NIOCIS Routing Layer: Typed Decisions and Node-Conditioned Option Spaces for Tool-Use Agents
ORIGINAL / NIOCIS: Typed-Decision Next-Hop Routing with Node-Conditioned Option Spaces for Tool-Use Agents
NIOCIS compresses tool-use agent logs into a loop-collapsed action graph and uses a small model to emit a typed option distribution in one pass, replacing costly per-hop LLM reasoning with very low latency.
01 ABSTRACT
NIOCIS is a routing layer that folds execution logs into a loop-collapsed action graph and uses a small model to output a calibrated distribution over typed options in a single pass, without generating text. On 1,615 production security-operations episodes, its encoder matched 32.2% of held-out next-hop decisions in 10 ms, versus 2.2 s per LLM hop. The authors argue the typed layer is chosen for cost, not a separated performance lead.
02 KEY FINDINGS
- On 3,875 held-out decisions, the typed encoder matched the logged next hop 32.2% of the time in 10 ms, while each LLM hop takes 2.2 s.
- Fine-tuned jpt-4b reaches 31.9% accuracy versus 9.9% zero-shot, with a paired test showing no significant difference (p=0.68).
- Option-space ablation attributes the gain to per-node candidate restriction (+3.5 points over a flat catalog) rather than graph traversal.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG