Researchers proposed Wiki Foundation Model for agent-native knowledge representation. It bridges fine-grained structures with dense contexts while keeping explicit topologies. A query-conditioned attentive aggregation enables rich wiki message passing. An NCCL boundary exchange protocol reduces distributed training overheads. The model showed strong performance on long-term memory and multi-hop reasoning tasks. It achieved a 10.5 times training speedup on distributed clusters. The paper was submitted to arXiv on 16 September 2026.

• 10.5x training acceleration on distributed clusters
• 5 long-term agent memory and multi-hop reasoning benchmarks