PaperlayerView exact revision on arXiv

Source-linked AI summary

SatOps-Decision: Low-Latency Discriminative Decisions for Onboard Satellite Communication Operations

Lifan Mei, Yuxuan Li, Jiaxian Tu, Shenghe Xu

arXiv:2610.38697v1cs.LGcs.AIcs.NI

TL;DR

Satellite operations require timely decisions from incomplete or aging evidence under contact and resource constraints, while supplied-answer generation adds a decoding stage. SatOps-Decision adapts a local Laya scorer and evaluates it with an auditable operational benchmark; it achieves higher raw accuracy than frozen Laya while retaining lower median request latency than local generative comparators.

Abstract

from arXiv · show

We propose SatOps-Decision, a method for low-latency decisions in satellite communication operations. To the best of our knowledge, it is the first application of Jev-style general-purpose discriminative decision models to satellite operations, and the accompanying benchmark is the first dedicated to this approach. The learned scorer, a locally adapted Laya encoder with its native decision head, selects among supplied operational answers in a single forward pass. The benchmark combines authored operational contracts, controlled state changes and independently checked labels in an auditable construction. Evaluation prioritizes complete-request latency and raw decision accuracy; all main accuracy comparisons use the same 1,024 held-out questions, separate from training and development. SatOps-Decision's local scorer has a median latency of 90.3 ms per decision, versus 493.9 ms for Qwen3.5-0.8B generation and 388.4 ms for OneJev1.6-0.8B: $5.47\times$ and $4.30\times$ speedups based on median request times in the measured FP32/MPS implementations. Its mean raw accuracy is 85.3%, versus 42.5% for frozen Laya, 40.7% for Qwen generation and 58.6% for OneJev. As cloud accuracy references, original Jev 1.14/1.13 reach 87.4/81.9%.

I. INTRODUCTION

Satellite communication operations require timely judgments from incomplete or aging evidence under limited contact time and onboard resources. SatOps-Decision combines direct candidate scoring with separate operational checks and introduces a dedicated benchmark for these decisions.

I. INTRODUCTION

  • Satellite communication judgments must account for incomplete or aging evidence, available contact time, and limited onboard resources.A backup path may be unusable if energy, readiness, or completion time is insufficient; recovery must be confirmed from later observations.Source: §I · p. 1
  • The method uses a locally adapted Laya scorer to select among supplied operational answers, with deterministic checks enforcing execution prerequisites.The scorer handles evidence assessment, action selection, switching permission, and recovery verification.Source: §I · p. 1
  • Direct scoring avoids answer decoding for small answer sets, motivating evaluation of whether domain adaptation improves accuracy while retaining a latency advantage.The comparison includes frozen Laya, local generative Qwen and OneJev, and cloud Jev references.Source: §I · p. 1
  • The benchmark uses explicit operational contracts, controlled changes in evidence, contact, and resources, and independently checked targets.All models receive the same policy, state, and candidate meanings.Source: §I · p. 1
  • To the authors’ knowledge, this is the first Jev-style application to satellite operations and the first dedicated benchmark for this approach.The benchmark has separate training, development, and test data.Source: §I · pp. 1–2

II. BACKGROUND AND SCOPE

Jev-style models score supplied candidate meanings for typed decisions, while different backbones can implement the interface. SatOps-Decision’s original Jev API comparisons provide accuracy references rather than local evaluations.

II. BACKGROUND AND SCOPE

  • Jev returns typed decisions and probabilities through a supplied-state interface, with Choice, Noul, and Score representing categorical, binary, and ordered judgments.Candidate descriptions specify answer meanings, allowing one model to serve multiple decisions without free-form generation.Source: §II · p. 2
  • Laya combines a pretrained ModernBERT-large encoder with a native decision head, and its components are adapted to satellite operational decisions.OneJev1.6-0.8B derives from Qwen3.5-0.8B, the generative LLM control.Source: §II · p. 2
  • Original TypeSafe Jev 1.13/1.14 API versions provide accuracy references only because their weights are unavailable for local evaluation.The local model comparisons include other implementations and backbones.Source: §II · p. 2
  • Prior work spans spacecraft language agents, satellite diagnosis and recovery, communication control, and wireless decisions; SatOps-Decision adds a domain-adapted method and benchmark for satellite operations.The cited DoomSat and Odyssey demonstrations concern game or flight-navigation choices.Source: §II · p. 2
  • The benchmark applies counterfactual examples and behavioral testing to contact-, resource-, and evidence-conditioned operations.These principles and controlled decision groups with independent solvers are established in prior work.Source: §II · p. 2

III. SATOPS-DECISION METHOD

SatOps-Decision supplies policy, timestamped evidence, a question, and candidate meanings to a domain-adapted Laya scorer, then keeps operational execution checks separate. The scorer uses a single forward pass to produce candidate scores without autoregressive decoding.

III. SATOPS-DECISION METHOD

  • The interface supplies policy, timestamped evidence, a question, and candidate meanings while excluding labels and evaluation metadata.The operational checker is separate from the learned scorer.Source: §III · p. 2

A. Local adaptation and direct answer scoring

  • All four operational judgments use Choice scoring, with Laya’s native decision head producing option logits.The model is based on a pinned 421M-parameter Laya English checkpoint.Source: A. Local adaptation and direct answer… · pp. 2–3
  • A single forward pass returns candidate scores without autoregressive decoding.Padded markers are masked, and the model is fine-tuned using masked option cross-entropy.Source: A. Local adaptation and direct answer… · p. 3
  • Training uses 1,536 questions, four epochs, and seeds 7, 17, and 29; checkpoints are selected by raw task-macro accuracy on 256 development questions.All three predesignated seeds are reported, with checkpoint selection completed before test evaluation.Source: A. Local adaptation and direct answer… · p. 3
  • Runs use 384 AdamW updates with effective batch 16, FP32 training, and token budgets of 1,024 total and 256 for questions and options.Truncation audits preserve policy and answer meanings; reloads must reproduce development predictions before confirmation testing.Source: A. Local adaptation and direct answer… · p. 3

B. Operational contract and execution boundary

The operational contract evaluates evidence at the time an answer becomes available, checks switching against resource and contact constraints, and requires fresh post-intervention observations for recovery. Deterministic checks veto infeasible positive proposals, while the learned scorer’s raw predictions remain the primary evaluation target.

B. Operational contract and execution boundary

  • Evidence measured at time tm is usable only when valid and no older than its maximum permitted age when the answer becomes available.The answer time is t′ = t + ℓ.Source: B. Operational contract and execution… · p. 3
  • Switching permission depends on primary availability, backup readiness, energy and reserve thresholds, switch cost and duration, completion margin, and the scheduled contact interval.Offline questions declare a 0.5 s processing interval; the event diagnostic substitutes its measured interval, and energy uses synthetic percentage points.Source: B. Operational contract and execution… · p. 3
  • Recovery requires an executed intervention completed before the snapshot, followed by at least K valid, healthy observations at distinct timestamps that remain fresh when the answer becomes available.The latest K observations must be obtained after intervention completion.Source: B. Operational contract and execution… · p. 3
  • The deterministic checker vetoes proposed switching or permission that violates the contract and positive recovery claims lacking required evidence, but preserves negative proposals.Dispatch requires consistent action and permission judgments; recovery is evaluated after new observations arrive, while primary evaluation reports raw scorer predictions before corrections.Source: B. Operational contract and execution… · p. 3

A. Motivation and construction

The benchmark tests decisions across operational stages using authored policies and states, controlled changes, and independently checked labels. Its construction preserves candidate meanings while varying operational facts and answer-slot assignments.

A. Motivation and construction

  • All states, policies, numerical ranges, text templates, and labels are authored for this benchmark, with no mission telemetry or existing benchmark examples reused.Its operational framing draws on contact-aware networking and command-effect verification,, and the separation of diagnosis, intervention, and recovery in SatelliteOpsGym.Source: A. Motivation and construction · p. 3
  • A seeded generator samples policy-compliant nominal states, controlled changes alter evidence, contact, resources, or recovery observations, and two independently expressed contract implementations must agree on every label.Question wording and answer-slot mappings are varied while candidate meanings are preserved.Source: A. Motivation and construction · p. 4
  • Changing only backup readiness from ready to false changes the targets from switch/allow to wait/block, while reordering answer slots must preserve the correct operational meaning.Authored energy and time ranges place examples on both sides of operational decision boundaries.Source: A. Motivation and construction · p. 4

B. Boundary cases and a worked example

The benchmark varies policy thresholds and operational conditions to test whether decisions follow the supplied contract. Its worked example shows both how readiness changes a target and how candidate-slot rotation changes its letter.

B. Boundary cases and a worked example

  • Authored nominal cases span evidence-age limits of 12–30 s, switch durations of 4–15 s, recovery requirements of 2–4 timestamps, and policy energy thresholds of 18–42 points.Nominal contact and energy satisfy constraints before controlled transformations, testing whether models follow thresholds supplied in each policy.Source: B. Boundary cases and a worked example · p. 4
  • Controlled cases cross decision boundaries by changing readiness, setting energy below reserve-plus-cost or policy thresholds, closing contact before completion, expiring evidence, or altering recovery observations.Targets follow full priority rules; an added fault need not change the answer when collecting fresh evidence has priority.Source: B. Boundary cases and a worked example · p. 4
  • In the released example, a ready backup makes switch correct, while changing only readiness to false makes wait correct.Rotating candidate meanings changes target letters from B/D to A/C, separating the operational change from its answer-slot change.Source: B. Boundary cases and a worked example · p. 4

C. Splits and suitability for the interface

The main confirmation test uses held-out roots and shared inputs across models, while auxiliary panels probe stress and recovery timing. The benchmark presents the authored operational logic through a common policy-state-question-candidate interface.

