A new arXiv preprint asks whether the way LLM scientists are connected changes how well they search for scientific laws. The study uses symbolic regression as a concrete testbed, letting the authors observe the process rather than just the final output.
The authors point out that team performance is often reported as a single score, but that score actually comes from two stages: a system generates candidate answers, and a selection rule picks one. This distinction matters for evaluating whether communication structure helps.
The abstract does not report specific results, so the main contribution at this stage is the framing: separating generation from selection when studying LLM teams. The paper is available on arXiv under identifier 2610.24131.