An arXiv preprint posted Oct. 20 reports that communication structure alone does not determine the ranking of fixed-budget large language model teams searching for scientific laws, according to the arXiv CS.LG listing.
The finding matters for compute budgeting and benchmark design. The paper argues team performance is often reported as one delivered score even though it is produced in two stages: candidate generation and final selection. In fixed-budget experiments, the largest measured gain came from distributing the call budget across more independent trajectories, not from peer-communication topology.
The study used symbolic regression as an observable scientific-search setting and compared LLM teams that differed in peer visibility and update timing. It evaluated frozen candidate archives and the expressions selected from them. Distributing the call budget across more independent trajectories improved mean held-out fitting score Q by 0.70 at the widest tested allocation. Under a common archive-wide selector, the implemented Ring policy reduced mean Q by 0.33 relative to Independent, while Star and Complete remained unresolved. Holding generated archives fixed, changing the delivery rule changed the measured architecture gaps, including the sign of the Complete point estimate.
The authors say the results support reporting candidate generation and final selection as separate endpoints. They do not identify a communication mechanism or a universally preferred topology. The arXiv comments say the paper was accepted at the NeurIPS 2026 Workshop on Foundations of Agentic Systems Theory (FAST), non-archival, with 7 pages of main text and 33 pages with appendices. The posting is a preprint and has not been peer reviewed. The abstract does not name the LLMs tested.