Jev: Semantic Decision Engines for Network Control
Collection
Jev decision models vs LLMs for intent interpretation at the edge and in 6G O-RAN, and a SoK of semantic decision engines in network control. • 7 items • Updated
Datasets:
Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: CastError
Message: Couldn't cast
screen_id: string
eligibility: string
explicit_loop_claim: string
supported_loop_claim: string
tail_latency: string
deadline_attainment: string
rationale: string
evidence: string
note: string
to
{'screen_id': Value('string'), 'eligibility': Value('string'), 'explicit_loop_claim': Value('string'), 'supported_loop_claim': Value('string'), 'tail_latency': Value('string'), 'deadline_attainment': Value('string'), 'evidence': Value('string'), 'note': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
for key, table in generator:
^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
for item in generator(*args, **kwargs):
~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
screen_id: string
eligibility: string
explicit_loop_claim: string
supported_loop_claim: string
tail_latency: string
deadline_attainment: string
rationale: string
evidence: string
note: string
to
{'screen_id': Value('string'), 'eligibility': Value('string'), 'explicit_loop_claim': Value('string'), 'supported_loop_claim': Value('string'), 'tail_latency': Value('string'), 'deadline_attainment': Value('string'), 'evidence': Value('string'), 'note': Value('string')}
because column names don't match
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1879, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
screen_id string | eligibility string | explicit_loop_claim string | supported_loop_claim string | tail_latency string | deadline_attainment string | evidence string | note string |
|---|---|---|---|---|---|---|---|
S0026 | eligible | other budget or multiple loops | insufficient evidence | not reported | not reported | §5.4, pp.110–113: LLM policies are executed; §6.9.2, p.147: control interval is “≈ 1 minute” despite “real-time” adjustment claim; no decision-latency distribution. | |
S0035 | eligible | not stated | N/A (no claim) | not reported | not reported | §III–IV: natural-language intents are converted into NSDs; the generated NSD was sent to the NFVO, which “deployed the network service on an edge cluster.” | |
S0070 | eligible | not stated | N/A (no claim) | not reported | not reported | §3.4 and §4: ISR “recommend[s] optimal solutions for automated network architecting,” selecting among ACL, VLAN, firewall, proxy, and monitoring options; evaluated against expert recommendations. | |
S0102 | eligible | not stated | N/A (no claim) | not reported | not reported | §3–4: DSM→LLM→CLI pipeline executes generated Cisco NX-OSv configurations; “All models produced syntactically valid and functionally correct CLI.” | |
S0105 | eligible | not stated | N/A (no claim) | not reported | not reported | §IV: GPT-3.5 converts high-level business requirements to NEST JSON and a TMF641 Service Order consumed by OpenSlice to provide a network slice. | |
S0151 | background | N/A | N/A | N/A | N/A | §3–6 and Appendix A: study generates and evaluates a Nile–English translation corpus; natural-language-to-Nile translation and operational configuration are explicitly future work. | |
S0303 | eligible | other budget or multiple loops | matched | not reported | not reported | Abstract; §5.3–5.5: NLP/LLM engine yields NSDs and policies; enforcement takes “less than 1.75 s” even for complex intents, supported by repeated-experiment box plots (Fig. 8). | |
S0375 | eligible | not stated | N/A (no claim) | not reported | not reported | §3–4 and Appendices B–C: DSMs are translated to multivendor CLI and independently checked by LLM evaluators; six scenarios, four profiles, and 72 executions were assessed. | |
S0476 | background | N/A | N/A | N/A | N/A | Methods and Technical Validation: BINS is a dataset study; BERT/DeBERTa models perform named-entity recognition, while generation and deployment of network policies are described only as prospective usage. | |
S0483 | background | N/A | N/A | N/A | N/A | Abstract and §§IV–VI: explicitly a “position paper” proposing an LLM/LTM orchestration architecture and research agenda; it reports no implemented or evaluated network decision engine. | |
S0486 | eligible | not stated | N/A (no claim) | N/A | N/A | Section 8, p.29: “use artificial intelligence” for “configuration creation”; proposes autonomous and supervised IBN models. No timing evaluation of this semantic decision task. | Semantic IBN contribution is an explicitly described conceptual design; other evaluated networking contributions are non-semantic. |
S0501 | eligible | not stated | N/A (no claim) | N/A | N/A | Abstract and §3.2: workflow maps an action’s “intent to a global OSM control policy”; dashboard supports policy management. No decision-step timing evaluation. | Platform/system description with prospective use cases; timing measurements are inapplicable. |
S0534 | eligible | other budget or multiple loops | mismatch | not reported | not reported | §4, pp.9–11: enforcement spans “Non-RT” and “Near-RT” RICs. §6.1, pp.26–27: fastest negotiation is 1.3 s; authors say models need work to meet “sub-second Near-RT RIC targets.” | |
S0645 | eligible | other budget or multiple loops | insufficient evidence | not reported | not reported | §1, p.2: framework claims “sub-20ms decision latency” for edge orchestrators. §4–5 report configuration/fault metrics and 82 ms retraining spikes, but no decision-boundary measurement supporting the sub-20 ms claim. | |
S0653 | eligible | other budget or multiple loops | insufficient evidence | not reported | not reported | §5.2.3, pp.67–68: LLM translates natural-language intents into Rego policies; Groq “enables sub-second response times.” §5.18 evaluates deployment duration, not intent-translation latency. | |
S0659 | background | N/A | N/A | N/A | N/A | Review article based on a systematic literature search; it explains intent-based networking but reports no primary semantic network-decision study. | |
S0678 | eligible | not stated | N/A (no claim) | not reported | not reported | §3.4, pp.41–49: OSDF parses high-level application policies and installs flow rules; §5.6 measures response time, but states no loop class or decision deadline. | |
S0777 | eligible | not stated | N/A (no claim) | not reported | not reported | Methods, pp.68–69: intent submission is decomposed into placement, chaining, and SDN rules; 360 runs measure mean instantiation time, without a decision-loop class, tail quantile, or deadline-attainment rate. | |
S0837 | eligible | RT/fast | insufficient evidence | not reported | not reported | Abstract/Introduction, pp.1–2: chatbot promises diagnostics and assistance “in real-time.” Results, pp.6–8, report MTTR and accuracy but no response-latency measurement at the diagnostic decision boundary. | |
S0838 | eligible | RT/fast | insufficient evidence | not reported | not reported | Abstract, p.11: NLP chatbot “answers user queries in real-time, providing instant support.” Table III, p.13 labels response “Instant” but supplies no measured latency or deadline results. | |
S0003 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0008 | eligible | non-RT | matched | not reported | not reported | Experiment 3/Fig. 4: “average TUT timing of 24.17s”; runtimes were “within acceptable limits, given that intent-based network configuration is a non-real-time operation.” | |
S0031 | eligible | other budget or multiple loops | matched | not reported | not reported | Section V-B: across “around 350 tests,” product proposal time was “about a few minutes” and “less than five minutes”; configuration after product choice was “around 30 seconds.” | |
S0087 | eligible | RT/fast | matched | not reported | not reported | Abstract/Section 5.2: algorithms claimed “fast response time”; server response time was measured while adding conflicting and nonconflicting intents as the database grew to 1,000 intents. | |
S0115 | eligible | RT/fast | matched | not reported | not reported | Section 5.4: smaller models “通常在 10~20 s 即可完成问答任务”，因此“更适合需要高频、实时交互的应用场景”。 | |
S0142 | eligible | other budget or multiple loops | matched | not reported | not reported | Abstract/Section VI-B: “average response time per configuration line remains below 15 milliseconds”; evaluation found it “approximately 15 milliseconds” across varying file sizes. | |
S0144 | eligible | RT/fast | insufficient evidence | not reported | not reported | Sections IV–V: update mechanism “invokes the ML model periodically in real-time”; evaluation reports routing and QoS outcomes but no decision-step execution measurement. | |
S0149 | background | N/A | N/A | N/A | N/A | Review article surveying intent translation; no primary decision study. | |
S0165 | eligible | other budget or multiple loops | matched | not reported | not reported | Fig. 3.2/Section 3.1.3.3: architecture separates “Non-RT RIC” and “Near-RT RIC”; measured edge loop 4.6 ms, hospital loop 3.8–87.9 ms, “well below the sub-second alarm targets.” | |
S0170 | eligible | RT/fast | mismatch | not reported | not reported | Sections V-B/VI-A: mean intent-processing responses were 89.67–163.36 s, yet authors state the “short response times enables real-time decision-making.” | |
S0255 | eligible | RT/fast | insufficient evidence | not reported | not reported | Results/Conclusion: DRL feedback “helps correct any deviations in real-time” and triggers “immediate feedback,” but no decision-step latency or cycle duration is measured. | |
S0258 | eligible | not stated | N/A (no claim) | not reported | not reported | Reported CDFs and threshold attainment concern service-path delay, not decision-step execution latency. | |
S0281 | eligible | not stated | N/A (no claim) | not reported | not reported | LLM chatbot generates DHCP configurations; evaluated with 30 sample conversations. | |
S0300 | eligible | RT/fast | insufficient evidence | not reported | not reported | Sec. III.J, p. 727: the assurance module "reallocates resources in real time"; Sec. IV reports service latency, not decision-loop execution latency. | |
S0373 | eligible | not stated | N/A (no claim) | not reported | not reported | ETS-Chatbot maps natural-language enterprise intents into a network intent model. | |
S0387 | eligible | RT/fast | insufficient evidence | not reported | not reported | Demonstration details, p. 2: a "fast-response agent/controller" triggers reconfiguration and the demo presents "real-time response"; no numerical decision-time measurement is reported. | |
S0453 | eligible | not stated | N/A (no claim) | not reported | not reported | PEIMS uses ONOS intents and OpenFlow rules to select and install DNS port mappings. | |
S0455 | eligible | not stated | N/A (no claim) | not reported | not reported | SentiNet translates and validates natural-language connectivity, isolation, and routing configurations. | |
S0474 | eligible | RT/fast | matched | not reported | not reported | Sec. V.B, pp. 432-433: optimized models are described as viable for "time-sensitive infrastructure operations"; planning latency and end-to-end task-completion time are measured over 50 runs. | |
S0495 | exclude | N/A | N/A | N/A | N/A | Classifies Arabic network-intent text into topic labels but does not produce, select, or check a network decision. | |
S0575 | eligible | not stated | N/A (no claim) | not reported | not reported | DIPS translates intents into VNF provisioning specifications, chains, and placement configurations. | |
S0740 | eligible | other budget or multiple loops | matched | not reported | not reported | Sec. IV.B, R4: LLM calls occur at "approximately every 100 steps" while RL controls between calls; post-change SLO miss and returns support this limited-call cadence. | |
S0746 | background | N/A | N/A | N/A | N/A | AISO selects compute resources and revises Kubernetes deployment descriptors; the evaluated decision is non-network application deployment. | |
S0833 | eligible | RT/fast | insufficient evidence | not reported | not reported | Sec. 5.2, p. 1451: the CNN executes in "negligible time, ensuring instant fault detection suitable for the latency-sensitive control plane"; no numerical latency measurement is supplied. | |
S0358 | background | N/A | N/A | N/A | N/A | Sec. 3, pp. 373–375: ‘Ongoing Work’ is a literature review; component configuration and the intentional layer are described under ‘Plan for Future Work.’ | Doctoral research proposal; no implemented or explicitly specified network decision engine. |
S0409 | eligible | not stated | N/A (no claim) | not reported | not reported | Sec. 4, p. 4: ‘parse the intent and derive the operational steps necessary to meaningfully configure the network’; Sec. 6.1, p. 8: prototype implemented with Mininet and Floodlight. | Evaluation reports average setup, deployment, RTT, and redirection times, but no p95-or-higher latency or deadline-attainment result. |
S0050 | eligible | RT/fast | insufficient evidence | N/A | N/A | §I–IV, pp.14–15: ‘check on errors in real-time and correct these errors in real-time’; the algorithm translates user intent and reroutes traffic, but reports no timing measurement. | |
S0064 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0099 | eligible | RT/fast | insufficient evidence | not reported | not reported | §5.1.1, p.177: ‘Allowed E2E Delay (ms)’ is 10/20 for L1/L2; §5.1.6 reports only ‘average E2E delay’ reductions, without showing decision-boundary attainment of those budgets. | |
S0183 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0248 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0287 | eligible | RT/fast | mismatch | not reported | not reported | §V-D, p.418: ‘senses information about the network data in real time to make the appropriate network path selection’; evaluation measures packet ‘average transmission time,’ not path-decision execution time. | |
