{ "ran": { "budget": { "grid": [ 0.01, 0.01122, 0.012589, 0.014125, 0.015849, 0.017783, 0.019953, 0.022387, 0.025119, 0.028184, 0.031623, 0.035481, 0.039811, 0.044668, 0.050119, 0.056234, 0.063096, 0.070795, 0.079433, 0.089125, 0.1, 0.112202, 0.125893, 0.141254, 0.158489, 0.177828, 0.199526, 0.223872, 0.251189, 0.281838, 0.316228, 0.354813, 0.398107, 0.446684, 0.501187, 0.562341, 0.630957, 0.707946, 0.794328, 0.891251, 1.0, 1.122018, 1.258925, 1.412538, 1.584893, 1.778279, 1.995262, 2.238721, 2.511886, 2.818383, 3.162278, 3.548134, 3.981072, 4.466836, 5.011872, 5.623413, 6.309573, 7.079458, 7.943282, 8.912509, 10.0 ], "series": [ { "key": "Jev-1.13.0", "label": "Jev-1.13.0", "colour": "#F5A524", "dash": "", "family": "decision model", "miss": [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.8525, 0.45104, 0.20521, 0.07875, 0.03271, 0.01417, 0.00937, 0.00604, 0.00479, 0.00396, 0.00271, 0.0025, 0.00208, 0.00167, 0.00167, 0.00146, 0.00125, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063, 0.00063 ], "n": 4800 }, { "key": "SemIf-Qwen3.5-4B", "label": "SemIf-Qwen3.5-4B", "colour": "#7CC4FA", "dash": "", "family": "decision model", "miss": [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.76646, 0.54958, 0.4675, 0.24458, 0.17083, 0.13375, 0.07687, 0.0675, 0.01417, 0.00208, 0.00104, 0.00063, 0.00021, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 ], "n": 4800 }, { "key": "AnyJev-L0", "label": "AnyJev-L0", "colour": "#3E86D6", "dash": "", "family": "decision model", "miss": [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.97708, 0.7925, 0.42917, 0.2875, 0.18396, 0.14438, 0.10812, 0.05625, 0.04271, 0.02604, 0.00958, 0.00042, 0.00021, 0.00021, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 ], "n": 4800 }, { "key": "Qwen3.5-4B-JSON", "label": "Qwen3.5-4B-JSON", "colour": "#A8B4C6", "dash": "2 4", "family": "JSON reference", "miss": [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9875, 0.25604, 0.17271, 0.0325, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 ], "n": 4800 }, { "key": "DeepSeek-V4.1-Flash", "label": "DeepSeek-V4.1-Flash", "colour": "#C9A0FF", "dash": "7 4", "family": "hosted LLM", "miss": [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99375, 0.79125, 0.49625, 0.23042, 0.14417, 0.08417, 0.05042, 0.03333, 0.025, 0.01937, 0.01667, 0.015, 0.01271, 0.01188, 0.01125, 0.00937, 0.00833, 0.00729, 0.00562, 0.005, 0.00438, 0.00292, 0.00271, 0.0025, 0.0025, 0.00187, 0.00167, 0.00125, 0.0 ], "n": 4800 }, { "key": "GLM-5.3-Flash", "label": "GLM-5.3-Flash", "colour": "#F28CC0", "dash": "7 4", "family": "hosted LLM", "miss": [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99854, 0.98062, 0.92292, 0.82125, 0.66021, 0.51042, 0.36771, 0.25438, 0.18271, 0.13396, 0.10521, 0.08208, 0.065, 0.05208, 0.04583, 0.0375, 0.02604, 0.01896, 0.015, 0.01188, 0.00896, 0.0075, 0.00729, 0.00708 ], "n": 4800 }, { "key": "Qwen3.8-Flash", "label": "Qwen3.8-Flash", "colour": "#E8EEF6", "dash": "7 4", "family": "hosted LLM", "miss": [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.995, 0.90667, 0.69979, 0.48146, 0.27646, 0.16208, 0.12812, 0.11458, 0.10833, 0.10604, 0.095, 0.08125, 0.05229, 0.02375, 0.00167, 0.00042, 0.00021 ], "n": 4800 } ] }, "load": { "rates": [ "0.1", "0.5", "1", "2" ], "models": [ { "key": "Jev-1.13.0", "label": "Jev-1.13.0", "colour": "#F5A524", "cells": { "0.1": { "within": 1.0, "q95": 0.00242321649871, "rho": 0.00774490851172, "ns": false }, "0.5": { "within": 0.996666666667, "q95": 0.00241110772514, "rho": 0.0376270884167, "ns": false }, "1": { "within": 1.0, "q95": 0.00212802178505, "rho": 0.0740113686964, "ns": false }, "2": { "within": 0.996666666667, "q95": 0.00213078528571, "rho": 0.148925406379, "ns": false } } }, { "key": "SemIf-Qwen3.5-4B", "label": "SemIf-Qwen3.5-4B", "colour": "#7CC4FA", "cells": { "0.1": { "within": 0.986666666667, "q95": 0.0271297540414, "rho": 0.0111078524474, "ns": false }, "0.5": { "within": 1.0, "q95": 0.00598722592751, "rho": 0.0397305522337, "ns": false }, "1": { "within": 0.973333333333, "q95": 0.00347122133854, "rho": 0.0953356399535, "ns": false }, "2": { "within": 0.886666666667, "q95": 0.259333422788, "rho": 0.260765866279, "ns": false } } }, { "key": "AnyJev-L0", "label": "AnyJev-L0", "colour": "#3E86D6", "cells": { "0.1": { "within": 0.743333333333, "q95": 0.0262772203122, "rho": 0.017490420259, "ns": false }, "0.5": { "within": 0.523333333333, "q95": 0.00600689061382, "rho": 0.119400981213, "ns": false }, "1": { "within": 0.29, "q95": 1.47430058827, "rho": 0.347232004282, "ns": false }, "2": { "within": 0.00666666666667, "q95": 34.5332568751, "rho": 1.16867251105, "ns": true } } }, { "key": "Qwen3.5-4B-JSON", "label": "Qwen3.5-4B-JSON", "colour": "#A8B4C6", "cells": { "0.1": { "within": 0.996666666667, "q95": 0.0281577825439, "rho": 0.0104900133269, "ns": false }, "0.5": { "within": 1.0, "q95": 0.00596181570505, "rho": 0.0520469891445, "ns": false }, "1": { "within": 1.0, "q95": 0.00320362293426, "rho": 0.105574852713, "ns": false }, "2": { "within": 1.0, "q95": 0.00235588505438, "rho": 0.217334982032, "ns": false } } }, { "key": "DeepSeek-V4.1-Flash", "label": "DeepSeek-V4.1-Flash", "colour": "#C9A0FF", "cells": { "0.1": { "within": 0.89, "q95": 0.00249332673194, "rho": 0.0201450919349, "ns": false }, "0.5": { "within": 0.926666666667, "q95": 0.00231907167591, "rho": 0.0953574258694, "ns": false }, "1": { "within": 0.84, "q95": 0.913890524863, "rho": 0.220726166939, "ns": false }, "2": { "within": 0.846666666667, "q95": 6.36626734026, "rho": 0.403441529108, "ns": false } } }, { "key": "GLM-5.3-Flash", "label": "GLM-5.3-Flash", "colour": "#F28CC0", "cells": { "0.1": { "within": 0.22, "q95": 0.00244800258121, "rho": 0.0388473183811, "ns": false }, "0.5": { "within": 0.246666666667, "q95": 1.43722221922, "rho": 0.247836861274, "ns": false }, "1": { "within": 0.706666666667, "q95": 0.00213500914872, "rho": 0.232061307419, "ns": false }, "2": { "within": 0.913333333333, "q95": 0.230924594052, "rho": 0.350004301618, "ns": false } } }, { "key": "Qwen3.8-Flash", "label": "Qwen3.8-Flash", "colour": "#E8EEF6", "cells": { "0.1": { "within": 0.0, "q95": 0.00233765345795, "rho": 0.0585137150097, "ns": false }, "0.5": { "within": 0.0, "q95": 0.00225338492724, "rho": 0.276709019023, "ns": false }, "1": { "within": 0.0, "q95": 1.25980753152, "rho": 0.524391550184, "ns": false }, "2": { "within": 0.0, "q95": 20.2953990169, "rho": 1.08503134966, "ns": true } } } ] }, "path": [ { "key": "Jev-1.13.0", "label": "Jev-1.13.0", "colour": "#F5A524", "deploy": "hosted", "dec": 285.56385450100004, "a1": 16.6313648224, "e2": 2.43294239044 }, { "key": "SemIf-Qwen3.5-4B", "label": "SemIf-Qwen3.5-4B", "colour": "#7CC4FA", "deploy": "self-hosted", "dec": 356.81902105, "a1": 16.7102813721, "e2": 3.13997268677 }, { "key": "AnyJev-L0", "label": "AnyJev-L0", "colour": "#3E86D6", "deploy": "self-hosted", "dec": 1068.05595849, "a1": 19.110918045, "e2": 3.058552742 }, { "key": "Qwen3.5-4B-JSON", "label": "Qwen3.5-4B-JSON", "colour": "#A8B4C6", "deploy": "self-hosted", "dec": 432.351958007, "a1": 17.7140235901, "e2": 2.78198719025 }, { "key": "DeepSeek-V4.1-Flash", "label": "DeepSeek-V4.1-Flash", "colour": "#C9A0FF", "deploy": "hosted", "dec": 526.540166698, "a1": 16.2822008133, "e2": 2.89964675903 }, { "key": "GLM-5.3-Flash", "label": "GLM-5.3-Flash", "colour": "#F28CC0", "deploy": "hosted", "dec": 1054.51387493, "a1": 15.5273675919, "e2": 2.5874376297 }, { "key": "Qwen3.8-Flash", "label": "Qwen3.8-Flash", "colour": "#E8EEF6", "deploy": "hosted", "dec": 2352.1580630900003, "a1": 16.3269042969, "e2": 2.62403488159 } ], "slots": 4, "load_n": 300, "path_n": 30, "quality": { "Fresh": [ { "key": "Jev-1.13.0", "correct": 0.947, "timely": 0.997, "p95": 0.38 }, { "key": "SemIf-Qwen3.5-4B", "correct": 0.1, "timely": 1.0, "p95": 0.52 }, { "key": "AnyJev-L0", "correct": 0.097, "timely": 1.0, "p95": 0.42 }, { "key": "DeepSeek-V4.1-Flash", "correct": 0.95, "timely": 0.99, "p95": 0.81 }, { "key": "GLM-5.3-Flash", "correct": 0.94, "timely": 0.237, "p95": 2.11 }, { "key": "Qwen3.8-Flash", "correct": 0.907, "timely": 0.0, "p95": 2.94 }, { "key": "Qwen3.5-4B-JSON", "correct": 0.067, "timely": 1.0, "p95": 0.44 } ], "Stale": [ { "key": "Jev-1.13.0", "correct": 0.843, "timely": 1.0, "p95": 0.42 }, { "key": "SemIf-Qwen3.5-4B", "correct": 0.1, "timely": 0.997, "p95": 0.86 }, { "key": "AnyJev-L0", "correct": 0.077, "timely": 0.593, "p95": 1.13 }, { "key": "DeepSeek-V4.1-Flash", "correct": 0.83, "timely": 0.91, "p95": 2.09 }, { "key": "GLM-5.3-Flash", "correct": 0.947, "timely": 0.063, "p95": 3.93 }, { "key": "Qwen3.8-Flash", "correct": 0.85, "timely": 0.0, "p95": 6.51 }, { "key": "Qwen3.5-4B-JSON", "correct": 0.08, "timely": 1.0, "p95": 0.53 } ], "Noisy": [ { "key": "Jev-1.13.0", "correct": 0.937, "timely": 0.997, "p95": 0.36 }, { "key": "SemIf-Qwen3.5-4B", "correct": 0.107, "timely": 1.0, "p95": 0.52 }, { "key": "AnyJev-L0", "correct": 0.093, "timely": 1.0, "p95": 0.44 }, { "key": "DeepSeek-V4.1-Flash", "correct": 0.963, "timely": 0.927, "p95": 1.16 }, { "key": "GLM-5.3-Flash", "correct": 0.94, "timely": 0.193, "p95": 3.0 }, { "key": "Qwen3.8-Flash", "correct": 0.913, "timely": 0.0, "p95": 6.97 }, { "key": "Qwen3.5-4B-JSON", "correct": 0.07, "timely": 1.0, "p95": 0.44 } ], "Contradictory": [ { "key": "Jev-1.13.0", "correct": 0.81, "timely": 1.0, "p95": 0.4 }, { "key": "SemIf-Qwen3.5-4B", "correct": 0.103, "timely": 1.0, "p95": 0.71 }, { "key": "AnyJev-L0", "correct": 0.067, "timely": 0.997, "p95": 0.5 }, { "key": "DeepSeek-V4.1-Flash", "correct": 0.817, "timely": 0.957, "p95": 0.95 }, { "key": "GLM-5.3-Flash", "correct": 0.813, "timely": 0.08, "p95": 2.95 }, { "key": "Qwen3.8-Flash", "correct": 0.767, "timely": 0.0, "p95": 7.14 }, { "key": "Qwen3.5-4B-JSON", "correct": 0.08, "timely": 1.0, "p95": 0.48 } ] }, "radio": { "points": [ { "rate": 0.1, "speed": 3, "models": { "Jev-1.13.0": { "throughput": 0.17, "packet_p95": 127.1, "handover_rate": 0.131, "handover_p95": 152.5, "rlf": 12.0, "stale_exposure": 1.97, "affected": 23.91, "network": 22.3 }, "SemIf-Qwen3.5-4B": { "throughput": 0.181, "packet_p95": 131.2, "handover_rate": 0.23, "handover_p95": 105.0, "rlf": 11.0, "stale_exposure": 2.49, "affected": 23.58, "network": 21.62 }, "AnyJev-L0": { "throughput": 0.182, "packet_p95": 130.2, "handover_rate": 0.136, "handover_p95": 85.0, "rlf": 10.0, "stale_exposure": 2.57, "affected": 20.11, "network": 20.81 }, "DeepSeek-V4.1-Flash": { "throughput": 0.175, "packet_p95": 130.7, "handover_rate": 0.197, "handover_p95": 113.2, "rlf": 11.0, "stale_exposure": 3.83, "affected": 26.15, "network": 22.6 }, "GLM-5.3-Flash": { "throughput": 0.171, "packet_p95": 134.6, "handover_rate": 0.117, "handover_p95": 100.0, "rlf": 16.0, "stale_exposure": 8.7, "affected": 28.65, "network": 23.73 }, "Qwen3.8-Flash": { "throughput": 0.179, "packet_p95": 135.3, "handover_rate": 0.146, "handover_p95": 95.6, "rlf": 17.0, "stale_exposure": 15.18, "affected": 26.42, "network": 22.25 }, "Qwen3.5-4B-JSON": { "throughput": 0.184, "packet_p95": 133.5, "handover_rate": 0.11, "handover_p95": 109.9, "rlf": 9.0, "stale_exposure": 2.81, "affected": 21.86, "network": 21.74 } } }, { "rate": 0.1, "speed": 60, "models": { "Jev-1.13.0": { "throughput": 0.168, "packet_p95": 177.3, "handover_rate": 0.668, "handover_p95": 95.0, "rlf": 64140.0, "stale_exposure": 1.96, "affected": 30.93, "network": 34.43 }, "SemIf-Qwen3.5-4B": { "throughput": 0.183, "packet_p95": 154.2, "handover_rate": 0.689, "handover_p95": 85.0, "rlf": 62698.0, "stale_exposure": 2.51, "affected": 25.97, "network": 31.23 }, "AnyJev-L0": { "throughput": 0.187, "packet_p95": 151.5, "handover_rate": 0.695, "handover_p95": 85.0, "rlf": 61976.0, "stale_exposure": 2.53, "affected": 25.6, "network": 30.56 }, "DeepSeek-V4.1-Flash": { "throughput": 0.171, "packet_p95": 151.5, "handover_rate": 0.69, "handover_p95": 85.0, "rlf": 62191.0, "stale_exposure": 4.01, "affected": 28.11, "network": 32.49 }, "GLM-5.3-Flash": { "throughput": 0.17, "packet_p95": 166.3, "handover_rate": 0.678, "handover_p95": 95.5, "rlf": 63452.0, "stale_exposure": 9.18, "affected": 31.98, "network": 34.39 }, "Qwen3.8-Flash": { "throughput": 0.187, "packet_p95": 152.8, "handover_rate": 0.697, "handover_p95": 85.0, "rlf": 62957.0, "stale_exposure": 15.58, "affected": 25.9, "network": 30.18 }, "Qwen3.5-4B-JSON": { "throughput": 0.171, "packet_p95": 162.3, "handover_rate": 0.676, "handover_p95": 85.0, "rlf": 62717.0, "stale_exposure": 2.79, "affected": 30.14, "network": 34.18 } } }, { "rate": 0.1, "speed": 120, "models": { "Jev-1.13.0": { "throughput": 0.177, "packet_p95": 168.2, "handover_rate": 1.402, "handover_p95": 90.0, "rlf": 67160.0, "stale_exposure": 1.94, "affected": 27.46, "network": 31.41 }, "SemIf-Qwen3.5-4B": { "throughput": 0.173, "packet_p95": 160.3, "handover_rate": 1.393, "handover_p95": 100.0, "rlf": 66937.0, "stale_exposure": 2.44, "affected": 32.54, "network": 33.75 }, "AnyJev-L0": { "throughput": 0.181, "packet_p95": 176.1, "handover_rate": 1.405, "handover_p95": 90.0, "rlf": 67513.0, "stale_exposure": 2.62, "affected": 29.52, "network": 31.21 }, "DeepSeek-V4.1-Flash": { "throughput": 0.181, "packet_p95": 169.0, "handover_rate": 1.385, "handover_p95": 88.2, "rlf": 67326.0, "stale_exposure": 3.99, "affected": 29.81, "network": 32.36 }, "GLM-5.3-Flash": { "throughput": 0.178, "packet_p95": 174.2, "handover_rate": 1.392, "handover_p95": 95.0, "rlf": 67016.0, "stale_exposure": 8.8, "affected": 29.82, "network": 31.99 }, "Qwen3.8-Flash": { "throughput": 0.171, "packet_p95": 187.8, "handover_rate": 1.367, "handover_p95": 105.0, "rlf": 67133.0, "stale_exposure": 16.14, "affected": 31.81, "network": 34.39 }, "Qwen3.5-4B-JSON": { "throughput": 0.181, "packet_p95": 169.7, "handover_rate": 1.396, "handover_p95": 85.0, "rlf": 66834.0, "stale_exposure": 2.73, "affected": 29.39, "network": 31.87 } } }, { "rate": 0.3, "speed": 3, "models": { "Jev-1.13.0": { "throughput": 0.211, "packet_p95": 133.3, "handover_rate": 0.06, "handover_p95": 83.8, "rlf": 4.0, "stale_exposure": 2.06, "affected": 21.04, "network": 16.08 }, "SemIf-Qwen3.5-4B": { "throughput": 0.073, "packet_p95": 126.1, "handover_rate": 0.042, "handover_p95": 81.5, "rlf": 5.0, "stale_exposure": 2.54, "affected": 21.02, "network": 16.81 }, "AnyJev-L0": { "throughput": 0.164, "packet_p95": 137.9, "handover_rate": 0.042, "handover_p95": 93.0, "rlf": 7.0, "stale_exposure": 2.78, "affected": 20.42, "network": 17.23 }, "DeepSeek-V4.1-Flash": { "throughput": 0.209, "packet_p95": 133.4, "handover_rate": 0.054, "handover_p95": 91.5, "rlf": 3.0, "stale_exposure": 4.1, "affected": 20.91, "network": 16.0 }, "GLM-5.3-Flash": { "throughput": 0.163, "packet_p95": 137.5, "handover_rate": 0.042, "handover_p95": 121.8, "rlf": 3.0, "stale_exposure": 9.89, "affected": 22.77, "network": 17.45 }, "Qwen3.8-Flash": { "throughput": 0.197, "packet_p95": 127.9, "handover_rate": 0.06, "handover_p95": 85.0, "rlf": 6.0, "stale_exposure": 16.23, "affected": 19.9, "network": 15.52 }, "Qwen3.5-4B-JSON": { "throughput": 0.195, "packet_p95": 128.8, "handover_rate": 0.058, "handover_p95": 85.1, "rlf": 4.0, "stale_exposure": 2.89, "affected": 21.44, "network": 16.78 } } }, { "rate": 0.3, "speed": 30, "models": { "Jev-1.13.0": { "throughput": 0.155, "packet_p95": 171.4, "handover_rate": 0.123, "handover_p95": 100.0, "rlf": 42.0, "stale_exposure": 1.89, "affected": 35.0, "affected_ci": [ 29.91, 40.24 ], "network": 31.17, "network_ci": [ 26.56, 35.92 ] }, "SemIf-Qwen3.5-4B": { "throughput": 0.186, "packet_p95": 159.2, "handover_rate": 0.131, "handover_p95": 117.5, "rlf": 37.0, "stale_exposure": 2.39, "affected": 34.08, "affected_ci": [ 30.07, 37.82 ], "network": 30.3, "network_ci": [ 26.2, 34.39 ] }, "AnyJev-L0": { "throughput": 0.168, "packet_p95": 164.9, "handover_rate": 0.128, "handover_p95": 120.0, "rlf": 26.0, "stale_exposure": 2.6, "affected": 34.96, "affected_ci": [ 29.75, 40.11 ], "network": 31.6, "network_ci": [ 27.17, 35.51 ] }, "DeepSeek-V4.1-Flash": { "throughput": 0.136, "packet_p95": 151.1, "handover_rate": 0.123, "handover_p95": 134.0, "rlf": 198.0, "stale_exposure": 3.78, "affected": 35.03, "affected_ci": [ 30.04, 39.76 ], "network": 30.51, "network_ci": [ 26.51, 33.99 ] }, "GLM-5.3-Flash": { "throughput": 0.225, "packet_p95": 160.8, "handover_rate": 0.138, "handover_p95": 104.3, "rlf": 17.0, "stale_exposure": 8.81, "affected": 32.56, "affected_ci": [ 27.75, 36.99 ], "network": 28.51, "network_ci": [ 25.36, 31.11 ] }, "Qwen3.8-Flash": { "throughput": 0.154, "packet_p95": 179.3, "handover_rate": 0.124, "handover_p95": 135.5, "rlf": 43.0, "stale_exposure": 14.05, "affected": 34.03, "affected_ci": [ 29.24, 38.68 ], "network": 30.51, "network_ci": [ 26.68, 33.67 ] }, "Qwen3.5-4B-JSON": { "throughput": 0.122, "packet_p95": 159.3, "handover_rate": 0.119, "handover_p95": 140.0, "rlf": 56.0, "stale_exposure": 2.68, "affected": 35.54, "affected_ci": [ 30.82, 40.76 ], "network": 31.41, "network_ci": [ 26.81, 36.01 ] } } }, { "rate": 0.3, "speed": 60, "models": { "Jev-1.13.0": { "throughput": 0.156, "packet_p95": 185.6, "handover_rate": 0.235, "handover_p95": 95.0, "rlf": 18345.0, "stale_exposure": 1.94, "affected": 33.47, "network": 32.04 }, "SemIf-Qwen3.5-4B": { "throughput": 0.153, "packet_p95": 150.0, "handover_rate": 0.24, "handover_p95": 85.0, "rlf": 18351.0, "stale_exposure": 2.4, "affected": 34.41, "network": 31.86 }, "AnyJev-L0": { "throughput": 0.151, "packet_p95": 151.5, "handover_rate": 0.242, "handover_p95": 88.5, "rlf": 18483.0, "stale_exposure": 2.67, "affected": 35.07, "network": 32.02 }, "DeepSeek-V4.1-Flash": { "throughput": 0.159, "packet_p95": 154.9, "handover_rate": 0.244, "handover_p95": 85.0, "rlf": 18922.0, "stale_exposure": 3.75, "affected": 32.41, "network": 30.66 }, "GLM-5.3-Flash": { "throughput": 0.148, "packet_p95": 157.3, "handover_rate": 0.238, "handover_p95": 85.0, "rlf": 18699.0, "stale_exposure": 9.28, "affected": 35.31, "network": 32.1 }, "Qwen3.8-Flash": { "throughput": 0.148, "packet_p95": 154.4, "handover_rate": 0.24, "handover_p95": 95.0, "rlf": 18665.0, "stale_exposure": 16.57, "affected": 34.67, "network": 31.5 }, "Qwen3.5-4B-JSON": { "throughput": 0.147, "packet_p95": 152.5, "handover_rate": 0.243, "handover_p95": 90.0, "rlf": 18256.0, "stale_exposure": 2.77, "affected": 35.48, "network": 32.29 } } }, { "rate": 0.3, "speed": 120, "models": { "Jev-1.13.0": { "throughput": 0.145, "packet_p95": 159.8, "handover_rate": 0.466, "handover_p95": 85.0, "rlf": 19964.0, "stale_exposure": 1.88, "affected": 29.14, "network": 32.33 }, "SemIf-Qwen3.5-4B": { "throughput": 0.143, "packet_p95": 160.7, "handover_rate": 0.461, "handover_p95": 100.0, "rlf": 19736.0, "stale_exposure": 2.31, "affected": 34.31, "network": 35.85 }, "AnyJev-L0": { "throughput": 0.143, "packet_p95": 170.4, "handover_rate": 0.453, "handover_p95": 100.0, "rlf": 19740.0, "stale_exposure": 2.53, "affected": 33.46, "network": 33.73 }, "DeepSeek-V4.1-Flash": { "throughput": 0.149, "packet_p95": 146.3, "handover_rate": 0.463, "handover_p95": 85.0, "rlf": 19939.0, "stale_exposure": 3.69, "affected": 30.7, "network": 32.61 }, "GLM-5.3-Flash": { "throughput": 0.169, "packet_p95": 150.9, "handover_rate": 0.474, "handover_p95": 85.0, "rlf": 19982.0, "stale_exposure": 8.46, "affected": 28.34, "network": 31.62 }, "Qwen3.8-Flash": { "throughput": 0.156, "packet_p95": 147.4, "handover_rate": 0.469, "handover_p95": 89.9, "rlf": 19873.0, "stale_exposure": 15.48, "affected": 31.21, "network": 32.7 }, "Qwen3.5-4B-JSON": { "throughput": 0.141, "packet_p95": 167.2, "handover_rate": 0.448, "handover_p95": 100.0, "rlf": 20202.0, "stale_exposure": 2.65, "affected": 33.56, "network": 35.8 } } }, { "rate": 1, "speed": 3, "models": { "Jev-1.13.0": { "throughput": 0.198, "packet_p95": 122.6, "handover_rate": 0.014, "handover_p95": 88.0, "rlf": 5.0, "stale_exposure": 1.98, "affected": 18.6, "network": 17.35 }, "SemIf-Qwen3.5-4B": { "throughput": 0.197, "packet_p95": 123.8, "handover_rate": 0.018, "handover_p95": 95.0, "rlf": 5.0, "stale_exposure": 2.59, "affected": 18.51, "network": 17.17 }, "AnyJev-L0": { "throughput": 0.198, "packet_p95": 121.8, "handover_rate": 0.018, "handover_p95": 82.5, "rlf": 5.0, "stale_exposure": 2.64, "affected": 18.34, "network": 17.17 }, "DeepSeek-V4.1-Flash": { "throughput": 0.198, "packet_p95": 123.9, "handover_rate": 0.014, "handover_p95": 150.0, "rlf": 5.0, "stale_exposure": 4.25, "affected": 19.17, "network": 17.26 }, "GLM-5.3-Flash": { "throughput": 0.197, "packet_p95": 123.0, "handover_rate": 0.014, "handover_p95": 167.0, "rlf": 5.0, "stale_exposure": 9.42, "affected": 18.97, "network": 17.16 }, "Qwen3.8-Flash": { "throughput": 0.183, "packet_p95": 125.0, "handover_rate": 0.012, "handover_p95": 84.8, "rlf": 5.0, "stale_exposure": 15.62, "affected": 19.66, "network": 17.34 }, "Qwen3.5-4B-JSON": { "throughput": 0.199, "packet_p95": 123.6, "handover_rate": 0.016, "handover_p95": 96.0, "rlf": 5.0, "stale_exposure": 2.9, "affected": 18.79, "network": 17.26 } } }, { "rate": 1, "speed": 60, "models": { "Jev-1.13.0": { "throughput": 0.1, "packet_p95": 178.9, "handover_rate": 0.078, "handover_p95": 86.0, "rlf": 3249.0, "stale_exposure": 2.16, "affected": 34.0, "network": 30.5 }, "SemIf-Qwen3.5-4B": { "throughput": 0.098, "packet_p95": 181.0, "handover_rate": 0.074, "handover_p95": 92.8, "rlf": 3039.0, "stale_exposure": 2.67, "affected": 33.55, "network": 30.4 }, "AnyJev-L0": { "throughput": 0.098, "packet_p95": 185.3, "handover_rate": 0.075, "handover_p95": 92.2, "rlf": 3078.0, "stale_exposure": 2.8, "affected": 34.3, "network": 30.84 }, "DeepSeek-V4.1-Flash": { "throughput": 0.097, "packet_p95": 183.7, "handover_rate": 0.074, "handover_p95": 89.9, "rlf": 3255.0, "stale_exposure": 4.38, "affected": 34.58, "network": 31.7 }, "GLM-5.3-Flash": { "throughput": 0.097, "packet_p95": 181.5, "handover_rate": 0.074, "handover_p95": 120.0, "rlf": 3057.0, "stale_exposure": 9.65, "affected": 33.05, "network": 30.31 }, "Qwen3.8-Flash": { "throughput": 0.096, "packet_p95": 187.6, "handover_rate": 0.075, "handover_p95": 97.0, "rlf": 3203.0, "stale_exposure": 16.65, "affected": 34.49, "network": 31.28 }, "Qwen3.5-4B-JSON": { "throughput": 0.096, "packet_p95": 175.6, "handover_rate": 0.071, "handover_p95": 90.0, "rlf": 3399.0, "stale_exposure": 3.06, "affected": 33.83, "network": 31.23 } } }, { "rate": 1, "speed": 120, "models": { "Jev-1.13.0": { "throughput": 0.124, "packet_p95": 173.2, "handover_rate": 0.147, "handover_p95": 100.0, "rlf": 5033.0, "stale_exposure": 1.73, "affected": 32.41, "network": 32.88 }, "SemIf-Qwen3.5-4B": { "throughput": 0.132, "packet_p95": 187.5, "handover_rate": 0.149, "handover_p95": 90.0, "rlf": 4988.0, "stale_exposure": 2.16, "affected": 31.94, "network": 32.94 }, "AnyJev-L0": { "throughput": 0.121, "packet_p95": 168.4, "handover_rate": 0.148, "handover_p95": 94.7, "rlf": 5028.0, "stale_exposure": 2.33, "affected": 30.05, "network": 32.05 }, "DeepSeek-V4.1-Flash": { "throughput": 0.131, "packet_p95": 163.5, "handover_rate": 0.149, "handover_p95": 90.0, "rlf": 5162.0, "stale_exposure": 3.02, "affected": 31.01, "network": 31.9 }, "GLM-5.3-Flash": { "throughput": 0.119, "packet_p95": 158.5, "handover_rate": 0.143, "handover_p95": 97.0, "rlf": 4868.0, "stale_exposure": 8.8, "affected": 35.74, "network": 33.7 }, "Qwen3.8-Flash": { "throughput": 0.118, "packet_p95": 169.6, "handover_rate": 0.143, "handover_p95": 111.8, "rlf": 5156.0, "stale_exposure": 13.57, "affected": 35.46, "network": 34.24 }, "Qwen3.5-4B-JSON": { "throughput": 0.117, "packet_p95": 160.4, "handover_rate": 0.141, "handover_p95": 125.1, "rlf": 5237.0, "stale_exposure": 2.45, "affected": 33.58, "network": 33.62 } } } ], "headroom": [ 3.96, 2.17, 7.02 ], "fixed_delay_effect": [ 3.93, 1.22, 6.37 ] }, "kpm": { "Jev-1.13.0": { "p50": 1.347, "p95": 2.497 }, "SemIf-Qwen3.5-4B": { "p50": 1.249, "p95": 2.954 }, "AnyJev-L0": { "p50": 1.247, "p95": 4.483 }, "DeepSeek-V4.1-Flash": { "p50": 1.351, "p95": 2.835 }, "GLM-5.3-Flash": { "p50": 1.35, "p95": 3.712 }, "Qwen3.8-Flash": { "p50": 1.455, "p95": 3.539 }, "Qwen3.5-4B-JSON": { "p50": 1.247, "p95": 4.638 } } }, "edge": { "models": [ "Jev-1.13.0", "SemIf-Qwen3.5-4B", "Laya", "DeepSeek-V4.1-Flash", "GLM-5.3-Flash", "Qwen3.8-Flash", "Qwen3.5-4B-JSON" ], "load": { "1": { "completion": [ 0.94, 0.63, 0.117, 0.907, 0.917, 0.9, 0.193 ], "latency": [ 0.49, 1.28, 0.31, 0.67, 1.16, 1.76, 1.78 ], "supported": 300, "correct_count": [ 282, 189, 35, 272, 275, 270, 58 ] }, "2": { "completion": [ 0.953, 0.087, 0.117, 0.977, 0.823, 0.66, 0.003 ], "latency": [ 0.48, 1.96, 0.3, 0.77, 1.52, 1.9, 1.3 ], "supported": 300, "correct_count": [ 286, 26, 35, 293, 247, 198, 1 ] }, "4": { "completion": [ 0.937, 0.01, 0.117, 0.983, 0.807, 0.027, 0.003 ], "latency": [ 0.47, 1.85, 0.3, 0.67, 1.87, 1.96, 1.26 ], "supported": 300, "correct_count": [ 281, 3, 35, 295, 242, 8, 1 ] }, "8": { "completion": [ 0.943, 0.01, 0.117, 0.603, 0.077, 0.013, 0.003 ], "latency": [ 0.51, 1.8, 0.32, 1.89, 1.96, 1.87, 1.33 ], "supported": 300, "correct_count": [ 283, 3, 35, 181, 23, 4, 1 ] }, "16": { "completion": [ 0.907, 0.01, 0.093, 0.093, 0.023, 0.003, 0.003 ], "latency": [ 1.74, 1.9, 1.82, 1.87, 1.94, 1.65, 1.33 ], "supported": 300, "correct_count": [ 272, 3, 28, 28, 7, 1, 1 ] } }, "deadline": { "0.5": { "completion": [ 0.89, 0.0, 0.117, 0.377, 0.0, 0.0, 0.0 ], "latency": [ 0.46, null, 0.29, 0.49, null, null, null ], "supported": 300, "correct_count": [ 267, 0, 35, 113, 0, 0, 0 ] }, "1": { "completion": [ 0.927, 0.003, 0.117, 0.913, 0.727, 0.0, 0.0 ], "latency": [ 0.47, 0.89, 0.29, 0.83, 0.97, null, null ], "supported": 300, "correct_count": [ 278, 1, 35, 274, 218, 0, 0 ] }, "4": { "completion": [ 0.943, 0.027, 0.117, 0.987, 0.99, 0.05, 0.007 ], "latency": [ 0.48, 3.83, 0.3, 1.01, 1.45, 3.95, 2.89 ], "supported": 300, "correct_count": [ 283, 8, 35, 296, 297, 15, 2 ] }, "2": { "completion": [ 0.937, 0.01, 0.117, 0.983, 0.807, 0.027, 0.003 ], "latency": [ 0.47, 1.85, 0.3, 0.67, 1.87, 1.96, 1.26 ], "supported": 300, "correct_count": [ 281, 3, 35, 295, 242, 8, 1 ] } }, "cache": { "changing, on": { "completion": [ 0.943, 0.01, 0.117, 0.957, 0.77, 0.007, 0.003 ], "latency": [ 0.47, 1.65, 0.3, 0.8, 1.79, 1.93, 1.32 ], "supported": 300, "correct_count": [ 283, 3, 35, 287, 231, 2, 1 ] }, "repeated, off": { "completion": [ 1.0, 0.013, 0.113, 0.997, 0.78, 0.033, 0.003 ], "latency": [ 0.48, 1.89, 0.3, 0.7, 1.83, 1.86, 1.51 ], "supported": 300, "correct_count": [ 300, 4, 34, 299, 234, 10, 1 ] }, "repeated, on": { "completion": [ 1.0, 0.953, 0.113, 1.0, 1.0, 0.97, 0.577 ], "latency": [ 0.2, 0.19, 0.2, 0.2, 0.2, 0.19, 0.19 ], "supported": 300, "correct_count": [ 300, 286, 34, 300, 300, 291, 173 ] }, "changing, off": { "completion": [ 0.937, 0.01, 0.117, 0.983, 0.807, 0.027, 0.003 ], "latency": [ 0.47, 1.85, 0.3, 0.67, 1.87, 1.96, 1.26 ], "supported": 300, "correct_count": [ 281, 3, 35, 295, 242, 8, 1 ] } }, "ocr": { "steady": { "changing, off": { "completion": [ 0.511, 0.15, 0.256, 0.411, 0.45, 0.317, 0.011 ], "latency": [ 0.42, 1.68, 0.23, 1.07, 1.05, 1.84, 1.15 ], "supported": 180, "correct_count": [ 92, 27, 46, 74, 81, 57, 2 ] }, "changing, on": { "completion": [ 0.511, 0.039, 0.256, 0.506, 0.45, 0.278, 0.011 ], "latency": [ 0.48, 1.72, 0.22, 0.69, 1.57, 1.95, 1.2 ], "supported": 180, "correct_count": [ 92, 7, 46, 91, 81, 50, 2 ] }, "repeated, off": { "completion": [ 0.5, 0.011, 0.233, 0.494, 0.45, 0.322, 0.011 ], "latency": [ 0.47, 1.14, 0.22, 0.76, 1.32, 1.92, 1.83 ], "supported": 180, "correct_count": [ 90, 2, 42, 89, 81, 58, 2 ] }, "repeated, on": { "completion": [ 0.5, 0.5, 0.233, 0.5, 0.5, 0.5, 0.489 ], "latency": [ 0.13, 0.13, 0.13, 0.12, 0.13, 0.13, 0.13 ], "supported": 180, "correct_count": [ 90, 90, 42, 90, 90, 90, 88 ] } }, "bursty": { "changing, off": { "completion": [ 0.511, 0.056, 0.256, 0.467, 0.167, 0.039, 0.022 ], "latency": [ 0.46, 1.24, 0.25, 1.86, 1.95, 1.75, 1.54 ], "supported": 180, "correct_count": [ 92, 10, 46, 84, 30, 7, 4 ] }, "changing, on": { "completion": [ 0.511, 0.056, 0.256, 0.506, 0.211, 0.044, 0.022 ], "latency": [ 0.45, 1.26, 0.26, 1.36, 1.96, 1.98, 1.54 ], "supported": 180, "correct_count": [ 92, 10, 46, 91, 38, 8, 4 ] }, "repeated, off": { "completion": [ 0.5, 0.056, 0.233, 0.444, 0.111, 0.028, 0.028 ], "latency": [ 0.52, 1.84, 0.28, 1.87, 2.0, 1.69, 1.13 ], "supported": 180, "correct_count": [ 90, 10, 42, 80, 20, 5, 5 ] }, "repeated, on": { "completion": [ 0.5, 0.5, 0.233, 0.5, 0.5, 0.5, 0.5 ], "latency": [ 0.14, 0.16, 0.1, 0.14, 0.15, 0.13, 0.15 ], "supported": 180, "correct_count": [ 90, 90, 42, 90, 90, 90, 90 ] } } }, "ocr_reference": { "supported": 180, "unsupported": 60, "jev_compliant_timely": 180, "jev_correct_rejections": 60, "recognizer_correct": { "changing": 92, "repeated": 90 } } }, "sok": { "groups": { "Timing & load": [ "Named timing endpoint", "Specified metric denominator", "p95 or higher", "Deadline attainment", "Input load", "Queue / waiting" ], "Execution & checks": [ "Named execution location", "Network round trip included", "Format check reported", "Semantic check reported", "Final-state check reported", "Three gates distinguished" ], "Stability & reuse": [ "Stability evidence", "Non-learning comparison", "Public code pointer", "Public data pointer", "Explicit loop claim", "Supported loop claim" ] }, "reporting": { "Named timing endpoint": { "count": 92, "n": 139, "ci95_pct": [ 58.3, 74.1 ] }, "Specified metric denominator": { "count": 49, "n": 139, "ci95_pct": [ 27.3, 43.2 ] }, "p95 or higher": { "count": 9, "n": 139, "ci95_pct": [ 2.9, 10.8 ] }, "Deadline attainment": { "count": 4, "n": 139, "ci95_pct": [ 0.7, 5.8 ] }, "Input load": { "count": 48, "n": 139, "ci95_pct": [ 26.6, 42.4 ] }, "Queue / waiting": { "count": 6, "n": 139, "ci95_pct": [ 1.4, 7.9 ] }, "Stability evidence": { "count": 31, "n": 139, "ci95_pct": [ 15.8, 29.5 ] }, "Non-learning comparison": { "count": 48, "n": 139, "ci95_pct": [ 26.6, 42.4 ] }, "Explicit loop claim": { "count": 50, "n": 139, "ci95_pct": [ 28.1, 43.9 ] }, "Supported loop claim": { "count": 4, "n": 139, "ci95_pct": [ 0.7, 5.8 ] }, "Named execution location": { "count": 86, "n": 139, "ci95_pct": [ 54.0, 69.8 ] }, "Network round trip included": { "count": 23, "n": 139, "ci95_pct": [ 10.8, 23.0 ] }, "Format check reported": { "count": 71, "n": 139, "ci95_pct": [ 42.4, 59.0 ] }, "Semantic check reported": { "count": 133, "n": 139, "ci95_pct": [ 92.1, 98.6 ] }, "Final-state check reported": { "count": 54, "n": 139, "ci95_pct": [ 30.9, 46.8 ] }, "Three gates distinguished": { "count": 9, "n": 139, "ci95_pct": [ 2.9, 10.8 ] }, "Public code pointer": { "count": 34, "n": 139, "ci95_pct": [ 17.3, 31.7 ] }, "Public data pointer": { "count": 51, "n": 139, "ci95_pct": [ 28.8, 44.6 ] } }, "families": [ { "id": 1, "title": "LLNet: An Intent-Driven Approach to Instructing Softwarized Network Devices Using a Small Language Model", "interfaces": [ "G", "C" ], "workflow": "Intent JSON to confirmed switch rules", "path": "Intent handler; SDN/device manager. Request-driven JSON interpretation, user confirmation, and rule deployment on softwarized switches.", "checks": { "Observation state": "NE", "Feasibility": "Network Manager detects conflicts with installed intents.", "Candidate coverage": "Intent Converter checks named devices and supported scenario types." }, "note": "Human confirmation and syntax checks precede deployment.", "url": "https://doi.org/10.1109/TNSM.2025.3570017", "year": "2025" }, { "id": 2, "title": "Evaluating Agentic Configuration Repair for Computer Networks", "interfaces": [ "G" ], "workflow": "Offline configuration repair with feedback", "path": "Configuration-repair assistant. Offline iterative edits and diagnosis over Cornetto configurations; forwarding properties are simulated.", "checks": { "Observation state": "NE", "Feasibility": "Batfish/Config2Spec verification, used either only for final scoring or also as agent feedback depending on the condition.", "Candidate coverage": "NE" }, "note": "Verification evaluates the final configuration in all conditions and guides intermediate repairs only in the feedback condition.", "url": "https://arxiv.org/abs/2606.06212", "year": "2026" }, { "id": 3, "title": "Don't Mind the Gap: Bridging Network-Wide Objectives and Device-Level Configurations", "interfaces": [ "C" ], "workflow": "Policy compilation to BGP configurations", "path": "Routing-policy compiler. Operator policies compile into device-level BGP configurations.", "checks": { "Observation state": "NE", "Feasibility": "Propane compiler computes topology-constrained permitted paths and compiles routing preferences.", "Candidate coverage": "NE" }, "note": "", "url": "https://doi.org/10.1145/2934872.2934909", "year": "2016" }, { "id": 4, "title": "Intent-Based Infrastructure and Service Orchestration Using Agentic-AI", "interfaces": [ "S", "G" ], "workflow": "Agent plans to cluster/service deployment", "path": "Intent handler; infrastructure/service orchestrator. Agent plan and dispatch, Kubernetes cluster/resource changes, and network-service deployment.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The workflow uses cluster-state feedback and human approval for sensitive tools.", "url": "https://doi.org/10.1109/OJCOMS.2025.3600706", "year": "2025" }, { "id": 5, "title": "Software-Defined Network Assimilation: Bridging the Last Mile towards Centralized Network Configuration Management with NAssim", "interfaces": [ "S", "C" ], "workflow": "Offline device-model and parameter mapping", "path": "Device-model preparation and configuration mapping. Offline manual parsing, CLI/model validation, and top-k vendor-to-unified parameter mapping.", "checks": { "Observation state": "NE", "Feasibility": "NAssim Validator checks CLI syntax and hierarchy and tests snippets against device/configuration evidence.", "Candidate coverage": "NE" }, "note": "Parameter top-k recommendation is selection; validation concerns device-model consistency.", "url": "https://doi.org/10.1145/3544216.3544244", "year": "2022" }, { "id": 6, "title": "Towards End-to-End Network Intent Management with Large Language Models", "interfaces": [ "G" ], "workflow": "Business intent to RAN/core service configuration", "path": "Business/service intent handler. Natural-language demand becomes CFS, then RAN/core RFS configuration; translation is the measured boundary.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Resolver verification is architectural; evaluation measures translation.", "url": "https://arxiv.org/abs/2504.13589", "year": "2025" }, { "id": 7, "title": "Intent Assurance Using LLMs Guided by Intent Drift", "interfaces": [ "S", "G" ], "workflow": "Intent fulfillment and drift-triggered repair", "path": "Intent handler; assurance orchestrator. Intent/KPI extraction followed by collection-service fulfillment and drift-triggered repair.", "checks": { "Observation state": "NE", "Feasibility": "A separate LLM validator inspects the generated policy tree for omissions, ordering and attributes.", "Candidate coverage": "NE" }, "note": "Health/drift feedback triggers assurance and repair.", "url": "https://arxiv.org/abs/2402.00715", "year": "2024" }, { "id": 8, "title": "Network-Wide Configuration Synthesis", "interfaces": [ "C" ], "workflow": "Solver-based network configuration synthesis", "path": "Network configuration synthesizer. Datalog/SMT synthesis of OSPF, BGP, and static forwarding configurations.", "checks": { "Observation state": "NE", "Feasibility": "SyNET encodes specified forwarding requirements and protocol interactions into solver constraints.", "Candidate coverage": "The synthesis contract returns a satisfying network configuration if one exists, within its formal input model." }, "note": "", "url": "https://doi.org/10.1007/978-3-319-63390-9_14", "year": "2017" }, { "id": 9, "title": "NetComplete: Practical Network-Wide Configuration Synthesis with Autocompletion", "interfaces": [ "C" ], "workflow": "Routing-sketch completion before activation", "path": "Network configuration synthesizer. OSPF, BGP, and static-route sketch completion; synthesis runtimes are distinct from activation.", "checks": { "Observation state": "NE", "Feasibility": "NetComplete solves symbolic sketch constraints and checks BGP announcement-preference consistency.", "Candidate coverage": "NE" }, "note": "Feasibility is conditional on the operator sketch and modeled routing requirements.", "url": "https://www.usenix.org/conference/nsdi18/presentation/el-hassany", "year": "2018" }, { "id": 10, "title": "NetIntent: Leveraging Large Language Models for End-to-End Intent-Based SDN Automation", "interfaces": [ "G", "C" ], "workflow": "Flow generation, installation and drift repair", "path": "Intent handler; SDN controller. Flow JSON generation, LLM conflict detection, deterministic resolution/installation, and drift repair.", "checks": { "Observation state": "NE", "Feasibility": "LLM compares new and installed flow rules; controller policy resolves conflicts or escalates. Structural validation is a separate non-LLM component.", "Candidate coverage": "NE" }, "note": "The workflow escalates untranslatable outputs.", "url": "https://doi.org/10.1109/OJCOMS.2025.3642642", "year": "2025" }, { "id": 11, "title": "Intent Engine: Natural-Language Intent Translation for Intent-Driven Orchestration in the Compute Continuum", "interfaces": [ "G" ], "workflow": "State-grounded service-objective construction", "path": "Intent handler above compute-continuum orchestrator. State-grounded SLO construction.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Deterministic snapshot encoding, periodic monitoring and schema checks support SLO construction.", "url": "https://doi.org/10.1016/j.sysarc.2026.103938", "year": "2026" }, { "id": 12, "title": "Hey, Lumi! Using Natural Language for Intent-Based Network Management", "interfaces": [ "G", "S", "C" ], "workflow": "Confirmed intent to compiler/controller", "path": "Intent handler; SDN compiler/controller. NER/assembly and user-confirmed Nile, Merlin compilation/deployment, plus a separate exploratory contradiction classifier.", "checks": { "Observation state": "NE", "Feasibility": "Random Forest flags intent-pair contradictions in the reported exploratory task; formal policy verification is prospective.", "Candidate coverage": "NE" }, "note": "The exploratory contradiction classifier is evaluated in Sec. 10.1.", "url": "https://www.usenix.org/conference/atc21/presentation/jacobs", "year": "2021" }, { "id": 13, "title": "ITBench: Evaluating AI Agents across Diverse Real-World IT Automation Tasks", "interfaces": [ "G" ], "workflow": "Network-fault diagnosis and mitigation in ITBench", "path": "SRE/assurance agent. Network-fault root-cause reports and mitigation actions in the ITBench SRE environment.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Reported aggregate timing covers the full SRE task set; network-fault tasks form a subset.", "url": "https://arxiv.org/abs/2502.05352", "year": "2025" }, { "id": 14, "title": "JAUNT: Joint Alignment of User Intent and Network State for QoE-Centric LLM Tool Routing", "interfaces": [ "S" ], "workflow": "Per-request tool selection with latency prediction", "path": "Service/tool router. Per-request MCP tool selection combines semantic matching, a user profile, and predicted network latency.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "EWMA latency predictions and candidate retrieval feed the selector.", "url": "https://arxiv.org/abs/2510.18550", "year": "2025" }, { "id": 15, "title": "Fast Intent-Driven Service Orchestration with Jev for 6G Edge Networks", "interfaces": [ "S", "G", "C" ], "workflow": "Versioned intent contracts to edge placement", "path": "Intent handler; edge service scheduler. Versioned contract activation and numerical placement; modeled NR access and modeled/real OCR service paths remain distinct.", "checks": { "Observation state": "Controller versions contracts and cache identity; numerical scheduler reads current queues/capacity after interpretation.", "Feasibility": "Shared execution layer checks placement permissions; numerical scheduler scores permitted placements under current state.", "Candidate coverage": "NE" }, "note": "This classification describes the 6G version included in the review corpus. The RIC experiments analyzed in Sec. VII come from a subsequent code and data release. Observation handling here covers state reads and versioning.", "url": "https://arxiv.org/abs/2609.23136v1", "year": "2026" }, { "id": 16, "title": "Replacing Large Language Models with Jev Decision Models for Low-Latency Edge Service Orchestration", "interfaces": [ "S", "G", "C" ], "workflow": "Intent interpretation to edge admission/execution", "path": "Intent handler; admission controller and edge scheduler. Four-field interpretation followed by shared admission and modeled or real two-node OCR execution.", "checks": { "Observation state": "Scheduler reads worker/queue state after interpretation; interpretation reuse is keyed by request text and extraction policy.", "Feasibility": "Scheduler filters by service, placement, queue capacity, and predicted deadline; observed completion can still fail after admission.", "Candidate coverage": "Scheduler rejects when no eligible node has a predicted completion within the deadline." }, "note": "", "url": "https://arxiv.org/abs/2609.22753", "year": "2026" }, { "id": 17, "title": "Towards Intent-Based Network Management: Large Language Models for Intent Extraction in 5G Core Networks", "interfaces": [ "S", "G" ], "workflow": "Intent extraction; proposed core placement", "path": "Core-network intent handler (functional placement). Six-class multi-label intent extraction with explanations; activation is future work.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The evaluation measures intent extraction; core-network placement is architectural.", "url": "https://doi.org/10.1109/DRCN60692.2024.10539172", "year": "2024" }, { "id": 18, "title": "Advanced LLM-Enhanced Intent-Based 5G Network Management using Dynamic Semantic Routes", "interfaces": [ "S", "G" ], "workflow": "Static route selection and dynamic parameters", "path": "Core-network intent router/handler. Encoder-based static routing and generated dynamic-route parameters; early and expanded report scopes are retained.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The router includes None among its semantic route options.", "url": "https://arxiv.org/abs/2608.22644v1", "year": "2026" }, { "id": 19, "title": "NLP Powered Intent Based Network Management for Private 5G Networks", "interfaces": [ "S" ], "workflow": "Registered workflow/model to slice/NFV action", "path": "Intent handler; private-5G slice/NFV orchestrator. Registered workflow/model matching leads to slice configuration, LoS classification, or NS/VIM service instantiation.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Matching selects a registered operation/model; its workflow specifies the execution actions and parameters.", "url": "https://doi.org/10.1109/ACCESS.2023.3265894", "year": "2023" }, { "id": 20, "title": "Chat-Driven Optimal Management for Virtual Network Services", "interfaces": [ "S", "G", "C" ], "workflow": "Intent parameters to placement/routing solver", "path": "Intent handler; virtual-network optimizer. Chat parameters drive ILP placement/routing and an explicit accepted-request-subset decision.", "checks": { "Observation state": "NE", "Feasibility": "ILP enforces link, compute, latency, and placement constraints parameterized by interpreted intent.", "Candidate coverage": "Optimizer arbitrates an acceptable request subset under joint resource constraints in the expanded report." }, "note": "BERT/SVM labels and generated Llama JSON are alternative interpreters supplying parameters to the solver.", "url": "https://doi.org/10.1109/icc52391.2025.11160902", "year": "2025" }, { "id": 21, "title": "What Do LLMs Need to Synthesize Correct Router Configurations?", "interfaces": [ "G" ], "workflow": "Configuration translation/generation with feedback", "path": "Router-configuration translator/synthesizer. Cross-vendor translation and multi-router BGP policy generation, refined with verification feedback.", "checks": { "Observation state": "NE", "Feasibility": "Campion checks configuration equivalence; Batfish checks the scoped routing policy. Syntax and semantic feedback are iterated separately.", "Candidate coverage": "NE" }, "note": "Campion and Batfish verify modeled routing/forwarding behavior.", "url": "https://doi.org/10.1145/3626111.3628194", "year": "2023" }, { "id": 22, "title": "LLM-Based AI Agent for Virtual Network Function Deployment", "interfaces": [ "G" ], "workflow": "Deployment-workflow generation and twin testing", "path": "VNF deployment orchestrator. MOP-to-Python/Ansible workflow generation and iterative repair in a network digital twin.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Digital-twin execution tests generated workflows and returns logs for iterative repair; validation covers those executions.", "url": "https://doi.org/10.1007/s10922-026-10078-x", "year": "2026" }, { "id": 23, "title": "Toward Intent-Based Network Automation for Smart Environments: A Healthcare 4.0 Use Case", "interfaces": [ "G" ], "workflow": "Intent entity extraction; proposed SDN activation", "path": "Intent handler above SDN (proposed placement). BERT extracts user, goal, action, target, and time-frame entities; SDN activation/assurance remains architectural.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Token labels extract open entity values.", "url": "https://doi.org/10.1109/ACCESS.2023.3338189", "year": "2023" }, { "id": 24, "title": "Practical Intent-Driven Routing Configuration Synthesis", "interfaces": [ "C" ], "workflow": "BGP compilation, staging and activation", "path": "Routing-policy compiler and activation controller. Aura compiles per-switch BGP policies, stages configurations, and changes active routing policy after convergence checks.", "checks": { "Observation state": "NE", "Feasibility": "Routing-policy validation uses container-based switch emulation before deployment.", "Candidate coverage": "NE" }, "note": "Compilation, operator review, deployment, and convergence are distinct boundaries.", "url": "https://www.usenix.org/conference/nsdi23/presentation/ramanathan", "year": "2023" }, { "id": 25, "title": "Snowcap: Synthesizing Network-Wide Configuration Updates", "interfaces": [ "C" ], "workflow": "Safe update ordering with convergence waits", "path": "Configuration-update planner/controller. Snowcap synthesizes an ordered update plan, then applies commands with convergence waits.", "checks": { "Observation state": "NE", "Feasibility": "LTL safety requirements constrain every intermediate update state; a forwarding supergraph overapproximates convergence states.", "Candidate coverage": "NE" }, "note": "The computed update order enforces safety across intermediate states.", "url": "https://doi.org/10.1145/3452296.3472915", "year": "2021" }, { "id": 26, "title": "Safely and Automatically Updating In-Network ACL Configurations with Intent Language", "interfaces": [ "C" ], "workflow": "ACL verification, repair and migration", "path": "ACL verification and update controller. Jinjing checks, repairs, migrates, or generates ACL updates under scoped reachability intent.", "checks": { "Observation state": "NE", "Feasibility": "ACL reachability predicates and placement constraints are checked/solved with formal constraints and SMT.", "Candidate coverage": "NE" }, "note": "", "url": "https://doi.org/10.1145/3341302.3342088", "year": "2019" }, { "id": 27, "title": "FaulT-Bench: Towards Benchmarking Network Troubleshooting LLM Agents under Unreliable User Tickets", "interfaces": [ "G" ], "workflow": "Tool-driven diagnosis and proposed remediation", "path": "Network diagnostic assistant. Agents inspect a live emulated network, return diagnosis, and propose remediation under unreliable user tickets.", "checks": { "Observation state": "The diagnostic agent is tasked with checking ticket premises against tool-observed network state; false-premise and wrong-device tickets test this responsibility.", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Evaluation tests the agent's premise checks, diagnoses and proposed fixes.", "url": "https://arxiv.org/abs/2608.27021", "year": "2026" }, { "id": 28, "title": "Intent-Based Network Configuration Using Large Language Models", "interfaces": [ "G", "C" ], "workflow": "Confirmed intent translation to NFV deployment", "path": "Intent handler; NFV orchestrator. VNF/SFC or formal-policy JSON translation, human confirmation, and NI-NFVO deployment.", "checks": { "Observation state": "NE", "Feasibility": "NFV-Intent and NI-MANO check infrastructure-supported resource flavors and user capability before deployment.", "Candidate coverage": "NE" }, "note": "Verification ownership is task-specific; the early and expanded translation datasets remain separate evidence scopes.", "url": "https://doi.org/10.1109/noms59830.2024.10575237", "year": "2024" }, { "id": 29, "title": "NetConfEval: Can LLMs Facilitate Network Configuration?", "interfaces": [ "G" ], "workflow": "Task-specific policy, API and routing generation", "path": "Intent/configuration handler; routing-program assistant. JSON/Datalog policies, API calls with arguments, routing code, FRR/RIFT/P4 configuration, and a scoped live BGP change.", "checks": { "Observation state": "NE", "Feasibility": "LLM conflict diagnosis is one scoped output; syntax/tests and network verifiers provide task-specific feedback. Operator confirmation remains needed for language-to-intent correctness.", "Candidate coverage": "NE" }, "note": "API calls construct arguments as well as selecting a function name. Evaluation and activation boundaries vary across benchmark tasks.", "url": "https://arxiv.org/abs/2309.06342v1", "year": "2023" }, { "id": 30, "title": "A Network Arena for Benchmarking AI Agents on Network Troubleshooting", "interfaces": [ "G" ], "workflow": "Tool-driven anomaly localization and diagnosis", "path": "Network diagnostic assistant. NIKA uses tool-driven network observation for anomaly detection, localization, and root-cause reports.