NIOCIS picks tool-use agent next hops without generating text
Original title: NIOCIS: Typed-Decision Next-Hop Routing with Node-Conditioned Option Spaces for Tool-Use Agents
On 3,875 held-out security-operations decisions, a typed encoder matched the logged hop 32.2% of the time in about 10 ms.
NIOCIS is a routing layer that turns execution logs into a loop-collapsed action graph and answers which declared option, plus stop, a fleet takes next. A small model emits a calibrated distribution in one pass and no text. On 1,615 production security-operations episodes, the typed encoder matched 32.2% of 3,875 held-out decisions, or 34.9% within declared options, versus 2.2 seconds per LLM hop. The authors choose it for cost: paired tests do not separate it from feature readouts, and a cross-trace ceiling is 25.6–44.9%.
Unconfirmed: Paired tests do not resolve whether the typed encoder beats the feature readouts or whether fine-tuning beats the reported zero-shot gap.
- jpt-4b scored 9.9% zero-shot and 31.9% fine-tuned, a gap with p=0.68, at 244 ms.
- The realised hop fell outside declared options for 7.9% of decisions and outside the top twenty for 9.3%.
- The most confident 15% of predictions held 91.9% accuracy.
- Per-node candidate restriction added 3.5 points; graph traversal added none within its interval.
From the original article
1 min read at arxiv.orgTool-use agents pay per decision: at each hop a language model re-reads the state and thinks for seconds, though its actions are finite and decisions recur. We present NIOCIS, a routing layer that folds execution logs into a loop-collapsed action graph (forward transitions plus recovery back-edges) answering one typed question: which declared option at a state, plus stop, does the fleet take next?
By Jin Han · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.