S0328 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0359 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0389 | eligible | other budget or multiple loops | matched | not reported | not reported | AICL section, pp.151–154: Performance, Semantic, and Feedback loops; metrics use ‘sliding window (W = 5s)’ and experiments show the L1 loop ‘detected performance drift within seconds’ and restored bandwidth. | |
S0477 | eligible | RT/fast | matched | not reported | not reported | Abstract: ‘comprehensive checking within a few seconds per network with intent updates’; §VI-C reports average rechecking ‘less than 10 seconds’ and the conclusion confirms ‘under 10 seconds per network update.’ | |
S0578 | eligible | other budget or multiple loops | insufficient evidence | N/A | N/A | §II–IV, pp.111–114: hierarchical ‘E2E CCL’ coordinates domain CCLs, and domain policies ‘can make local real-time, closed-loop decisions’; no corresponding timing evaluation is reported. | |
S0682 | eligible | RT/fast | mismatch | not reported | not reported | §II-E, pp.17875–17876 claims ‘real-time monitoring’ and ‘low-latency inference’; §III-C measures segmentation inference (0.04 s Jetson Nano, 1.50 s Raspberry Pi), not the IBN prioritization/control decision boundary. | |
S0787 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0061 | eligible | RT/fast | insufficient evidence | not reported | not reported | Sections IV–V: orchestration is "validated using a 7-minute iperf3 test"; the conclusion claims "real-time orchestration and monitoring," but no decision-step latency is measured. | |
S0072 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0122 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0200 | eligible | RT/fast | insufficient evidence | not reported | not reported | Introduction: "real-time decision-making capabilities." Figure 10 reports end-to-end OSM VNF instantiation/configuration, not the LLM decision step or a stated real-time threshold. | |
S0250 | eligible | RT/fast | insufficient evidence | not reported | not reported | Abstract: "centralized real-time reinforced detection scheme." Section V evaluates accuracy, TPR/FPR, rewards and convergence episodes, but reports no decision-step latency. | |
S0289 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0335 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0363 | eligible | RT/fast | matched | not reported | reported | Section 4.2.2, Tables 6–7: a "time bound of 25 min"; INCS synthesis takes 5.32–53.69 s, while timeout counts are reported as b/8. | |
S0406 | eligible | RT/fast | insufficient evidence | not reported | not reported | Introduction/conclusion: NMT lets operators complete transformations "quickly" and "greatly speed up the processing process"; evaluation reports translation quality only, not translation latency. | |
S0505 | eligible | RT/fast | matched | not reported | not reported | Discussion: "Initial Mininet trials indicate that mitigation intents are consistently converted into OpenFlow rules with sub-millisecond latency." No latency quantile or deadline-attainment count is given. | |
S0581 | background | N/A | N/A | N/A | N/A | One-page OFC abstract; no methods or evaluation and no described semantic decision engine. | |
S0700 | background | N/A | N/A | N/A | N/A | Systematic comparative literature review and mitigation guide; no primary semantic decision engine. | |
S0802 | eligible | not stated | N/A (no claim) | not reported | not reported | ||
S0062 | eligible | not stated | N/A (no claim) | not reported | not reported | Section V, pp. 147–149: the hybrid algorithm configures service chains from requested intent QoS and reports intent-deployment execution times. | |
S0079 | eligible | other budget or multiple loops | insufficient evidence | not reported | not reported | Sections 3 and 6, pp. A15–A18: vLink data are “collected periodically (e.g., every minute)” and the agent determines capacity “for the next period”; decision-runtime support is absent. | Invited tutorial containing primary implementation and numerical studies. |
S0133 | eligible | not stated | N/A (no claim) | not reported | not reported | Chapters 4 and 7: GPT-4 translates natural-language requests into YANG configuration; evaluation reports 50% correct across 12 operator requests. | |
S0203 | exclude | N/A | N/A | N/A | N/A | Cryptographic LSSS access-policy reconstruction accepts formal policies and attributes; no semantic intent or high-level-objective engine produces a network configuration, orchestration, or diagnostic decision. | |
