TIPEX finds AI agent parallelism peaks in the middle

TIPEX splits AI agent parallelism into two mechanisms and unifies them under one execution semantics, according to a paper accepted at ICML 2026.

The split matters for anyone setting fan-out width in a production system. Two different levers, not one.

Replica Parallelism runs multiple complete solution attempts side by side, then picks the best. Structural Parallelism keeps one attempt running and breaks the work into pieces that run at the same time.

On the GAIA benchmark, adding parallelism raised accuracy and cut latency, at a higher token cost. Tasks of intermediate difficulty gained the most. Pushing parallelism harder on those same tasks, or applying it to tasks that were already easy or already too hard, didn't reliably perform better.

Fan-out width isn't a dial you can just turn up. The paper puts the biggest returns in the middle of the difficulty curve, not at the top. Teams that scale parallelism by default instead of by task difficulty pay more in tokens for accuracy gains that stop improving before the fan-out does.

ICML's 2026 acceptance makes this the newest measured account of that tradeoff.

The paper doesn't say what the cutoff difficulty looks like outside GAIA, so translating this into your own task distribution takes some guessing.