C. Splits and suitability for the interface

  • The confirmation test contains 1,024 questions from 256 independent roots, with confirmation roots and normalized states disjoint from training, development, and earlier tests.Training and development contain 1,536 and 256 questions, respectively; confirmation questions are newly sampled from the same authored grammar.Source: C. Splits and suitability for the… · p. 4
  • Every main model receives the same standard test inputs and labels, while the proposed result averages three fixed training seeds and other model rows use one fixed configuration.A task-stratified paired root bootstrap keeps correlated questions from each root together.Source: C. Splits and suitability for the… · p. 4
  • Auxiliary stress and recovery-timing panels probe limits separately from the main standard accuracy comparison.The standard, two stress, and recovery-timing panels include 2,048 questions from 416 distinct roots; operator-note renderings reuse the same 256 standard roots.Source: C. Splits and suitability for the… · p. 4
  • Policy, state, question, and changing candidate meanings provide a common decision interface across four tasks, with Qwen receiving the same information and candidate meanings as decision models.Independent contract implementations generate gold labels, and exact canonical rules reach 100%.Source: C. Splits and suitability for the… · p. 4

V. EVALUATION

The evaluation compares the adapted local Laya scorer with local and cloud references, using complete-request latency and accuracy measures. On the stated desktop setup, its median request time is lower than those of the local generative models.

A. Models and measurement

  • Latency timing uses an Apple M6 with 24 GiB memory, macOS 27.0, PyTorch MPS, FP32, and four host threads.The balanced 128-question subset is timed over three shuffled passes after eight warmups; complete-request time includes formatting, tokenization, transfer, synchronized inference, and answer extraction.Source: A. Models and measurement · p. 5

B. Primary benefit: short local request latency

  • 90.3 ms median complete-request latency for SatOps-Decision compares with 493.9 ms for Qwen generation and 388.4 ms for OneJev.These correspond to 5.47× and 4.30× control-to-proposed time ratios on the same inputs and local setup.Source: B. Primary benefit: short local… · p. 5
  • Domain adaptation raises mean raw standard-test accuracy from 42.5% to 85.3% while retaining approximately the same measured inference cost.Frozen Laya takes 91.2 ms; the timings characterize the evaluated checkpoints and implementations on the stated desktop FP32/MPS setup.Source: B. Primary benefit: short local… · p. 5

A. Models and measurement

C. Decision accuracy after domain adaptation

On identical held-out questions, domain-adapted Laya substantially outperforms frozen Laya and improves accuracy across all four operational tasks. The controls and directional-error results qualify what adaptation changes and what still requires operational checking.

C. Decision accuracy after domain adaptation

  • 85.3±0.8% raw accuracy for SatOps-Decision compares with 42.5% for frozen Laya, 40.7% for Qwen, and 58.6% for OneJev on identical held-out questions.Original Jev 1.14 reaches 87.4% and Jev 1.13 reaches 81.9%; the paired root-bootstrap 95% interval for the adaptation gain over frozen Laya is [40.14, 45.15] percentage points.Source: C. Decision accuracy after domain… · p. 5
  • Adapted accuracy exceeds frozen Laya on all four tasks: 74.5% versus 37.1% for assessment, 89.1% versus 25.8% for action, 78.9% versus 53.9% for permission, and 98.6% versus 53.1% for recovery.Each task has 256 questions; assessment and action have four candidate meanings, while permission and recovery are binary.Source: C. Decision accuracy after domain… · p. 5

D. What changes through adaptation

  • Full adaptation answers 865 questions correctly versus 468 for head-only adaptation, supporting joint adaptation of the encoder and decision layers.The no-augmentation control answers 867 correctly (84.67%), compared with 865 (84.47%) for full adaptation under the same seed; Table IV covers the seed-7 controls on the standard test.Source: D. What changes through adaptation · pp. 5–6, Table IV · p. 6
  • 54 false authorizations among 128 gold-ineligible permission questions (42.2%) and 28 unwarranted switches among 192 non-switch cases motivate deterministic checks before dispatch.Adding deterministic corrections raises checked-system mean standard accuracy to 93.3%; the main accuracy comparison is for the learned scorer before these corrections.Source: D. What changes through adaptation · p. 6

E. Input transfer and recovery diagnostics

Diagnostics show exact rule-based performance on fully specified contracts but weaker transfer to operator-note inputs and recovery decisions. The conclusion identifies real telemetry, mission-specific constraints, and onboard hardware as validation steps.

E. Input transfer and recovery diagnostics

  • 47.4% adapted Laya accuracy on matched operator-note inputs compares with 56.9% for OneJev, while checked adapted Laya restores 4–5 of 24 clocked cases per seed versus 8 for checked frozen Laya.Canonical rules solve the fully specified contract exactly, but recovery still depends on the scorer proposing a useful action; the diagnostics prioritize input-representation transfer and recovery decisions.Source: E. Input transfer and recovery… · p. 6

VI. CONCLUSION

  • SatOps-Decision combines a domain-adapted local Laya scorer with independently checked benchmark labels, and reports 90.3 ms median latency and 85.3±0.8% raw accuracy after adaptation.Original Jev 1.14 reaches 87.4% as an accuracy-only cloud reference on the same questions.Source: §VI · p. 6
  • Further validation is planned with real telemetry, mission-specific constraints, and onboard hardware.The conclusion also names input-representation transfer and recovery decisions as next steps.Source: §VI · p. 6

Related papers

Loading 2610.38697v1…