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Mitigation safety and execution are explicitly future work. The early four-switch demonstration and expanded benchmark remain separate scopes.", "url": "https://arxiv.org/abs/2507.01997v2", "year": "2025" }, { "id": 31, "title": "INTA: Intent-Based Translation for Network Configuration with LLM Agents", "interfaces": [ "G" ], "workflow": "Retrieved manuals to multivendor CLI", "path": "Cross-vendor configuration translator. INTA combines manual retrieval, target-CLI generation, syntax correction, and a semantic-consistency report.", "checks": { "Observation state": "NE", "Feasibility": "Command-tree matching checks syntax/view structure; an LLM checks source-target semantic equivalence and refines disagreements.", "Candidate coverage": "NE" }, "note": "The LLM judges source-target semantic consistency.", "url": "https://doi.org/10.1109/ICNP65844.2025.11192391", "year": "2025" }, { "id": 32, "title": "NetLLM: Adapting Large Language Models for Networking", "interfaces": [ "S" ], "workflow": "Learned bitrate and cluster-action selection", "path": "Application/networking controller; cluster scheduler. A learned networking head selects video bitrate or a scheduling stage and executor count.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The decision interfaces cover bitrate and scheduling actions. Prediction, training, and aggregate performance measurements have separate evaluation scopes.", "url": "https://doi.org/10.1145/3651890.3672268", "year": "2024" }, { "id": 33, "title": "NetArena: Dynamic Benchmarks for AI Agents in Network Automation", "interfaces": [ "G" ], "workflow": "Agent planning/repair with environment feedback", "path": "Network planning and repair agent. NetArena agents modify topology plans, emulated routing configuration, or Kubernetes network policies.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Environment-grounded evaluation provides feedback on agent actions. The review consolidates the NetPress alias under this family.", "url": "https://arxiv.org/abs/2506.03231", "year": "2025" }, { "id": 34, "title": "OmniPlan: An Adaptive Framework for Timely and Near-Optimal Network Planning Optimization", "interfaces": [ "G", "S", "C" ], "workflow": "Intent preferences to network-planning experts", "path": "Intent handler; network-planning optimizer. Preference extraction and expert selection feed weight adaptation and device-placement/link-routing optimization.", "checks": { "Observation state": "NE", "Feasibility": "MIP experts compile explicit planning constraints; fast experts use scoped resource-constrained heuristics.", "Candidate coverage": "NE" }, "note": "Learned weight adaptation constructs continuous values. Semantic validation is proposed future work.", "url": "https://arxiv.org/abs/2606.18105", "year": "2026" }, { "id": 35, "title": "OSDF: An Intent-based Software Defined Network Programming Framework", "interfaces": [ "C" ], "workflow": "Policies to SDN paths and conflict advice", "path": "Intent-based SDN controller. OSDF maps policies to paths/meters and computes conflict classifications and repair recommendations.", "checks": { "Observation state": "NE", "Feasibility": "Policy conflict module applies explicit rules for redundancy, shadowing, generalization, correlation, and overlap; recommendations go to the administrator.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/1807.02205v1", "year": "2018" }, { "id": 36, "title": "VIVoNet: Visually-represented, Intent-based, Voice-assisted Networking", "interfaces": [ "G", "C" ], "workflow": "Confirmed voice intent to OpenFlow rules", "path": "Voice intent handler; SDN controller. Alexa extracts operation/endpoints and confirms intent; deterministic path computation installs OpenFlow rules.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Successful rule push marks the activation acknowledgement.", "url": "https://arxiv.org/abs/1904.03228v1", "year": "2019" }, { "id": 37, "title": "GP4P4: Enabling Self-Programming Networks", "interfaces": [ "C" ], "workflow": "Behavior-guided search for P4 programs", "path": "Programmable-data-plane synthesizer. Genetic search constructs P4 programs against packet-behavior rules.", "checks": { "Observation state": "NE", "Feasibility": "Program fitness evaluates the specified behavioral examples/rules during search.", "Candidate coverage": "NE" }, "note": "Search stops on a valid program or at the generation limit.", "url": "https://arxiv.org/abs/1910.00967v1", "year": "2019" }, { "id": 38, "title": "Refining Network Intents for Self-Driving Networks", "interfaces": [ "G", "C" ], "workflow": "Confirmed intent to NFV service-chain deployment", "path": "Intent handler; NFV service-chain deployer. Entity extraction and seq2seq Nile generation precede user confirmation and SONATA-NFV deployment.", "checks": { "Observation state": "NE", "Feasibility": "Intent Deployer asserts consistency with existing network configuration and warns the operator before deployment.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2008.05509v1", "year": "2020" }, { "id": 39, "title": "The Role of Intent-Based Networking in ICT Supply Chains", "interfaces": [ "C" ], "workflow": "Conceptual access-policy parsing and refinement", "path": "Supply-chain access-policy handler (conceptual). Controlled language is parsed with regex/vocabulary rules and refined using organizational access policies.", "checks": { "Observation state": "NE", "Feasibility": "Proposed policy configurator resolves authorization conflicts using organizational scope/hierarchy before calling the controller.", "Candidate coverage": "NE" }, "note": "The paper presents conceptual examples.", "url": "https://arxiv.org/abs/2105.05179v1", "year": "2021" }, { "id": 40, "title": "Intent-based Network Management and Orchestration for Smart Distribution Grids", "interfaces": [ "U" ], "workflow": "Conceptual smart-grid intent to network slices", "path": "Intent/service orchestrator; slice MANO (conceptual). Smart-grid intent and SLA become a service request, GST/NEST profile, and proposed network-slice resource configuration.", "checks": { "Observation state": "NE", "Feasibility": "The proposed architecture assigns SLA validation and profile trade-off checks to intent/service orchestration.", "Candidate coverage": "NE" }, "note": "Functional transformations are specified without enough implementation detail to choose one decision-interface class.", "url": "https://arxiv.org/abs/2105.05594v1", "year": "2021" }, { "id": 41, "title": "Mission-Critical Public Safety Networking: An Intent-Driven Service Orchestration Perspective", "interfaces": [ "C" ], "workflow": "Controlled intent to simulated public-safety service", "path": "Intent handler; public-safety service orchestrator. Controlled-language processing maps requests to LTE/PTT service templates, followed by ns-3 access and voice-service execution.", "checks": { "Observation state": "NE", "Feasibility": "The architecture assigns resource, subscription, and conflict checks to intent validation using telemetry, with user dialogue on failure.", "Candidate coverage": "NE" }, "note": "The paper reports an architectural validation design and simulated service outcomes.", "url": "https://arxiv.org/abs/2205.03932v1", "year": "2022" }, { "id": 42, "title": "Experimental Demonstration of Partially Disaggregated Optical Network Control Using the Physical Layer Digital Twin", "interfaces": [ "C" ], "workflow": "Optical feasibility to lightpath deployment", "path": "Optical path computation and SDN controller. Digital-twin optical tuning and modulation feasibility feed route/spectrum assignment and ONOS/NETCONF lightpath deployment.", "checks": { "Observation state": "NE", "Feasibility": "L-PCE and RSA restrict modulation/path/channel combinations by optical feasibility and wavelength continuity.", "Candidate coverage": "NE" }, "note": "Optical feasibility and post-configuration transmission checks are tied to the measured photonics setup.", "url": "https://arxiv.org/abs/2212.11874v1", "year": "2022" }, { "id": 43, "title": "Knowledge-based Intent Modeling for Next Generation Cellular Networks", "interfaces": [ "C" ], "workflow": "Knowledge-based intent to simulated services", "path": "Knowledge-based intent handler; service orchestrator. KG queries form service/network intent; ns-3 deployment and KPI-threshold compliance reporting follow.", "checks": { "Observation state": "NE", "Feasibility": "Network-layer validation compares resource requirements with available resources and service models before orchestration.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2302.08544v2", "year": "2023" }, { "id": 44, "title": "Grooming Connectivity Intents in IP-Optical Networks Using Directed Acyclic Graphs", "interfaces": [ "C" ], "workflow": "IP-optical intent compilation in simulation", "path": "IP-optical intent compiler/orchestrator. Graph algorithms groom connectivity intents into shared path/modulation/spectrum resources.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "SAP/JML/LDJML are alternative algorithms for the same compilation responsibility; the reported evaluation is simulation.", "url": "https://arxiv.org/abs/2304.09711v1", "year": "2023" }, { "id": 45, "title": "An Intent-based Framework for Vehicular Edge Computing", "interfaces": [ "C" ], "workflow": "Constrained vehicular-edge service mapping", "path": "Vehicular-edge intent orchestrator. Declarative requests are mapped to constrained nodes/links, then installed or reconfigured using ONOS/Mininet.", "checks": { "Observation state": "NE", "Feasibility": "PAI/LAM heuristics map CPU, memory, location, bandwidth, latency and priority constraints in the request/network models.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2304.09916v1", "year": "2023" }, { "id": 46, "title": "Wireless Multi-Agent Generative AI: From Connected Intelligence to Collective Intelligence", "interfaces": [ "G" ], "workflow": "Distributed transmit-power proposals", "path": "Distributed wireless power controller. GPT-4 agents propose transmit powers using prior-round actions in a four-user power-game case.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The four-user case evaluates constructed transmit-power actions.", "url": "https://arxiv.org/abs/2307.02757v1", "year": "2023" }, { "id": 47, "title": "Direct-Conflict Resolution in Intent-Driven Autonomous Networks", "interfaces": [ "C" ], "workflow": "Modeled RAN tilt-conflict resolution", "path": "RAN intent-conflict resolver. Bargaining methods choose antenna tilt to reconcile coverage/capacity utility in an ns-3 LTE scenario.", "checks": { "Observation state": "NE", "Feasibility": "Explicit feasible utility sets and bargaining/Jain-fairness selection resolve the modeled two-intent parameter conflict.", "Candidate coverage": "NE" }, "note": "Bargaining resolves the modeled coverage/capacity utility conflict.", "url": "https://arxiv.org/abs/2401.08341v1", "year": "2024" }, { "id": 48, "title": "Intent Profiling and Translation Through Emergent Communication", "interfaces": [ "S" ], "workflow": "Device-intent symbols to QoS slices", "path": "Device intent encoder; network slice allocator. MAPPO maps device intent to discrete symbols and then maps messages to QoS slices.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Constraints are specified in the optimization problem.", "url": "https://arxiv.org/abs/2402.02768v1", "year": "2024" }, { "id": 49, "title": "PreConfig: A Pretrained Model for Automating Network Configuration", "interfaces": [ "G" ], "workflow": "Configuration generation and vendor translation", "path": "Device-configuration assistant. PreConfig generates configurations/commands, translates vendors, and explains device configuration semantics.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Evaluation concerns configuration artifacts and their explanations.", "url": "https://arxiv.org/abs/2403.09369v1", "year": "2024" }, { "id": 50, "title": "Predictive Intent Maintenance with Intent Drift Detection in Next Generation Network", "interfaces": [ "S", "C" ], "workflow": "Time-series monitoring of intent drift", "path": "Intent-assurance monitor. Predictive models and a greedy comparator flag intent drift from monitored throughput time series.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The monitor detects KPI drift in throughput time series.", "url": "https://arxiv.org/abs/2404.15091v1", "year": "2024" }, { "id": 51, "title": "Knowledge Graph Embedding in Intent-Based Networking", "interfaces": [ "G", "S" ], "workflow": "Entity/KG completion to intent templates", "path": "Knowledge-based intent handler. Entity extraction and KG completion form intent templates; a learned triple classifier outputs validity.", "checks": { "Observation state": "NE", "Feasibility": "KG2E performs learned plausibility classification of each intent triple as valid/invalid.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2405.07850v1", "year": "2024" }, { "id": 52, "title": "LLM-Based Intent Processing and Network Optimization Using Attention-Based Hierarchical Reinforcement Learning", "interfaces": [ "G", "S", "C" ], "workflow": "Intent validity to hierarchical RAN control", "path": "Intent handler; hierarchical RAN controller. KPI extraction and validity checks drive application selection, traffic steering, sleep, beam, power, and handover policies.", "checks": { "Observation state": "NE", "Feasibility": "Traffic prediction and explicit QoS-drift thresholds determine intent validity before hierarchical control.", "Candidate coverage": "NE" }, "note": "The validity rule uses QoS drift; the reported system result jointly covers the selected applications.", "url": "https://arxiv.org/abs/2406.06059v2", "year": "2024" }, { "id": 53, "title": "Online Learning for Autonomous Management of Intent-based 6G Networks", "interfaces": [ "S" ], "workflow": "Parent-bandit selection of joint service priorities", "path": "Intent-conflict resolver. A parent bandit combines service-agent predictions into joint service-priority choices.", "checks": { "Observation state": "NE", "Feasibility": "A learned parent-agent policy chooses among priority combinations to resolve modeled resource conflicts.