PaperlayerView exact revision on arXiv

Source-linked AI summary

A Token Service Interface for AI-Native RANs

Jianan Zhang, Peng Li, Xiang Cheng

arXiv:2610.02618v1cs.NI

TL;DR

AI services can have distinct token-set delivery needs even when average rate and delay targets match, and conventional flow descriptors provide little context about usefulness, timing, or retained execution state. The article proposes TSI to bind service, temporal, and stateful information to schedulable token sets. A drone-inspection case study illustrates how those semantics inform radio allocation, advance preparation, and selective migration.

  • Problem

    AI services may require different delivery treatment for token sets, but conventional flow descriptors provide little context about partial usefulness, predictable emissions, or retained execution state.

    Source: §I · p. 1

  • Method

    The article organizes service, temporal, and stateful semantics and proposes TSI to bind delivery requirements, readiness forecasts, and execution-state references to each schedulable token set.

    Source: §I · p. 1

  • Results

    The drone-inspection case study illustrates that importance weighting changes uplink resource benefits, advance preparation preserves usable transmission time, and selective migration balances forwarding delay against interruption.

    Source: §VI · p. 7

  • Takeaways & Limitations

    The evaluations quantify network utility and cost under explicit assumptions, while application-level benefits require delivery-surrogate calibration against measured task outcomes.

    Source: §VI · p. 7

    Establishing application-level benefits requires calibrating delivery surrogates against measured task outcomes and evaluating metadata overhead, prediction uncertainty, and radio-edge coordination.

    Source: §VI · p. 7

Abstract

from arXiv · show

Generative and embodied AI services exchange token streams within continuing inference and control loops. Their communication requirements depend on each token set's purpose, useful timing, and execution context. This article organizes these properties into service, temporal, and stateful semantics and proposes a token service interface (TSI) between applications, radio access networks (RANs), and edge runtimes. TSI binds delivery requirements, readiness forecasts, and execution-state references to each schedulable token set, specifying field ownership, versioned updates, and admission feedback. A drone inspection case study over a decoupled RAN illustrates importance-aware radio allocation, advance preparation for timely delivery, and selective state migration that balances interruption against forwarding delay.

I. INTRODUCTION

AI services can have different delivery needs even when their average rate and delay targets match, while conventional flow descriptors provide little context about token usefulness, timing, and execution state. The article organizes these needs into three semantic dimensions and proposes an interface connecting applications, RANs, and edge runtimes.

I. INTRODUCTION

  • Conventional flow descriptors provide little context about whether partial delivery is useful, future emissions are predictable, or token sets depend on retained execution state.The article proposes binding delivery requirements, runtime readiness forecasts, and execution-state references to each token set, with field ownership and update rules.Source: §I · p. 1
  • The article organizes communication-relevant properties into service, temporal, and stateful semantics: requested outcome, useful timing, and execution context.These requirements can vary within one session and even when average rate and delay targets match.Source: §I · p. 1
  • The application requests an outcome, while the RAN and edge runtime translate it into protection, scheduling, and placement decisions and return supported service.The article distinguishes measured task outcomes, application-aware delivery surrogates, and network performance metrics.Source: §I · p. 1
  • A drone-inspection case study over a decoupled RAN illustrates how importance, advance timing, and retained state inform allocation, scheduling, and placement.The article notes that decoupling improves the uplink.Source: §I · p. 1

II. COMMUNICATION SEMANTICS OF AI SERVICES

Communication semantics are application properties that change communication decisions. Token-set delivery requirements vary with the requested service, while evaluation distinguishes task outcomes from delivery surrogates and network metrics.

II. COMMUNICATION SEMANTICS OF AI SERVICES

  • A token set groups representation units that share a delivery context, and one set may span several packets while remaining independently schedulable.Selective treatment requires sets to remain identifiable.Source: §II · p. 1

A. Service Semantics

  • Service requirements vary within a VLA session: perception may produce data-heavy bursts, reasoning exchanges can be sequential, and action outputs can be small yet critical.Token sets can have unequal importance; soft sets permit selective transmission when partial evidence remains useful, whereas hard sets require complete, correctly ordered reconstruction.Source: A. Service Semantics · p. 1
  • For an additive utility profile, expected delivered importance weights each token set by its delivery probability, sums across the collection, and normalizes by total weight.Delivery after an applicable deadline does not count.Source: A. Service Semantics · pp. 1–2
  • Evaluation separates measured task outcomes, application-aware delivery surrogates such as importance-weighted timely delivery, and network metrics such as reliability, latency, and cost.A delivery metric remains a surrogate unless calibrated and validated against task outcomes; this separation allows recalibration without redesigning the controller.Source: A. Service Semantics · p. 2

B. Temporal Semantics

Temporal semantics describe when service matters, while stateful semantics identify the execution context to which token sets belong. Forecasts can support preparation for timely delivery, and state placement can balance forwarding delay against migration cost.

B. Temporal Semantics

  • Autoregressive inference has a prefill phase followed by sequential decode; time to first token measures initial responsiveness, while time per output token measures subsequent cadence.Input bursts followed by output trickles create directional demands that average rates miss.Source: B. Temporal Semantics · p. 2
  • Runtime forecasts can predict aperiodic token-set emissions from execution progress, with validity intervals and confidence, enabling preparation for prompt transmission upon arrival.Timely delivery of selected perception sets can be more useful than complete delivery after a control deadline.Source: B. Temporal Semantics · p. 2

C. Stateful Semantics

  • Stateful semantics identify the execution context of a token set; losing access to the serving instance may require state transfer or repeated prefill, stalling inference.Session identity and compute affinity indicate where successive sets should continue.Source: C. Stateful Semantics · p. 2
  • Radio association and execution-state placement can evolve separately, so after handover requests may be forwarded or state migrated when expected forwarding savings justify migration cost.Figure 1 summarizes the three semantic dimensions and evaluation hierarchy.Source: C. Stateful Semantics · p. 2

III. TOKEN SERVICE INTERFACE

The Token Service Interface (TSI) makes communication semantics available to the RAN and edge runtime through a descriptor for each schedulable token set. Field ownership and admission feedback coordinate decisions while leaving resource authority with network controllers.

III. TOKEN SERVICE INTERFACE

  • TSI provides a boundary through which field owners export a Token Service Descriptor for each schedulable token set and controllers return an admission result.The descriptor associates each set with delivery requirements, temporal information, and execution context.Source: §III · p. 2

A. Architectural Principles and Existing Mechanisms

  • TSI exports communication-relevant properties while keeping model internals and task content private, allowing models, tokenizers, and serving pipelines to evolve independently.Source: A. Architectural Principles and… · p. 2
  • Applications, adapters, and runtimes supply their fields; the RAN controls radio admission and allocation, while edge control manages placement and consistent state transfer.Controllers may accept a descriptor, propose a feasible alternative, or reject it using observed network and compute state.Source: A. Architectural Principles and… · p. 2
  • Compared with Release 18 flow-direction PDU-set targets and TSCAI flow-level assistance, TSI binds per-set targets, runtime readiness forecasts, and execution affinity with update rules and admission response.It builds on TokCom’s architectural vision [5].Source: A. Architectural Principles and… · p. 2

B. Token Service Descriptor and Field Ownership

The Token Service Descriptor (TSD) binds metadata to each schedulable token set, combining reusable defaults with per-set overrides. Field ownership separates application requirements from runtime forecasts and state references.

B. Token Service Descriptor and Field Ownership

  • A TSD identifies the session, token set, direction, payload extent, and descriptor version; per-set descriptors reference session defaults and supply overrides.The token-set identifier binds metadata to packetized data.Source: B. Token Service Descriptor and Field… · p. 2
  • The model or task adapter owns importance, while the application owns delivery regime and requested utility or completion reliability.Importance may be carried as a weight or class, or referenced through a versioned reusable table.Source: B. Token Service Descriptor and Field… · p. 2
  • The application owns deadline and cadence; the runtime producing a token set owns revisable readiness forecasts with validity, confidence, and lead time.A reference event and clock domain determine the absolute deadline and residual budget; deadline policies may persist while forecasts are revised for advance preparation.Source: B. Token Service Descriptor and Field… · pp. 2–3
  • The application supplies session identity and continuity requirements, while the runtime supplies the opaque state handle and execution affinity without requiring the RAN to inspect model state.Network measurements provide transport delay and transfer bandwidth; owners update fields when requirements or state change.Source: B. Token Service Descriptor and Field… · p. 3
  • Token sets may contain one or multiple tokens: finer granularity enables selective allocation but increases signaling, so quantization and aggregation trade selectivity against overhead.Token-set decisions must ultimately map to radio transport blocks.Source: B. Token Service Descriptor and Field… · p. 3

