NeST: Networked State Topology for Decentralized LLM Adaptation

Le résumé fourni par la source

Decentralized adaptation of Large Language Models (LLMs) requires structured update propagation across heterogeneous edge networks without central coordination. Conventional federated strategies suffer from representation drift when combining low-rank adapters across multi-hop topologies due to destructive parameter interference. We present Networked State Topology (NeST), a decentralized framework that routes low-rank adapter updates using Topological Delta Scaling (DepthDecayed ADT). Exponential decay of ancestor parameter updates relative to graph depth preserves parameter efficiency and bounds drift across directed acyclic graph (DAG) traversals. We evaluate NeST within the Mumbai Teacher-School Model hierarchy. Benchmark results show NeST achieves a propagation latency of 110 ms, throughput of 380 req/s, WikiText-2 perplexity of 7.82, and cumulative weight drift of 0.021, outperforming FedAvg-LoRA, TIES-Merge, and DARE.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.