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2407.17767v1", "year": "2024" }, { "id": 54, "title": "Large Language Models for Zero Touch Network Configuration Management", "interfaces": [ "S", "G" ], "workflow": "Intent to iterative verified configuration", "path": "Intent handler; configuration orchestrator. LLM intent classification/translation feed iterative device-configuration generation with verification feedback.", "checks": { "Observation state": "NE", "Feasibility": "A separate Batfish verifier checks generated configuration behavior; orchestrator coordinates feedback and regeneration.", "Candidate coverage": "NE" }, "note": "Reported time spans the whole pipeline through final generation and verification.", "url": "https://arxiv.org/abs/2408.13298v1", "year": "2024" }, { "id": 55, "title": "CAIP: Detecting Router Misconfigurations with Context-Aware Iterative Prompting of LLMs", "interfaces": [ "G" ], "workflow": "Retrieved configuration to error reports", "path": "Configuration diagnostic assistant. CAIP iteratively requests relevant configuration context and generates error reports/explanations.", "checks": { "Observation state": "NE", "Feasibility": "The LLM diagnoses syntax and dependency conflicts using progressively retrieved configuration context.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2411.14283v1", "year": "2024" }, { "id": 56, "title": "Poster: Could Large Language Models Perform Network Management?", "interfaces": [ "S" ], "workflow": "Wireless-mesh action recommendation", "path": "Wireless-mesh management assistant. Zero-shot model recommendations select a management action for observed events in a three-node scenario.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Evaluation measures management recommendations.", "url": "https://arxiv.org/abs/2411.16232v1", "year": "2024" }, { "id": 57, "title": "End-to-End Edge AI Service Provisioning Framework in 6G ORAN", "interfaces": [ "G", "S" ], "workflow": "Intent/model selection to service and xApp creation", "path": "Intent handler; service orchestrator; PCF interaction; xApp construction. Developer-intent profiling and model recommendation lead to API-service deployment, PCF policy interaction, and a generated QoS-monitoring xApp.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The prototype uses OAI/FlexRIC components for LLM-assisted provisioning.", "url": "https://arxiv.org/abs/2503.11933v1", "year": "2025" }, { "id": 58, "title": "RAG-Enabled Intent Reasoning for Application-Network Interaction", "interfaces": [ "G" ], "workflow": "Retrieved context to structured network intent", "path": "Application-network intent handler. RAG refines service scenarios and constructs structured network intent from retrieved context.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Intent validation and policy mapping are architectural components beyond the measured translation layer.", "url": "https://arxiv.org/abs/2505.09339v2", "year": "2025" }, { "id": 59, "title": "ALLSTaR: Automated LLM-Driven Scheduler Generation and Testing for Intent-Based RAN", "interfaces": [ "G", "C" ], "workflow": "Lua scheduler construction before slot execution", "path": "Scheduler construction assistant; RAN device-side scheduler/dApp. LLMs generate/review Lua schedulers; generated code performs per-slot PRB allocation and intent-composed scheduling.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "LLM construction/review precedes runtime scheduling; scheduler construction uses a pre-tested component library.", "url": "https://arxiv.org/abs/2505.18389v4", "year": "2025" }, { "id": 60, "title": "Intelligent Channel Allocation for IEEE 802.11be Multi-Link Operation: When MAB Meets LLM", "interfaces": [ "S", "C" ], "workflow": "Wi-Fi channel initialization followed by search", "path": "Wi-Fi channel allocator. LLM channel initialization is followed by BAI-MCTS search over remaining station configurations.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The carrier-sensing feasibility model supports throughput evaluation.", "url": "https://arxiv.org/abs/2506.04594v1", "year": "2025" }, { "id": 61, "title": "Toward an Intent-Based and Ontology-Driven Autonomic Security Response in Security Orchestration Automation and Response", "interfaces": [ "S", "U" ], "workflow": "Proposed intent choice; implemented enforcement", "path": "Security intent decision/enforcement agents. IDA selects persistent/transient mitigation intents; IEA maps security intent to Kubernetes filtering configuration.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "Ontology queries restrict candidate techniques to instantiated, operable artifacts with compatible defensive properties." }, "note": "IDA planning is a design assumption; the implementation example covers low-level enforcement.", "url": "https://arxiv.org/abs/2507.12061v1", "year": "2025" }, { "id": 62, "title": "Intent-Based Network for RAN Management with Large Language Models", "interfaces": [ "S", "G" ], "workflow": "History-based transmit-power strategy over O1", "path": "RAN management over O1. History analysis informs generated TxPower strategy/configuration, applied using NETCONF and assessed with performance-management data.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The prompt supplies configuration limits for O1 management.", "url": "https://arxiv.org/abs/2507.14230v2", "year": "2025" }, { "id": 63, "title": "A Novel Integrated Architecture for Intent Based Approach and Zero Touch Networks", "interfaces": [ "G", "S" ], "workflow": "Intent translation to learned throughput control", "path": "Intent handler; throughput controller. LLM/Nile translation and conflict checks feed a Q-learning traffic-shaping controller in an OAI environment.", "checks": { "Observation state": "NE", "Feasibility": "The intent translation component includes a basic check against previously configured intents; implementation/coverage is only briefly described.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2509.21026v1", "year": "2025" }, { "id": 64, "title": "Understanding Network Behaviors through Natural Language Question-Answering", "interfaces": [ "G", "C" ], "workflow": "Fact graph and generated network queries", "path": "Network behavior query assistant. Configuration is converted to a fact-graph representation; generated Python/Datalog queries are executed to answer network questions.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The generated query defines the network question evaluated by routing simulation.", "url": "https://arxiv.org/abs/2510.21894v1", "year": "2025" }, { "id": 65, "title": "A Secured Intent-Based Networking (sIBN) with Data-Driven Time-Aware Intrusion Detection", "interfaces": [ "S" ], "workflow": "Malicious-intent classification at ingress", "path": "Intent ingress/security filter. XGBoost flags malicious or tampered intent before downstream processing.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2511.05133v1", "year": "2025" }, { "id": 66, "title": "5G Network Automation Using Local Large Language Models and Retrieval-Augmented Generation", "interfaces": [ "S", "G" ], "workflow": "5G request classification to command generation", "path": "5G command intent handler. Local LLM request classification selects a category; retrieved examples guide executable command generation.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2511.21084v1", "year": "2025" }, { "id": 67, "title": "Efficient Asynchronous Federated Evaluation with Strategy Similarity Awareness for Intent-Based Networking in Industrial Internet of Things", "interfaces": [ "G", "S", "C" ], "workflow": "Strategies to predicted deployment approval", "path": "Intent translator; deployment evaluator; network controller. Structured strategy generation is followed by learned deployability approval and command mapping/deployment.", "checks": { "Observation state": "NE", "Feasibility": "Federated predictive evaluators estimate strategy deployability/performance before the controller installs a configuration.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2512.20627v2", "year": "2025" }, { "id": 68, "title": "Network Self-Configuration based on Fine-Tuned Small Language Models", "interfaces": [ "S", "G" ], "workflow": "Iterative configuration generation with verification", "path": "Configuration-generation agent. Intent classification and step planning lead to iterative Cisco configuration generation and a verification report.", "checks": { "Observation state": "NE", "Feasibility": "The Verifier is assigned syntax, dependency and objective checks with regeneration feedback.", "Candidate coverage": "NE" }, "note": "The verification module's implementation is unspecified. Reported time spans the whole configuration-generation pipeline.", "url": "https://arxiv.org/abs/2512.02861", "year": "2025" }, { "id": 69, "title": "Graph-Symbolic Policy Enforcement and Control (G-SPEC): A Neuro-Symbolic Framework for Safe Agentic AI in 5G Autonomous Networks", "interfaces": [ "G", "C" ], "workflow": "Language-planned actions through graph admission", "path": "5G action planner; controller admission guard. A language planner proposes actions; SHACL gates a hypothetical next-state graph before Open5GS orchestration.", "checks": { "Observation state": "A SHACL/SPARQL freshness guard rejects stale graph entities by lastUpdated age and requires refreshed telemetry.", "Feasibility": "G-SPEC checks topology, resource, operational-state, and bounded state-change policies before execution.", "Candidate coverage": "NE" }, "note": "Evaluation reports policy examples and tests of the graph guards.", "url": "https://arxiv.org/abs/2512.20275v1", "year": "2025" }, { "id": 70, "title": "Agentic AI Empowered Intent-Based Networking for 6G", "interfaces": [ "S", "G" ], "workflow": "RAN/core recommendations to slice provisioning", "path": "Cross-domain slice orchestrator; RAN/core specialists. An orchestrator combines RAN band/sector and UPF recommendations into provisioning arguments.", "checks": { "Observation state": "NE", "Feasibility": "Specialist LLMs reason over latency/capacity; the orchestrator checks compatibility of the combined recommendation before calling the provisioning tool.", "Candidate coverage": "NE" }, "note": "The orchestrator agent performs the consistency check.", "url": "https://arxiv.org/abs/2601.06640v1", "year": "2026" }, { "id": 71, "title": "Vision Language Models for Optimization-Driven Intent Processing in Autonomous Networks", "interfaces": [ "G", "C" ], "workflow": "Optimization-code generation and SDN demonstration", "path": "Optimization-code assistant; SDN controller. VLMs generate Gurobi programs; a scoped example solves flows, translates them into rules, and deploys/tests with Ryu/Mininet.", "checks": { "Observation state": "NE", "Feasibility": "Gurobi solves the constraints encoded in the generated program.", "Candidate coverage": "NE" }, "note": "The 85-instance program benchmark and a separate two-node executed demonstration are distinct evidence scopes.", "url": "https://arxiv.org/abs/2601.12744v1", "year": "2026" }, { "id": 72, "title": "IntAgent: NWDAF-Based Intent LLM Agent Towards Advanced Next Generation Networks", "interfaces": [ "G", "C" ], "workflow": "Planned, timed slice-rate changes via tools", "path": "NWDAF-based intent agent; core/session-management tools. An agent plans timed slice AMBR changes; tools apply and restore the settings.", "checks": { "Observation state": "NE", "Feasibility": "Dedicated feasibility and session-management tools evaluate action constraints and allowed QoS/state transitions before execution.", "Candidate coverage": "NE" }, "note": "NWDAF analytics, tool planning, timed actions, and observed throughput have different timing boundaries.", "url": "https://arxiv.org/abs/2601.13114v1", "year": "2026" }, { "id": 73, "title": "LEAD-Drift: Real-time and Explainable Intent Drift Detection by Learning a Data-Driven Risk Score", "interfaces": [ "S", "C" ], "workflow": "Risk-based drift alerts and attribution", "path": "Intent-assurance monitor. Learned risk and EMA thresholds raise drift alerts; SHAP and horizon rules supply attribution and time-to-failure ranges.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Risk scoring and time-series smoothing feed the alert decision.", "url": "https://arxiv.org/abs/2602.13672v1", "year": "2026" }, { "id": 74, "title": "MILD: Multi-Intent Learning and Disambiguation for Proactive Failure Prediction in Intent-based Networking", "interfaces": [ "S", "C" ], "workflow": "Multi-intent risk alerts and root-cause attribution", "path": "Multi-intent assurance monitor. Per-intent risk/gating selects alerts and root causes, with SHAP and horizon-based diagnostic outputs.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Primary evaluation uses labeled synthetic data, supplemented by a telemetry proof of concept.", "url": "https://arxiv.org/abs/2602.14283v1", "year": "2026" }, { "id": 75, "title": "Performance Comparison of IBN orchestration using LLM and SLMs", "interfaces": [ "G", "S" ], "workflow": "Agent plans/code to emulated-network deployment", "path": "Intent workflow orchestrator; emulated-network controller. Junior agents propose topology/service choices; a senior generates code; a policy agent recommends OSPF/DUAL before deployment.", "checks": { "Observation state": "NE", "Feasibility": "A senior model compares proposals under policy constraints; emulation provides an instantiation failure signal.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2603.06647v1", "year": "2026" }, { "id": 76, "title": "Contract-based Agentic Intent Framework for Network Slicing in O-RAN", "interfaces": [ "G", "C" ], "workflow": "Intent contract to rApp policy and xApp control", "path": "Intent contract handler; SLA rApp; near-RT xApps. Agent translation becomes a TMF921 contract; an rApp prepares A1 policy; KPM-driven xApps calculate PRB ratios and execute E2 control.", "checks": { "Observation state": "NE", "Feasibility": "An evaluator agent checks extracted intent; the SLA rApp separately checks requested throughput against PRB/CQI performance history before generating A1 policy.", "Candidate coverage": "NE" }, "note": "The SLA rApp estimates feasibility from performance history; numerical xApps act on current KPM data.", "url": "https://arxiv.org/abs/2603.01663v1", "year": "2026" }, { "id": 77, "title": "AI-driven Intent-Based Networking Approach for Self-configuration of Next Generation Networks", "interfaces": [ "G", "S", "U" ], "workflow": "Policy activation; proposed ranked assurance repair", "path": "Intent handler; SDN activation; assurance (mixed implemented/proposed). Schema-constrained policy realization and conflict-aware activation are combined with predictive multi-intent assurance; ranked repair remains conceptual.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The thesis implements selected components; ranked repair and general conflict governance remain proposals.", "url": "https://arxiv.org/abs/2603.23772v1", "year": "2026" }, { "id": 78, "title": "NetAgentBench: A State-Centric Benchmark for Evaluating Agentic Network Configuration", "interfaces": [ "G" ], "workflow": "Interactive routing commands to converged state", "path": "Network configuration/repair agent. Interactive FRR commands are evaluated against converged RIP/OSPF/BGP network state.