Skip to content
InfoResearchPreprintLLM-specific

When Do Decision Models Help LLM-Driven Network Discovery?

Published
Record updated
View JSON

Summary

This paper measures whether inserting fast, non-generative decision models to validate LLM-proposed actions improves network discovery across networks of 10, 100 and 1,000 nodes, using a simulation calibrated on a 30-node containerized testbed. Chunked batch dispatch recovers 92% to 100% of open ports at large scale, while single-host dispatch falls under 6% because of context saturation and scan latency. At the local model's measured proposal quality, decision gating raises port recall from 75.5% to 89.0% on small networks and intercepts decoy tarpits, but it adds latency for frontier models during early host discovery.