S0251 | exclude | N/A | N/A | N/A | N/A | Blockchain sharding, consensus, and malicious-node detection are fixed mechanisms; user-intent adaptation is asserted but no semantic intent or explicit high-level-objective processing method is described. | |
S0305 | eligible | not stated | N/A (no claim) | N/A | N/A | Sections III and V: connectivity intents are compiled into hierarchical path, spectrum, router, and border intents across multiple IP-optical domains. | Conceptual architecture and worked deployment example; no timing evaluation. |
S0342 | eligible | not stated | N/A (no claim) | not reported | not reported | Sections II–VI: human-readable intent graphs compile into local flow entries; MD-IDN is deployed on SAVI and intent-compilation time is evaluated. | |
S0370 | eligible | RT/fast | matched | reported | not reported | Sections 3–5, pp. 414–422: system claims “real-time detection”; Table 8 measures the detection-to-mitigation boundary: total mean 12.4 ms and “95%ile” 18.7 ms. | |
S0414 | eligible | RT/fast | insufficient evidence | not reported | not reported | Sections II–V: ICC-Bandit is designed for “real-time agent selection”; evaluation uses a 120 ms latency budget but reports cumulative constraint violation rather than selector execution latency. | |
S0522 | eligible | not stated | N/A (no claim) | not reported | not reported | Sections III–IV: IMR monitors intent flow statistics, an external application computes optimized routes, and ONOS enforces rerouted intent paths. | Demo paper with implemented system but no timing evaluation. |
S0618 | eligible | RT/fast | insufficient evidence | not reported | not reported | Sections 4.2 and 4.3.1, pp. 140–160: controller provides “real-time resource management” and analyzes metrics “in real time”; experiments do not measure decision-to-enforcement latency. | Book chapter containing proposed models, implementation, and experiments. |
S0708 | eligible | RT/fast | insufficient evidence | not reported | not reported | Section 3.4, p. 6: pipeline promises automated mitigation “with minimal delay” and “rapid” containment; results report segmentation and rule validation, not containment-loop latency. | |
S0821 | eligible | not stated | N/A (no claim) | not reported | not reported | Sections 4–6: ComBERT scores protocol-component semantics; thresholded, constraint-screened fusion selects reconstructed service units for scenario-specific RAN processing chains. | |
S0063 | eligible | not stated | N/A (no claim) | N/A | N/A | Sections 2 and 8: "Mapping of the user intents to available services"; "service topology is passed to the user." | Conceptual design; no evaluation. |
S0088 | background | N/A | N/A | N/A | N/A | Abstract discusses how "MR improves AIOps" but supplies no concrete system, method, benchmark, measurement study, or explicitly described conceptual design. | Position/background article. |
S0168 | eligible | not stated | N/A (no claim) | not reported | not reported | Section 2 evaluates LLM generation of "complete configurations of a specified number of network-wide intentions" on virtual and physical ROADM networks. | No decision-latency evaluation. |
S0230 | eligible | RT/fast | insufficient evidence | not reported | not reported | Sections III-E/F: pipeline is "executed continuously on every time window t"; GNN can "compute the optimal path in real-time." Section IV reports path selections, not decision time. | |
S0267 | eligible | RT/fast | insufficient evidence | not reported | not reported | Section IV says the system adapts "to model performance variations in real time"; Table II reports average generation latency of 48.8 s and 52.8 s, but no real-time threshold. | |
S0317 | eligible | RT/fast | insufficient evidence | not reported | not reported | Section I: IDN "monitors user requirements and network status in real time"; Section III demonstrates a created end-to-end slice but reports no decision-step timing. | |
S0357 | eligible | non-RT | insufficient evidence | N/A | N/A | Section 3.2: centralized DTN "no real-time operation is feasible"; the proposed design uses an "asynchronous and delay-tolerant mode of operation." No timing evaluation. | Conceptual design. |
S0383 | eligible | RT/fast | insufficient evidence | N/A | N/A | Abstract: network can "recognize the users' intent in real-time and then update the network slicing." Sections III-IV provide the framework and scenarios without timing evaluation. | Conceptual framework; no evaluation. |
