The Weight of Teamwork: Does Communication Structure Help LLM Scientists Search for Scientific Laws?

We study how communication structure affects the search conducted by LLM agents, using symbolic regression as a controlled instance of hypothesis proposal and exchange. At 50 calls per problem, we compare policies on 100 Qwen problems and a 41-problem Terra subset. In exploratory Qwen comparisons under a common selector, Ring reduced mean held-out fitting score Q by 0.3276 (95% CI [-0.4889, -0.1836]). Star and Complete had mean differences +0.0265 and +0.0958, with intervals spanning zero ([-0.0820, +0.1321] and [-0.0810, +0.3373]). As an allocation reference, ten independent trajectories improved mean Q over one by 0.6960 ([0.4105, 1.0338]), with median gain 0.1060. Fixed-archive delivery interactions passed Holm-adjusted signed-rank tests in 9/20 Qwen comparisons and 0/20 Terra comparisons, whose estimates were imprecise. The policies combine visibility with update timing, and the focal delivery comparisons are exploratory. The results support examining information access and candidate selection separately when interpreting team performance; they do not identify a communication mechanism.