6G slice placement stops relying on one fixed plan: ORION feeds verifier-checked outcomes back to its LLM planner
Synopsis
Placing, routing and admitting multi-domain 6G network slices is an NP-hard online problem, fixed planning strategies reverse rank across topologies, and most rejected requests do have a feasible placement. ORION's hierarchical orchestrator uses two LLM agents to translate service intents and generate abstract placement plans, and records each attempt's network conditions, plan shape and verifier feasibility outcome in an episodic memory that returns to the planner as labelled examples. With selection and execution held fixed on held-out topology families, the memory-equipped planner admits more than both the memory-off and first-in-first-out variants on two seeds, though the magnitude is not established.
Interpretation
The bottleneck in cross-domain slice orchestration sits at the plan layer: deciding which domain hosts each virtual network function and where the chain is cut affects admission more than improving the downstream partition selector. Multi-domain slice placement has largely been treated as an optimization problem, where exact formulations are intractable online and heuristics are fast but encode a fixed strategy that stops adapting when the substrate, workload mix or load regime changes. On a five-domain, tier-scarce substrate, every request rejected by first-fit-decreasing had a feasible placement according to an exhaustive feasibility verifier, and 84.7% of those were search failures in which the heuristic missed an existing placement.
ORION places the LLM planner inside a three-layer structure where it plans only over per-domain aggregate information, while concrete node placement is left to a learned selection layer and per-domain execution layer. LLMs have mostly served as intent translators or as advisors held fixed after deployment, and the few approaches that adapt online use only scalar reward feedback on their own proposals. The planner never sees per-node state because domain residuals are summed over nodes, hiding single-node fitness; the selection layer additionally observes the maximum single-node headroom per domain and routes cross-domain flows on the full per-node graph at commit time.
The episodic memory retrieves by topology signature and injects verified successes and violation-tagged counter-examples into the planner's prompt, so planning is conditioned on plan shapes that actually survived on similar substrates rather than on the model's own guesses. The first-in-first-out store shares the retrieval key, tag schema and prompt format with the full memory; the two differ only in the write rule, the eviction rule and recency weighting. On four topology families never seen in training with selection and execution held fixed, the full memory version admits 4.6 and 11.5 percentage points more than the memory-off version on seeds 0 and 1, and 11.1 and 12.3 points more than the first-in-first-out version.
Perspective
The demonstrated scope is a plan-layer isolation on parametrized multi-domain topology families: selection and execution are held fixed at a deterministic prior-following path, every arm commits the plan's suggested partition unchanged, and evaluation uses four topology families never seen in training across two seeds. Full-stack evaluation, the reward-labelled memory control and the routing-critical boundary probe follow a pre-specified protocol and are reserved for the extended version.
The direction of the admission gain is consistent across both seeds, but the magnitude is not established; full-stack evaluation, the control that labels memory by scalar admission reward, and the boundary probe where the memory's advantage may invert on routing-critical families are all reserved for future work under a pre-specified protocol.