S0419 | eligible | other budget or multiple loops | insufficient evidence | not reported | not reported | Section 5 models traffic "with an analysis time of 5 min"; Figure 8 collects data "every dt" before GNN routing decisions. No execution-time measurement is reported. | |
S0536 | eligible | other budget or multiple loops | matched | not reported | not reported | Abstract/Section IV: Oz plans cache deployment "in less than 10 seconds"; evaluation measures "latency to translate a single intent into configuration" for guided and exhaustive search. | |
S0642 | eligible | not stated | N/A (no claim) | not reported | not reported | Sections 4-5: EASYACL interprets natural language and synthesizes Cisco/Juniper ACL commands; three case studies report generated configurations but no decision timing. | |
S0779 | eligible | other budget or multiple loops | matched | not reported | reported | Section V-A/D: UIA predicts SDVs "every 100 ms"; NOA latency remains "below 80 ms"; Figure 20 reports "over 90%" of resolutions within the "500 ms service-level time budget." | |
S0829 | eligible | not stated | N/A (no claim) | not reported | not reported | Sections III-V: the LLM agent diagnoses failures, executes recovery commands, and verifies restoration on a CML digital twin; Table III reports mean recovery time with standard deviation. | No explicit decision deadline or control-loop class. |
S0026 | eligible | not stated | N/A (no claim) | not reported | not reported | Ch5/6: GPT-3.5/4 few-shot decomposes intents into executable policies on SAVI OpenStack; Tables 5.1/6.4 report average times over 5 trials only (e.g., 'GPT API Execution Time (s) 23'). | PhD thesis. Sec 3.9.2 'we consider having a dual-loop mechanism' (fast/slow) is a discussion-only proposal with no budget, not tied to the LLM step. Sec 6.9 SLA spikes concern a non-semantic core-allocation controller's app response time. |
S0035 | eligible | not stated | N/A (no claim) | not reported | not reported | Sec III-IV: Code Llama translates natural-language intents into NSDs deployed via NFVO on EURECOM 5G facility; Fig. 4 'NSD generation time' vs. number of applications, no quantile. |
End of preview.
SoK-JEV
Data and results for SoK: Semantic Decision Engines in Network Control Loops, by Delong Li, Chen Li, Xu Wang, Haochen Gong, Rui Lang, and Guangsheng Yu, University of Technology Sydney (UTS). arXiv: 2610.06425.
Code and reproduction commands: SoK-JEV.
Frozen packages
Six archives in data/ retain their original bytes, member paths, and SHA-256 manifests.
| Archive | Contents |
|---|---|
analysis.tar.gz |
Frozen estimates, 111,372 normalized fixed records, source manifests and numeric table baselines |
fixed-workloads.tar.gz |
Selected contract, policy, placement and forwarding request records and fixtures |
network-workloads.tar.gz |
Network task inputs, responses, timing and execution ledgers selected by source manifests |
network-source.tar.gz |
Recorded network-suite Python/C++ code, tests, configuration and licenses |
coding.tar.gz |
Final coding, two aligned independent coding records, scope boundaries, agreement input, family view and eligibility |
interventions.tar.gz |
Frozen scenes/images/code, 2,400 physical executions and all 166,432 records in 84 FIFO queues |
Literature-review records
Counts and sizes below are computed from this staged tree. Full texts are omitted. The coverage audit retains the original seed-42 sample, including two draws whose full text was not retrieved.
Folder under runs/literature-review/ |
Files | Bytes |
|---|---|---|
coverage-audit/ |
41 | 377,335 |
q2/ |
42 | 1,020,779 |
recall/ |
7 | 6,856,216 |
screening/ |
34 | 3,343,207 |
taxonomy/ |
58 | 778,308 |
Download data/* with --local-dir studies/sok in the code repository and runs/* with --local-dir .. This lands archives in studies/sok/data/ and records in runs/literature-review/.
License: CC BY 4.0. Third-party material in frozen archives retains its original terms.
77 abstracts longer than 3,000 characters, which contain full-text sections in the index, are omitted. Every released JSON/JSONL abstract longer than 3,000 characters is replaced by [omitted: indexed abstract longer than 3,000 characters].
The archives carry third-party and sister-release material under the licences included with it, including the 6G release’s GPL-2.0-only 5G-LENA patches in studies/sok/data/network-source.tar.gz (HF data/network-source.tar.gz), with their licence text at data/source/sok/cross-study/6g/github/src/ranbench/ns3/patches/LICENSE, and the sister release’s own LICENSE at data/source/sok/cross-study/6g/github/LICENSE.
- Downloads last month
- 40