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Evaluation scores converged network state after configuration actions.", "url": "https://arxiv.org/abs/2604.09678v1", "year": "2026" }, { "id": 79, "title": "A Reproducible Semantic Benchmark for Multivendor DSM-to-CLI Translation", "interfaces": [ "G", "S" ], "workflow": "Multivendor CLI generation and model judging", "path": "DSM-to-CLI translator; semantic evaluation layer. Models generate multivendor CLI; three model judges adjudicate semantic equivalence.", "checks": { "Observation state": "NE", "Feasibility": "A model-judge panel scores DSM/CLI semantic consistency during evaluation.", "Candidate coverage": "NE" }, "note": "Judge agreement and token cost concern the benchmark evaluation pipeline.", "url": "https://arxiv.org/abs/2606.20564v1", "year": "2026" }, { "id": 80, "title": "Agentic AI-Based Joint Computing and Networking via Mixture of Experts and Large Language Models", "interfaces": [ "S", "G" ], "workflow": "Expert/weight selection to resource allocation", "path": "Joint communications/computing optimizer. An LLM gate selects experts and constructs weights; learned expert outputs produce transmit/compute power and CPU allocations.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Continuous power and CPU allocations are constructed values.", "url": "https://arxiv.org/abs/2605.02911v1", "year": "2026" }, { "id": 81, "title": "Enhancing Secure Intent-Based Networking with an Agentic AI: The EU Project MARE Approach", "interfaces": [ "S", "G" ], "workflow": "Security intent completion to domain agents", "path": "Security intent handler; domain-agent orchestrator. The orchestrator checks intent support, fills missing parameters with the operator, plans action graphs, and invokes domain agents.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "Orchestrator classifies whether the system supports the intent and alerts the operator on unsupported requests." }, "note": "", "url": "https://arxiv.org/abs/2604.06856v1", "year": "2026" }, { "id": 82, "title": "Validated Intent Compilation for Constrained Routing in LEO Mega-Constellations", "interfaces": [ "G", "C", "S" ], "workflow": "Offline intent compilation to certified LEO routing", "path": "LEO intent compiler; routing certifier/controller. Offline language compilation produces ConstraintProgram JSON; deterministic validation precedes constrained GNN routing.", "checks": { "Observation state": "NE", "Feasibility": "Eight-pass validator grounds entities and detects conflicts; supported fragments obtain constructive routing witnesses for simultaneous hard constraints.", "Candidate coverage": "Certifier distinguishes ACCEPT, REJECT, and ABSTAIN; REJECT/acceptance guarantees are limited to supported fragments. Unsupported combinations use fallback without constructive certification." }, "note": "ABSTAIN marks an unresolved certification outcome; feasibility and fallback success remain undetermined.", "url": "https://arxiv.org/abs/2604.07264v1", "year": "2026" }, { "id": 83, "title": "Benchmarking LLM-Driven Network Configuration Repair", "interfaces": [ "S", "G" ], "workflow": "Configuration-context selection and repair", "path": "Configuration diagnostic/repair assistant. Models select relevant files, locate routers, explain faults, and produce search/replace repairs in Cornetto.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Ground-truth forwarding checks score benchmark outputs.", "url": "https://arxiv.org/abs/2604.22513v1", "year": "2026" }, { "id": 84, "title": "TIO-SHACL: Comprehensive SHACL validation for TMF Intent Ontologies", "interfaces": [ "C" ], "workflow": "RDF intent representation/ontology validation", "path": "Intent-ontology validator. TIO-SHACL returns conformance and violations for RDF intents and ontology functions.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Type, function-argument and hierarchy checks validate the intent representation.", "url": "https://arxiv.org/abs/2604.27359v1", "year": "2026" }, { "id": 85, "title": "Role-Based Agentic AI for Intent-Driven Network and Service Orchestration", "interfaces": [ "G", "C" ], "workflow": "Business intent to governed service orders", "path": "Business intent handler; supervisor; service orchestrator. Business RDF and governance-guided DAGs lead to OSL/OSM service orders; general resource agents remain architectural.", "checks": { "Observation state": "NE", "Feasibility": "MCP service tools are assigned feasibility/resource checks; supervisor governs admission, while the customer confirms semantic interpretation.", "Candidate coverage": "NE" }, "note": "Evaluation measures the three-agent service-order workflow.", "url": "https://arxiv.org/abs/2606.20580", "year": "2026" }, { "id": 86, "title": "$E^3$-Agent: An Executable and Evolving Agent for Resource Management of Edge Generative Inference", "interfaces": [ "G", "C" ], "workflow": "Event-driven parameter updates to task routing", "path": "Event-driven meta-controller; edge task router. An LLM adjusts risk TTL/calibration/router parameters; deterministic fast-path scoring dispatches each task.", "checks": { "Observation state": "Performance model consumes only completed feedback available at the current timestamp; delayed/missing feedback leaves estimates unchanged until a valid record arrives.", "Feasibility": "NE", "Candidate coverage": "Router restricts to available devices, avoids risky ones if alternatives exist, and routes anyway with penalties when all are risky; admission/drop decisions belong to an outer policy." }, "note": "", "url": "https://arxiv.org/abs/2605.27428v1", "year": "2026" }, { "id": 87, "title": "AgentxGCore: Agentic AI for Next-Generation Mobile Core Network", "interfaces": [ "G" ], "workflow": "Core-network plans to UPF/user reassignment", "path": "Core-network orchestration agent. Plans create/remove UPFs and reassign users through SMF/PCF execution tools.", "checks": { "Observation state": "NE", "Feasibility": "The executor agent assesses plan feasibility before selecting execution tools.", "Candidate coverage": "NE" }, "note": "The agent uses predictive telemetry and bounded context memory.", "url": "https://arxiv.org/abs/2606.00417v1", "year": "2026" }, { "id": 88, "title": "RAG-driven Multi-Agent LLM Framework with Task Decomposition for Beyond 5G Auto-Configuration", "interfaces": [ "G" ], "workflow": "gNB configuration/repair to approved startup", "path": "Configuration agent; gNB deployment orchestrator. Agents generate gNB configuration, return segmented review/repair reports, and prepare approved startup commands.", "checks": { "Observation state": "NE", "Feasibility": "A configuration-verification agent reviews generated segments and returns repair feedback; human approval gates startup commands.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2606.01222v1", "year": "2026" }, { "id": 89, "title": "Bridging High-Level Intent and Network Execution: Detecting Violations and Intent Drift Through Low-Level Traffic Analysis", "interfaces": [ "C" ], "workflow": "Offline port-policy compliance and drift monitoring", "path": "Data-plane intent-assurance monitor. Port-policy rules flag violating flows; a fixed empirical baseline flags drift in offline traffic records.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Offline compliance experiments use a four-field projection of a conceptual seven-tuple.", "url": "https://arxiv.org/abs/2606.05076v1", "year": "2026" }, { "id": 90, "title": "Privacy-Preserving Intent Fulfilment and Assurance for 6G RAN", "interfaces": [ "S", "C" ], "workflow": "Intent profiles to O1 assurance counters", "path": "RAN management/assurance over O1. Intent categories map to resource profiles; aggregate performance counters determine fulfillment/degradation.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The O1 assurance interface uses aggregate performance counters.", "url": "https://arxiv.org/abs/2607.08809v1", "year": "2026" }, { "id": 91, "title": "Distilling Knowledge from Large Language Models into Lightweight Reinforcement Learning Agents for Autonomous Cyber Operations", "interfaces": [ "S" ], "workflow": "Distilled policy for isolation/restoration actions", "path": "Network defense policy. A language-model teacher and distilled PPO student choose network isolation/restoration actions in CybORG.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Teacher-guided action masking is applied during distillation.", "url": "https://arxiv.org/abs/2607.28826v1", "year": "2026" }, { "id": 92, "title": "Learning Not to Optimize: Physics-Informed Action-Space Reshaping for Intent-Based Network Control", "interfaces": [ "C", "S" ], "workflow": "Candidate screening to network deployment", "path": "Candidate-set guard; network deployment selector. Residual screening, symmetry quotienting, and dominance filtering form a frontier before action selection.", "checks": { "Observation state": "NE", "Feasibility": "Executor recomputes exact network-law residuals; invalid selected actions are rejected, repaired, or mapped to fallback.", "Candidate coverage": "An empty executable set among generated candidates triggers fallback. The paper separates finite-generation coverage loss from screening correctness." }, "note": "Screening certifies members of the generated candidate set.", "url": "https://arxiv.org/abs/2608.00908v1", "year": "2026" }, { "id": 93, "title": "NetConfArena: An Executable Benchmark for LLM Agents in Closed-Loop Network Configuration", "interfaces": [ "G" ], "workflow": "Interactive CLI changes scored in emulation", "path": "Interactive network configuration agent. Agents produce and submit device CLI changes in an emulated network; task predicates score the resulting behavior.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Negative tests cover the benchmark's stated prohibitions.", "url": "https://arxiv.org/abs/2608.23179v1", "year": "2026" }, { "id": 94, "title": "Uncertainty Signals for Network Intent Translation: Risk Ranking and Ambiguity Localization", "interfaces": [ "G", "C" ], "workflow": "Command generation with uncertainty-based review", "path": "Intent translator; uncertainty-based review selector. Juniper command candidates are ranked by uncertainty for abstention/review; token entropy locates ambiguity.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Uncertainty scores determine candidate ranking and abstention.", "url": "https://arxiv.org/abs/2609.04486v1", "year": "2026" }, { "id": 95, "title": "On Identifying Adversarial Intent Injection in AI-Native 6G Networks", "interfaces": [ "S" ], "workflow": "Security classification of intent windows", "path": "Intent security filter. A classifier marks an intent sequence window malicious when attack content is detected.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Evaluation uses synthetic attack-intent windows.", "url": "https://arxiv.org/abs/2609.12144v1", "year": "2026" }, { "id": 96, "title": "NetInspector: Measuring and Improving LLM Capabilities for Reliable Intent-Based Networking Policy Generation", "interfaces": [ "G", "S", "C" ], "workflow": "Intent translation to tool-based policy inspection", "path": "Intent translator; policy-inspection agent. Entity prefiltering narrows policies; a tool-using agent queries network facts and returns Conflict or Safe.", "checks": { "Observation state": "A dedicated state manager versions and applies updates atomically so each query sees a consistent network snapshot.", "Feasibility": "Deterministic tools resolve groups and query reachability, middleboxes, rule priorities, bandwidth, and temporal overlap; the LLM combines the replies into a terminal verdict.", "Candidate coverage": "NE" }, "note": "Each query reads a consistent snapshot; the LLM assembles tool replies into the final verdict.", "url": "https://arxiv.org/abs/2609.21103v1", "year": "2026" }, { "id": 97, "title": "TelecomGPT-R1: Unified Post-Training for Reasoning Across Heterogeneous Telecom Tasks", "interfaces": [ "S", "G" ], "workflow": "Telecom fault classification with reasoning", "path": "Telecom diagnostic assistant. TeleLogs inputs produce a fault-class choice with supporting reasoning.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Rule/answer checks are used in training-data construction.", "url": "https://arxiv.org/abs/2609.25356v1", "year": "2026" }, { "id": 98, "title": "From Intents to Algorithms: Verified Algorithm Discovery for Transport Networks", "interfaces": [ "G", "C" ], "workflow": "Numerical algorithm search to path allocation", "path": "Intent compiler; transport algorithm search and allocator. Conceptual LLM/AST discovery is separated from measured numerical ranking-function evolution and deterministic path allocation.", "checks": { "Observation state": "NE", "Feasibility": "Trusted network functions retain capacity, latency, path-validity, and single-path feasibility; discovered algorithms rank alternatives without redefining those constraints.", "Candidate coverage": "NE" }, "note": "Evaluation covers the ten-parameter numerical search; free algorithm generation remains proposed.", "url": "https://arxiv.org/abs/2609.27386v1", "year": "2026" }, { "id": 99, "title": "Intent2Tc: Automated Intent-to-Traffic Control Translation with Language Models", "interfaces": [ "G", "C" ], "workflow": "Intent compilation to corrected traffic-shaping rules", "path": "Traffic-shaping intent compiler. Language models produce metadata, declarative sub-intents, and Linux tc rules; deterministic correction intervenes between stages.", "checks": { "Observation state": "NE", "Feasibility": "A template-based correction module enforces scoped policy consistency and tc hierarchy, priority, filtering, and syntax constraints.", "Candidate coverage": "NE" }, "note": "Template correction applies a fixed traffic semantic model.", "url": "https://arxiv.org/abs/2609.31397v1", "year": "2026" }, { "id": 100, "title": "Automatic Intent-Based Secure Service Creation Through a Multilayer SDN Network Orchestration", "interfaces": [ "C" ], "workflow": "Multilayer SDN service setup and teardown", "path": "Multilayer SDN service orchestrator. Encryption/domain constraints select a layer and installation plan, followed by connection setup or teardown.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Encryption, path provisioning, installation acknowledgements, and teardown have separate responsibilities and timing endpoints.", "url": "https://arxiv.org/abs/1803.03106v1", "year": "2018" }, { "id": 101, "title": "Cross-stakeholder service orchestration for B5G through capability provisioning", "interfaces": [ "U" ], "workflow": "Conceptual domain/capability/service orchestration", "path": "Domain/capability/cross-domain service orchestrators (conceptual). Business goals map to domain configuration, capability/SLA instances, and end-to-end provider combinations.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The architecture specifies three functional transformations; decision-interface implementations are unspecified.", "url": "https://arxiv.org/abs/2008.07162v1", "year": "2020" }, { "id": 102, "title": "NetOrchLLM: Mastering Wireless Network Orchestration with Large Language Models", "interfaces": [ "S", "G" ], "workflow": "Wireless tool selection and allocation generation", "path": "Wireless model/tool orchestrator. Task/model selection and function calls support constructed bandwidth and multiuser power-allocation vectors.