AI schedulers shift agent evidence choices by 24 percentage points

Researchers Ruiyang Ren, Yuhao Wang, Yinan Zhang and Chunyan Miao said AI search agents’ result timing changed evidence acquisition by up to 24 percentage points in an arXiv preprint submitted Oct. 19.

The Oct. 20 arXiv posting matters because agent systems increasingly gather information in parallel, and the authors said their method improved accuracy over early commitment by as much as 12.6 percentage points.

The tests replayed agent runs using two families of models on three benchmarks, according to the preprint. Acquisition disagreement refers to divergence in what evidence an agent chooses to seek or use, compared with variation seen when the same order is repeated.

The 24-percentage-point gap was measured above same-order repeat variation, giving the paper its scale comparison. The authors said Confluent Evidence Acquisition later brought that disagreement to within 1.0 percentage point of the repeat-variation baseline.

The researchers focus on schedulers, the system components that determine when search results become available to an agent. Their finding is that timing can shape later searches even when the starting evidence is the same.

That effect did not end when the first batch of evidence was complete, the authors said. It continued into a second search, meaning a temporary difference in what arrived first could influence what the agent later acquired.

The paper introduces Confluent Evidence Acquisition, or CEA, a scheduling approach that waits for each subquestion’s evidence to be complete before releasing local reasoning. It also allocates shared search capacity independently of which subtask finishes first.

Under fixed plans and what the authors call a stable conditional response law, they said they proved invariance across decisions, acquisitions and answers. With question-only plans on three benchmarks, CEA produced accuracy comparable to complete-input controls, according to the preprint.

The preprint also tested a weaker version that allocated resources to the first ready task when there was contention. That change added 21.6 percentage points of acquisition disagreement, while local query-target disagreement stayed close to repeat variation, according to the authors.

The limits are material. The work is an arXiv preprint in artificial intelligence, computation and language, and information retrieval, and has not been peer reviewed. The formal result also depends on a fixed plan and a stable response rule.

For AI developers, the paper points to a reproducibility issue in agents that combine language models with search or retrieval. The authors also said timed comparisons showed CEA answered faster than the same program waiting for the full initial evidence set.

Source: arXiv preprint, arXiv cs.AI, submitted Oct. 19, 2026; listed Oct. 20, 2026.