C. From Semantics to Decisions

The three semantic dimensions primarily inform protection and allocation, scheduling, and state management, while remaining coupled through shared time and radio budgets. TSI operation combines versioned per-set descriptors with observed network and compute state to return admission decisions.

C. From Semantics to Decisions

  • Service, temporal, and stateful semantics primarily guide protection and allocation, scheduling, and state management, respectively.Examples include utility-aware allocation, deadline ordering or predictive reservation, and forwarding or migration under mobility.Source: C. From Semantics to Decisions · p. 3
  • These decisions remain coupled: stronger protection takes time, and migration can reduce the budget available to an arriving token set.With a fixed radio budget, a robustness view maximizes normalized expected utility; with a fixed utility target, an efficiency view minimizes radio resources.Source: C. From Semantics to Decisions · p. 3
  • At session setup, the service registers TSD defaults and a state handle; for each forthcoming set, owners initialize its descriptor and the runtime may update its readiness forecast.On packet arrival, controllers combine the latest descriptor with observed network and compute state for admission and coordinated resource control.Source: C. From Semantics to Decisions · p. 3
  • Owners issue versioned updates when requirements, placement, retained state, or forecasts change; newer metadata supersedes older versions, expired forecasts are ignored, and missing optional fields use agreed defaults..Source: C. From Semantics to Decisions · p. 3

A. Architecture and Service Profiles

The inspection-drone illustration separates uplink perception reception, downlink action transmission, and edge execution under decoupled access. Its setup evaluates the semantic dimensions separately using a fixed radio snapshot, a derived patch-importance profile, and a prescribed mobility trajectory.

A. Architecture and Service Profiles

  • The drone uses a nearby uplink base station as perception receiver and a macro downlink base station to transmit actions under decoupled access.Service and temporal semantics use a fixed radio snapshot; stateful semantics use a prescribed flight trajectory, and joint control is outside the illustration’s scope.Source: A. Architecture and Service Profiles · pp. 3–4
  • The fixed snapshot places the macro and nearby uplink base stations 250 m and 60 m from the drone, respectively.The assumed transmit power is 23 dBm, noise power is −95 dBm, and path loss is 35 + 35 log10(d/1 m) dB.Source: A. Architecture and Service Profiles · p. 4
  • The importance profile derives from 5,000 COCO 2017 validation images resized into 196 patches per image, with spectral-residual saliency averaged, clipped, and mapped to [0.05, 1].The top 20% of weights carry 53.4% of importance.Source: A. Architecture and Service Profiles · p. 4

B. Resource Savings with Service Semantics

Importance-aware allocation reduces uplink resource cost relative to equal-importance allocation, with the size of the savings varying by utility target and radio conditions. The evaluation measures expected weighted delivery utility and excludes several sources of implementation overhead.

B. Resource Savings with Service Semantics

  • The target q is the minimum normalized expected utility, defined as expected delivered weight divided by total original weight, with each profile normalized using its own weights.For example, q = 0.75 requires delivery of at least 75% of total importance on average.Source: B. Resource Savings with Service… · p. 4
  • The solver minimizes uplink resource cost while meeting the expected weighted target under delivery probabilities modeled for independent Rayleigh blocks.Finite-blocklength effects, retransmissions, and signaling overhead are excluded; the common macro downlink is fixed.Source: B. Resource Savings with Service… · p. 4
  • 27.1% less uplink resource under coupled access and 41.3% less under decoupled access are required by the saliency-weighted profile than the equal-importance profile at q = 0.75.The coupled-to-decoupled cost ratios are 17.1 and 21.2, respectively, making the ratio 24.3% higher with saliency weighting.Source: B. Resource Savings with Service… · p. 4
  • 59% at q = 0.5 to 12% at q = 0.95 is the decrease in relative cost difference between profiles under decoupled access as the utility target rises.Higher targets leave less scope to omit low-value token sets, so savings depend on both target and radio conditions.Source: B. Resource Savings with Service… · pp. 4–5

