Communication Topology and Objective Alignment in Decentralized Language-Model Fine-Tuning

Decentralized language-model fine-tuning requires each receiver to select useful senders without access to their private text. We present a matched-budget evaluation protocol that separates three questions: whether a communication policy changes sender selection, improves model quality at an equal byte budget, and provides a benefit above its execution cost. We apply the protocol to numerical simulations with up to 1,024 learners, sealed Qwen3-0.6B experiments, and 48 runs on four L4 GPU hosts. Adaptive routing increased same-source transfers from 12.8% to 27.9%, but reduced quality-per-byte AUC in every physical comparison. A prospective follow-up that evaluated candidate merges before selection also increased final global NLL by 0.00416 and reduced AUC by 0.00181 relative to a data-independent rotating policy. Dynamic policies achieved only 0.902-0.907 of the reference applied-transfer rate. The routing signal therefore changed communication as intended, while its receiver-local objective conflicted with the network-wide endpoint. These results identify objective alignment as a systems requirement for source-sensitive routing.