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Selecting a model/tool and constructing its numerical allocation output are distinct interfaces.", "url": "https://arxiv.org/abs/2412.10107v1", "year": "2024" }, { "id": 103, "title": "An Autonomous Network Orchestration Framework Integrating Large Language Models with Continual Reinforcement Learning", "interfaces": [ "S", "G" ], "workflow": "Semantic objectives to sequential resource actions", "path": "Network orchestration controller. Semantic objective/QoE decisions and user sequencing feed RL placement, capacity, and path allocation.", "checks": { "Observation state": "State history is updated after each user allocation before subsequent users are processed.", "Feasibility": "Per-user/service state masking restricts the action executor to the represented feasible space.", "Candidate coverage": "NE" }, "note": "The model constructs a user ordering and selects resource actions. Feasibility depends on the represented infrastructure and mask.", "url": "https://arxiv.org/abs/2502.16198v1", "year": "2025" }, { "id": 104, "title": "KP-A: A Unified Network Knowledge Plane for Catalyzing Agentic Network Intelligence", "interfaces": [ "G", "S" ], "workflow": "Network queries and service subscriptions", "path": "Network knowledge/query plane; edge-service agent. Live/static knowledge endpoints support state answers, service recommendations, and subscription creation.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The prototype demonstrates live queries; freshness and versioning are design requirements.", "url": "https://arxiv.org/abs/2507.08164v1", "year": "2025" }, { "id": 105, "title": "Agentic AI for SAGIN Resource Management_Semantic Awareness, Orchestration, and Optimization", "interfaces": [ "G", "S", "C" ], "workflow": "Semantic rewards to RL placement/resource solving", "path": "SAGIN semantic manager; RL controller and resource solver. Semantic analysis shapes RL rewards; task placement/routing feeds continuous resource solving; command generation has a separate architectural scope.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Performance results cover the resource case study; command generation remains architectural.", "url": "https://arxiv.org/abs/2603.16458v1", "year": "2026" }, { "id": 106, "title": "WirelessBench: A Tolerance-Aware LLM Agent Benchmark for Wireless Network Intelligence", "interfaces": [ "G", "S" ], "workflow": "Tool-assisted wireless allocation and QoS decisions", "path": "Wireless allocation/feasibility assistant. Tool-augmented agents produce slice, CQI, bandwidth, throughput, mobility-conditioned allocation, and QoS decisions.", "checks": { "Observation state": "NE", "Feasibility": "The agent compares predicted throughput with a service minimum to produce the scoped QoS verdict; this remains part of the task being evaluated.", "Candidate coverage": "NE" }, "note": "Benchmark scoring uses a tolerance criterion. Mobility/CQI dependency errors can propagate into the final QoS verdict.", "url": "https://arxiv.org/abs/2603.21251v1", "year": "2026" }, { "id": 107, "title": "Intent-Driven 6G Service Orchestration: Grounded Translation, Validation, and Decomposition", "interfaces": [ "G", "S", "C" ], "workflow": "Catalog discovery to service-profile decomposition", "path": "TMF-style intent/service-management boundary. Catalog discovery and RDF construction feed requirement extraction, minimum-cost CFSS choice, and RFSS coverage/decomposition.", "checks": { "Observation state": "NE", "Feasibility": "RequirementCapability predicates test all extracted requirements against profile capabilities; SHACL separately validates RDF structure.", "Candidate coverage": "The capability layer determines whether any current catalog profile satisfies requirements; RFSS decomposition covers requirements using a greedy heuristic compared with CP-SAT." }, "note": "Coverage is evaluated against supplied catalog profiles and requirement predicates.", "url": "https://arxiv.org/abs/2606.28348", "year": "2026" }, { "id": 108, "title": "MOSS: End-to-End Dialog System Framework with Modular Supervision", "interfaces": [ "G", "S" ], "workflow": "Troubleshooting dialogue acts and responses", "path": "Network troubleshooting dialogue manager. Utterances update dialogue state, choose clarification/repair acts, and generate responses.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The network subset evaluates dialogue success in laptop-network troubleshooting.", "url": "https://arxiv.org/abs/1909.05528v1", "year": "2019" }, { "id": 109, "title": "MSADM: Large Language Model (LLM) Assisted End-to-End Network Health Management Based on Multi-Scale Semanticization", "interfaces": [ "S", "G", "C" ], "workflow": "Network-health diagnosis and proposed mitigation", "path": "Network health/diagnosis assistant. Anomaly/fault classification and semantic state summaries support root-cause reports and tc mitigation suggestions.", "checks": { "Observation state": "NE", "Feasibility": "LLM prompts request evidence checks and constrained mitigation choices.", "Candidate coverage": "NE" }, "note": "Mitigation outputs are suggested scripts.", "url": "https://arxiv.org/abs/2406.08305v4", "year": "2024" }, { "id": 110, "title": "LLMcap: Large Language Model for Unsupervised PCAP Failure Detection", "interfaces": [ "S", "C" ], "workflow": "Packet reconstruction to failure/localization labels", "path": "Packet-trace diagnostic monitor. Masked-language-model reconstruction errors drive call-failure and frame/chunk localization decisions.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Generative reconstruction is an internal signal; the retained diagnostic outputs are detection/localization labels.", "url": "https://arxiv.org/abs/2407.06085v1", "year": "2024" }, { "id": 111, "title": "Adapting Network Information into Semantics for Generalizable and Plug-and-Play Multi-Scenario Network Diagnosis", "interfaces": [ "G", "C", "S" ], "workflow": "Network knowledge to diagnostic plans/reports", "path": "Network knowledge and diagnostic assistant. Semantic/symbolic descriptions and KG updates support diagnosis plans, anomaly decisions, and explanatory reports.", "checks": { "Observation state": "NE", "Feasibility": "A Z3-assisted symbolic component reasons over represented network rules; the final diagnostic report remains a model output.", "Candidate coverage": "NE" }, "note": "", "url": "https://arxiv.org/abs/2501.16842v2", "year": "2025" }, { "id": 112, "title": "Leveraging Multi-Agent System (MAS) and Fine-Tuned Small Language Models (SLMs) for Automated Telecom Network Troubleshooting", "interfaces": [ "S", "G" ], "workflow": "Delegated telecom diagnosis and action planning", "path": "Telecom troubleshooting orchestrator. Coordinator and planner delegate diagnostic work, parameterize actions, and assemble root-cause reports.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Human approval gates executable plans; grounding and format checks also appear in training.", "url": "https://arxiv.org/abs/2511.00651v2", "year": "2025" }, { "id": 113, "title": "HYVE: Hybrid Views for LLM Context Engineering over Machine Data", "interfaces": [ "G", "S" ], "workflow": "Runbook, API and diagnostic-branch assistance", "path": "Network troubleshooting assistant. Runbooks, anomaly identification, reports, API arguments, summaries, branch decisions, and test IDs use different output interfaces.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "ValidateSourceCoverage checks preservation of serialized machine-data content.", "url": "https://arxiv.org/abs/2604.05400v1", "year": "2026" }, { "id": 114, "title": "Cross-Domain Query Translation for Network Troubleshooting: A Multi-Agent LLM Framework with Privacy Preservation and Self-Reflection", "interfaces": [ "S", "G" ], "workflow": "Clarification to cross-domain diagnostic queries", "path": "Cross-domain troubleshooting interface. Domain/intent classification and clarification precede anonymization, technical query translation, diagnosis, and user explanation.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Confidence scores trigger clarification; anonymization precedes diagnostic translation.", "url": "https://arxiv.org/abs/2604.13353v2", "year": "2026" }, { "id": 115, "title": "SADE: Symptom-Aware Diagnostic Escalation for LLM-Based Network Troubleshooting", "interfaces": [ "S", "C" ], "workflow": "Symptom probes to skill-based fault localization", "path": "Symptom-driven network diagnostic agent. Initial probes and deeper scans establish symptoms; a fault index routes to a skill that localizes a canonical fault/device.", "checks": { "Observation state": "The diagnostic workflow checks ambiguous reachability with direct-IP probes and inspects lower-layer evidence before accepting apparently healthy higher-layer observations.", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The agent performs skill-based interpretation at runtime. Verified fault injection prepares the benchmark.", "url": "https://arxiv.org/abs/2605.04530v1", "year": "2026" }, { "id": 116, "title": "PropLLM: Propagation-Aware Scene Reconstruction for Network Fault Diagnosis", "interfaces": [ "S", "G" ], "workflow": "Multimodal faults to root-cause/propagation chains", "path": "Propagation-aware network diagnostic model. Multimodal evidence produces a fault class and a generated root-cause/propagation chain.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Causal direction masks and timestamp biases constrain model reasoning.", "url": "https://arxiv.org/abs/2606.00582v1", "year": "2026" }, { "id": 117, "title": "Intent-Based Management of Next-Generation Networks: an LLM-Centric Approach", "interfaces": [ "G" ], "workflow": "Multidomain intent translation to service descriptors", "path": "Multidomain/domain intent handlers; NFV/RAN orchestrators. Language decomposition and translation form NSD/RAND descriptors for domain activation.", "checks": { "Observation state": "NE", "Feasibility": "The architecture assigns structure, conflict, and predicted-resource feasibility checks to intent handlers before deployment.", "Candidate coverage": "NE" }, "note": "Measured timing ends at descriptor translation. Validation mechanisms are proposed, and activation lies downstream.", "url": "https://doi.org/10.1109/mnet.2024.3420120", "year": "2024" }, { "id": 118, "title": "Configtrans: Network Configuration Translation Based on Large Language Models and Constraint Solving", "interfaces": [ "S", "C" ], "workflow": "Parameter/command mapping to target CLI trees", "path": "Cross-vendor configuration translator. Constraint solving maps parameters; reranking/LLM choice selects commands and keywords; view ordering constructs the target CLI tree.", "checks": { "Observation state": "NE", "Feasibility": "CSP enforces command/parameter correspondence and consistent view structure within the translation problem.", "Candidate coverage": "NE" }, "note": "", "url": "https://doi.org/10.1109/ICNP61940.2024.10858572", "year": "2024" }, { "id": 119, "title": "Network CoPilot: Intent-Driven Network Configuration Updating for Service Guarantee", "interfaces": [ "G" ], "workflow": "Graph-grounded routing-parameter updates", "path": "Network configuration-update assistant. A graph-enriched language model constructs OSPF/BGP parameter-update descriptions from intent, configuration, topology, and link state.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "", "url": "https://doi.org/10.1109/infocom55648.2025.11044495", "year": "2025" }, { "id": 120, "title": "Intent-Based Management for Open RAN: Intelligent Network Configuration Automation via Chatbot", "interfaces": [ "G", "S", "C" ], "workflow": "Dialogue-based intent to ONOS installation", "path": "Chatbot intent handler; ONOS controller. Entity/intent extraction and dialogue acts produce NILE that is installed through REST or CLI.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "ConflictResolver is planned for future implementation; the demonstrated path covers intent installation.", "url": "https://doi.org/10.1109/cloudnet62863.2024.10815823", "year": "2024" }, { "id": 121, "title": "NetLLMBench: A Benchmark Framework for Large Language Models in Network Configuration Tasks", "interfaces": [ "G" ], "workflow": "Command generation and probes in emulation", "path": "Network configuration benchmark agent. Models generate device-command JSON, request ping probes, and repair malformed responses.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Syntax verification and network-emulator feedback evaluate generated commands.", "url": "https://doi.org/10.1109/NFV-SDN61811.2024.10807499", "year": "2024" }, { "id": 122, "title": "Towards LLM-Based Failure Localization in Production-Scale Networks", "interfaces": [ "G", "S" ], "workflow": "Production alert diagnosis with operator assistance", "path": "Production-network diagnostic copilot. Alert summaries and device anomalies/scores feed ranked root-cause explanations and operator-assisted incident localization.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Component timing, 24-hour API scopes, and operator-assisted resolution time are distinct evidence boundaries.", "url": "https://doi.org/10.1145/3718958.3750505", "year": "2025" }, { "id": 123, "title": "NetAssistant: Dialogue Based Network Diagnosis in Data Center Networks", "interfaces": [ "G", "S", "C" ], "workflow": "Dialogue-selected workflows to health reports", "path": "Dialogue-based network diagnosis. Entity normalization and intent matching select workflows that produce network-health/root-cause reports for on-call users.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Workflow diagnosis and on-call resolution have different endpoints. Control/data-plane verification modules remain proposed.", "url": "https://www.usenix.org/conference/nsdi24/presentation/wang-haopei", "year": "2024" }, { "id": 124, "title": "Intent-Driven Network Management with Multi-Agent LLMs: The Confucius Framework", "interfaces": [ "G", "S" ], "workflow": "Workflow/DSL generation with diagnostic tools", "path": "Network-management workflow/DSL assistant. Confucius plans DAG/MOP workflows, selects blocks, generates topology/query DSL, and coordinates diagnostic tools.", "checks": { "Observation state": "NE", "Feasibility": "Custom parsers, API/ORM dry runs, and graph-invariant validators provide task-specific feedback before use.", "Candidate coverage": "NE" }, "note": "Retrieval filters stale documents; validators check specific DSL and network invariants.", "url": "https://doi.org/10.1145/3718958.3750537", "year": "2025" }, { "id": 125, "title": "Graph Structure-Enhanced Large Language Model for Optical Network Fault Diagnosis: An Explainable Alarm Root Cause Localization Approach", "interfaces": [ "G", "C", "S" ], "workflow": "Alarm propagation graphs to root-cause diagnosis", "path": "Optical-network alarm diagnostic assistant. Models construct event topology/propagation matrices; adjacency construction and root-cause localization use the resulting graph.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The model constructs a structured matrix that supports graph-derived root-cause explanations.", "url": "https://doi.org/10.1109/jiot.2025.3573056", "year": "2025" }, { "id": 126, "title": "LLM-Augmented Deep Reinforcement Learning for Dynamic O-RAN Network Slicing", "interfaces": [ "S" ], "workflow": "LLM-assisted learning for O-RAN slice allocation", "path": "O-RAN slicing policy (LLM-augmented MARL). Critic-only and fuller LLM-assisted learning variants produce joint slice/UE resource-block allocation.