C. Performance Improvements with Temporal and Stateful Semantics

Temporal semantics show how readiness forecasts affect timely delivery, while stateful semantics show how selective migration balances forwarding delay against interruption. The evaluations illustrate these effects under fixed delivery conditions and a common drone trace.

Temporal semantics

  • With sufficient forecast lead time, grant and resource preparation can finish before packet arrival and remain valid for transmission.Without a forecast, preparation consumes part of the 50-ms arrival-to-deadline window, reducing usable transmission resources.Source: C. Performance Improvements with… · p. 5
  • At q = 0.75 and η = 0.3, utility is 0.6364, a 15.1% relative loss without advance timing.The loss widens as preparation consumes more of the delivery window, even when allocation is reoptimized.Source: C. Performance Improvements with… · p. 5, Figure 5 · p. 6

Stateful semantics

  • The stateful comparison uses fixed anchoring, migration at every handover, and an offline oracle on a common drone trace.The oracle knows future requests and handovers and minimizes response latency within the retain-or-migrate decision model.Source: C. Performance Improvements with… · p. 5
  • At 8,192 context tokens, the oracle reduces mean additional latency by 26.4% versus fixed anchoring and 56.8% versus every-handover migration, with 448 MiB migrated versus 2,240 MiB.It reduces migration-induced execution pauses from five to one; all policies preserve state and complete every request.Source: C. Performance Improvements with… · p. 6, Figure 5 · p. 6

V. OPEN CHALLENGES

The open challenges concern calibrating service semantics, accounting for signaling and processing overhead, coordinating state under mobility, protecting shared interfaces, and standardizing field meanings and ownership.

Calibrating service semantics

  • Saliency and linguistic roles are proxies; token-set contributions can be assessed by deletion or perturbation and downstream-model reruns.Delivery targets can then be calibrated against task outcomes.Source: §V · pp. 6–7

Scaling temporal semantics

  • Finer descriptors and advance preparation must provide benefits that justify their signaling and processing overhead.Metadata, channel acquisition, clock alignment, scheduling, and retransmission costs shape useful token-set granularity and prediction horizons.Source: §V · p. 7

Maintaining state under mobility

  • Radio handover can outpace migration, so coordinated decisions require state size, transport conditions, compute queues, residence predictions, and consistent runtime ownership.Radio and edge controllers manage handover and migration separately.Source: §V · p. 7

Deploying a shared interface

  • Admission limits, authentication, and access control must protect resources, metadata, and state handles from inflated importance or urgency.Multi-agent deployments also add interference and congestion; privacy requires useful service classes without revealing task content.Source: §V · p. 7

Standardizing the boundary

  • AI interoperability requires common meanings and ownership for delivery, timing, and continuity fields, plus radio and edge protocol bindings.These bindings should permit control mechanisms to evolve; the proposal builds on Release 18 per-set QoS handling.Source: §V · p. 7

VI. CONCLUSION

TSI couples per-token-set delivery requirements, temporal information, and execution context to admission and network decisions. The case study illustrates importance-aware allocation, advance preparation, and selective migration trade-offs.

VI. CONCLUSION

  • TSI links each token set’s delivery requirements, temporal information, and execution context to coordinated RAN and edge admission decisions.Field owners initialize and update the descriptor; coordinated controllers return an admission result while retaining network authority over resource decisions.Source: §VI · p. 7
  • The case study illustrates that importance weighting changes decoupling’s resource benefit, advance preparation preserves usable transmission time, and selective migration balances forwarding delay against interruption.Source: §VI · p. 7
  • Application-level benefits still require calibration against measured task outcomes and evaluation of metadata overhead, prediction uncertainty, and radio-edge coordination.Source: §VI · p. 7

Related papers

Loading 2610.02618v1…