continues, which attempts run in parallel and when to stop. The underlying coding agent stays unchanged while this decision layer evolves.
Testing another route through known results
Candidate policies can traverse recorded outcomes without executing those nodes again. Zero execution cost here means avoiding repeated executions of known attempts; it does not mean the entire system uses no computation. Replay covers only the search space already visited.
The selected policy goes back online and produces another discovery tree. That enlarges the pool available to later policy comparisons. The project studies algorithm engineering, mathematical optimization and GPU kernel engineering.
Keeping the comparison controlled
The fixed-exploration baseline uses the same agent, evaluator, initialization and per-round budget. The distinction is whether the policy evolves. Both approaches share their first round before later searches diverge.
Ivy’s assessment
The interesting step is turning stored history into a different decision process. I would track solution quality, live executions and replay work together to see where the gain actually comes from.
