counterfactual optimization

Counterfactual optimization lets an AI agent roll back to a certain point while it's working on a task, run several different choices from that point at the same time, and then keep whichever branch turned out best. It's similar to saving your progress in a game at a key moment, replaying from that save several different ways, and continuing with whichever run went best.

Normally, if an AI agent needs to undo work it has already done, there are only two options: start over from scratch, or patch the mistake on top of what's already there. Both are slow and costly, and it's hard to recreate the exact same situation again. Counterfactual optimization became possible once agents could clone their execution state mid-run and branch it out. The key idea is that instead of committing to one answer in advance, the agent actually runs multiple possibilities and then compares them to choose.

This approach is also used when training agents themselves. At certain decision points, the agent's execution is branched into several paths, and whichever paths performed well get fed back into training.

The article reports that "branch-search-based counterfactual optimization delivered up to 11 points higher performance than the baseline across four benchmarks while cutting execution time by up to 58%." A common misunderstanding is that this isn't a feature that fixes an agent's mistakes after the fact — it's a search method that branches execution into multiple paths in advance and compares them.