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Training/critic assistance and per-action inference have separate evaluation scopes.", "url": "https://doi.org/10.1109/ICC52391.2025.11161572", "year": "2025" }, { "id": 127, "title": "R-IBN: A reinforcement learning-based intent-driven framework for end-to-end service orchestration and optimization", "interfaces": [ "G", "S", "C" ], "workflow": "Intent translation to learned service-chain placement", "path": "Intent handler; service-chain orchestrator. Language translation and contradiction classification feed GNN-assisted DRL SFC placement/routing.", "checks": { "Observation state": "NE", "Feasibility": "Embedding-similarity thresholds with KNN refinement flag intent-policy contradictions through a learned semantic check.", "Candidate coverage": "NE" }, "note": "GNN state forecasts supply the input to placement and routing decisions.", "url": "https://doi.org/10.1016/j.comnet.2025.111564", "year": "2025" }, { "id": 128, "title": "CEGS: Configuration Example Generalizing Synthesizer", "interfaces": [ "G", "S", "C" ], "workflow": "Example/template selection to formal synthesis", "path": "Configuration example retriever; formal synthesizer. Intent normalization/roles select examples and device mappings; templates are verified and symbolically completed.", "checks": { "Observation state": "NE", "Feasibility": "CEGS combines a Batfish-derived syntax parser, Python local attribute validation, NetComplete-based global verification, and formal template filling.", "Candidate coverage": "NE" }, "note": "Formal correctness is relative to encoded intent, topology, and templates. Timing covers the complete synthesis.", "url": "https://www.usenix.org/conference/nsdi25/presentation/liu-jianmin", "year": "2025" }, { "id": 129, "title": "Nissist: An Incident Mitigation Copilot based on Troubleshooting Guides", "interfaces": [ "G", "S" ], "workflow": "Clarified intent to guide-based mitigation actions", "path": "Incident mitigation copilot. Clarified intent selects troubleshooting-guide nodes and produces corrected diagnostic/mitigation actions.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The reviewed task is network connectivity; reported mitigation times aggregate mixed-cloud incidents.", "url": "https://arxiv.org/abs/2402.17531v2", "year": "2024" }, { "id": 130, "title": "A Holistic View of AI-driven Network Incident Management", "interfaces": [ "G", "S", "U" ], "workflow": "Conceptual incident diagnosis and mitigation", "path": "Network incident-management assistant (conceptual). A proposed pipeline ranks hypotheses, plans tests and interprets monitors, then recommends mitigation.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The paper proposes module roles and future verification research.", "url": "https://doi.org/10.1145/3626111.3628176", "year": "2023" }, { "id": 131, "title": "Towards Explainable Network Intrusion Detection Using Large Language Models", "interfaces": [ "S", "G" ], "workflow": "Intrusion classification with explanatory alerts", "path": "Network intrusion detector and explanation layer. Llama/GPT variants classify flows and produce explanatory alerts in separate scopes.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "The paper evaluates intrusion classification and explanations, and discusses inference limits for real-time detection.", "url": "https://arxiv.org/abs/2408.04342v1", "year": "2024" }, { "id": 132, "title": "FedLLMGuard: A Federated Large Language Model for Anomaly Detection in 5G Networks", "interfaces": [ "S" ], "workflow": "Federated traffic/log anomaly classification", "path": "5G network anomaly detector. A federated, adapted language model classifies traffic/log embeddings as normal or anomalous.", "checks": { "Observation state": "NE", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Detection metrics have defined endpoints. The reported attack-mitigation time lacks an operational definition and leaves the mitigation path unspecified.", "url": "https://doi.org/10.1016/j.comnet.2025.111473", "year": "2025" }, { "id": 133, "title": "Can Jev Make SRE Agents More Reliable?", "interfaces": [ "S", "G" ], "workflow": "Agent diagnosis with ranked tests and evidence gates", "path": "SRE diagnosis/mitigation agent. The agent proposes hypotheses and read-only tests; Jev ranks the tests and reviews diagnosis and mitigation evidence; the agent runs tests and repairs.", "checks": { "Observation state": "jev_plan adds a fresh, bounded namespace snapshot before ranking; the diagnosis review requires evidence of a current application failure.", "Feasibility": "Each mitigation-review question must reach probability 0.70, covering whether the repair addresses the cause, restored functionality, and appears durable. Rejection returns the agent to planning.", "Candidate coverage": "NE" }, "note": "Jev ranks only agent-proposed tests; a missing hypothesis cannot be recovered by ranking. Pass rate is measured; time-to-diagnosis is not.", "url": "https://sregym.com/blog/jev-sregym-lite", "year": "2026" }, { "id": 134, "title": "OrchestRAN: Network Automation through Orchestrated Intelligence in the Open RAN", "interfaces": [ "C", "S" ], "workflow": "Request-driven model placement and xApp dispatch", "path": "Non-RT RIC orchestrator (rApp); near-RT RIC/DU control apps. A BILP selects ML/AI Catalog models and nodes; containers are dispatched as xApps/dApps, and DRL agents select scheduling/slicing policies.", "checks": { "Observation state": "NE", "Feasibility": "Constraints enforce model performance, control-loop timescales, node resources, and connectivity to required input data; pruning removes infeasible model-node variables.", "Candidate coverage": "A request is admitted only when deployed models provide all requested functionalities; the branching variant can accept requests on a subset of clusters." }, "note": "Solver time comes from MATLAB/CPLEX simulation; the Colosseum prototype reports placement and network effects without orchestration timing.", "url": "https://arxiv.org/abs/2201.05632", "year": "2022" }, { "id": 135, "title": "Benchmarking Various ML Solutions in Complex Intent-Based Network Management Systems", "interfaces": [ "C", "S", "U" ], "workflow": "Intent policies; ML-hardware performance estimation", "path": "Intent-based orchestrator for an ICT supply chain. An NLP intent manager emits JSON intents, a policy configurator matches stored policies, and an SVD/SGD recommender estimates ML-technique performance on devices.", "checks": { "Observation state": "NE", "Feasibility": "After configuring policies, the policy configurator checks conflicts before triggering the network controller; no implementation or evaluation is reported.", "Candidate coverage": "The policy configurator checks whether the requested ML technique exists in the store and alerts the user when it does not; this check is described but not evaluated." }, "note": "Evaluation covers recommender accuracy (normalized RMSE on AI-benchmark data); intent translation and policy configuration are architectural.", "url": "https://arxiv.org/abs/2111.07724v1", "year": "2021" }, { "id": 136, "title": "Bridging Language Models and Formal Methods for Intent-Driven Optical Network Design", "interfaces": [ "G", "C" ], "workflow": "Offline intent to planned optical network design", "path": "Optical network design assistant (pre-deployment). LLM grammar translation and optical RAG feed LLM-generated PDDL problems over a manual domain; an LLM agent translates the plan into a design.", "checks": { "Observation state": "NE", "Feasibility": "A Lark grammar parser validates structure and routes errors to automatic correction or user clarification; the PDDL solver plans under budget, latency, and protection constraints.", "Candidate coverage": "When no plan satisfies the constraints, the solver reports infeasibility and the user receives feedback; the pipeline can fall back to disclosed LLM-only topology generation." }, "note": "Case studies report 96.7% grammar-error detection, with 2 of 15 vague-value cases undetected. No deployed network state is measured.", "url": "https://arxiv.org/abs/2509.22834v1", "year": "2025" }, { "id": 137, "title": "BLINC: Context-Specific Causal Learning for Automated RAN Configuration", "interfaces": [ "G", "C" ], "workflow": "Offline causal priors to RAN power-control configuration", "path": "Non-RT RIC rApps with CI/CD deployment. LLM-derived mandatory/prohibited edges constrain Bayesian-network structure learning; inference scores P0/target-SNR configurations applied over O1/O2.", "checks": { "Observation state": "Recommendation scores each candidate configuration given current measurements; incremental CPD updates use new data, and a past recommendation degrades for cell-edge users.", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "LLM constraints keep edges voted by more than half of independent runs and enter the structure score. Recommendation searches a predefined feasible configuration space.", "url": "https://arxiv.org/abs/2604.27084v1", "year": "2026" }, { "id": 138, "title": "Proving the Utility of Large Language Models in Cybersecurity Simulations: A Comprehensive Examination", "interfaces": [ "G" ], "workflow": "Offline emulator-scenario and attack-agent generation", "path": "Cyber-simulation scenario generator. ChatGPT-4 generates YAML network configurations from templates or golden examples; validated scenarios feed LLM-synthesized Python attack agents.", "checks": { "Observation state": "NE", "Feasibility": "Topological checks cover adjacency, bidirectional firewall consistency, vulnerabilities, and process mappings; configurations that fail or do not run in NASim are discarded.", "Candidate coverage": "NE" }, "note": "Generated attack agents print their own compromise verdict without an independent check. Validity rates combine schema, topology, and NASim outcomes.", "url": "https://arxiv.org/abs/2608.16422v1", "year": "2026" }, { "id": 139, "title": "Performance Evaluation of Intent-Based Networking Scenarios: A GitOps and Nephio Approach", "interfaces": [ "C" ], "workflow": "Declarative intents to GitOps/Nephio deployment", "path": "GitOps reconcilers (Argo CD, Flux CD, ConfigSync) and the Nephio/Porch package orchestrator. Desired-state manifests pushed to Git are synchronized and deployed on Kubernetes; Nephio first hydrates intent packages.", "checks": { "Observation state": "Operators compare tracked desired states with runtime states and reconcile drift after webhook-detected pushes; ConfigSync synchronization reflects its polling period.", "Feasibility": "NE", "Candidate coverage": "NE" }, "note": "Intents are declarative manifests without language interpretation. The benchmark measures sync, deploy, and healthy times and resource use.", "url": "https://arxiv.org/abs/2509.13901v1", "year": "2025" } ], "verification": { "transport": { "sparse": { "0": { "verified": 300, "n": 300, "ack": [ 0.044, 0.051 ], "outcome": [ 0.561, 0.586 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.0, "attainment": 100.0, "attainment_ci": [ 100.0, 100.0 ], "rho": null }, "0.1": { "verified": 300, "n": 300, "ack": [ 0.144, 0.151 ], "outcome": [ 0.659, 0.687 ], "decision_share_ack": 69.3, "decision_share_outcome": 15.2, "attainment": 100.0, "attainment_ci": [ 100.0, 100.0 ], "rho": null }, "0.5": { "verified": 300, "n": 300, "ack": [ 0.544, 0.551 ], "outcome": [ 1.06, 1.086 ], "decision_share_ack": 91.9, "decision_share_outcome": 47.2, "attainment": 100.0, "attainment_ci": [ 100.0, 100.0 ], "rho": null }, "2": { "verified": 300, "n": 300, "ack": [ 2.045, 2.051 ], "outcome": [ 2.561, 2.587 ], "decision_share_ack": 97.8, "decision_share_outcome": 78.1, "attainment": 100.0, "attainment_ci": [ 100.0, 100.0 ], "rho": null } }, "load": { "4.992": { "0": { "verified": 1016, "n": 1016, "ack": [ 0.044, 0.051 ], "outcome": [ 0.561, 0.588 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.0, "attainment": 100.0, "attainment_ci": [ 100.0, 100.0 ], "rho": 0.0 }, "0.1": { "verified": 1016, "n": 1016, "ack": [ 0.154, 0.348 ], "outcome": [ 0.683, 0.869 ], "decision_share_ack": 65.3, "decision_share_outcome": 14.7, "attainment": 100.0, "attainment_ci": [ 100.0, 100.0 ], "rho": 0.502 }, "0.5": { "verified": 1016, "n": 1016, "ack": [ 233.246, 372.255 ], "outcome": [ 233.749, 372.767 ], "decision_share_ack": 0.2, "decision_share_outcome": 0.2, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 2.5 }, "2": { "verified": 1016, "n": 1016, "ack": [ 1405.524, 2229.573 ], "outcome": [ 1406.028, 2230.085 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.1, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 9.987 } }, "7.987": { "0": { "verified": 1576, "n": 1576, "ack": [ 0.044, 0.051 ], "outcome": [ 0.561, 0.586 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.0, "attainment": 100.0, "attainment_ci": [ 100.0, 100.0 ], "rho": 0.0 }, "0.1": { "verified": 1576, "n": 1576, "ack": [ 0.287, 0.838 ], "outcome": [ 0.802, 1.355 ], "decision_share_ack": 35.0, "decision_share_outcome": 12.6, "attainment": 100.0, "attainment_ci": [ 100.0, 100.0 ], "rho": 0.804 }, "0.5": { "verified": 1576, "n": 1576, "ack": [ 508.04, 773.329 ], "outcome": [ 508.566, 773.853 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.1, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 3.999 }, "2": { "verified": 1576, "n": 1576, "ack": [ 2509.895, 3838.564 ], "outcome": [ 2510.421, 3839.087 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.1, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 15.981 } }, "9.484": { "0": { "verified": 1929, "n": 1929, "ack": [ 0.045, 0.051 ], "outcome": [ 0.561, 0.587 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.0, "attainment": 100.0, "attainment_ci": [ 100.0, 100.0 ], "rho": 0.0 }, "0.1": { "verified": 1929, "n": 1929, "ack": [ 1.48, 2.711 ], "outcome": [ 2.002, 3.231 ], "decision_share_ack": 6.8, "decision_share_outcome": 5.0, "attainment": 100.0, "attainment_ci": [ 100.0, 100.0 ], "rho": 0.955 }, "0.5": { "verified": 1929, "n": 1929, "ack": [ 627.838, 970.613 ], "outcome": [ 628.367, 971.128 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.1, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 4.749 }, "2": { "verified": 1929, "n": 1929, "ack": [ 2981.335, 4625.823 ], "outcome": [ 2981.864, 4626.337 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.1, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 18.977 } } } }, "edge": { "sparse": { "0": { "verified": 300, "n": 300, "ack": [ 0.005, 0.006 ], "outcome": [ 1.379, 2.777 ], "decision_share_ack": 1.2, "decision_share_outcome": 0.0, "attainment": 85.0, "attainment_ci": [ 78.3, 90.0 ], "rho": null }, "0.1": { "verified": 300, "n": 300, "ack": [ 0.105, 0.106 ], "outcome": [ 1.602, 2.739 ], "decision_share_ack": 94.9, "decision_share_outcome": 6.2, "attainment": 76.0, "attainment_ci": [ 66.3, 84.3 ], "rho": null }, "0.5": { "verified": 300, "n": 300, "ack": [ 0.506, 0.506 ], "outcome": [ 2.189, 3.001 ], "decision_share_ack": 98.9, "decision_share_outcome": 22.8, "attainment": 29.7, "attainment_ci": [ 17.3, 43.0 ], "rho": null }, "2": { "verified": 300, "n": 300, "ack": [ 2.006, 2.007 ], "outcome": [ 3.394, 4.758 ], "decision_share_ack": 99.7, "decision_share_outcome": 58.9, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": null } }, "load": { "4.992": { "0": { "verified": 1016, "n": 1016, "ack": [ 0.005, 0.006 ], "outcome": [ 1.382, 2.734 ], "decision_share_ack": 0.6, "decision_share_outcome": 0.0, "attainment": 85.2, "attainment_ci": [ 83.6, 87.0 ], "rho": 0.0 }, "0.1": { "verified": 1016, "n": 1016, "ack": [ 0.115, 0.307 ], "outcome": [ 1.555, 2.867 ], "decision_share_ack": 87.9, "decision_share_outcome": 6.5, "attainment": 71.9, "attainment_ci": [ 69.0, 75.0 ], "rho": 0.502 }, "0.5": { "verified": 1016, "n": 1016, "ack": [ 233.21, 372.216 ], "outcome": [ 235.246, 374.051 ], "decision_share_ack": 0.2, "decision_share_outcome": 0.2, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 2.5 }, "2": { "verified": 1016, "n": 1016, "ack": [ 1405.488, 2229.533 ], "outcome": [ 1407.557, 2230.965 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.1, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 9.987 } }, "7.987": { "0": { "verified": 1576, "n": 1576, "ack": [ 0.005, 0.006 ], "outcome": [ 1.382, 2.798 ], "decision_share_ack": 0.6, "decision_share_outcome": 0.0, "attainment": 84.6, "attainment_ci": [ 82.6, 86.6 ], "rho": 0.0 }, "0.1": { "verified": 1576, "n": 1576, "ack": [ 0.248, 0.795 ], "outcome": [ 1.774, 3.063 ], "decision_share_ack": 40.6, "decision_share_outcome": 5.7, "attainment": 61.2, "attainment_ci": [ 56.8, 65.6 ], "rho": 0.804 }, "0.5": { "verified": 1576, "n": 1576, "ack": [ 507.999, 773.286 ], "outcome": [ 509.794, 774.647 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.1, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 3.999 }, "2": { "verified": 1576, "n": 1576, "ack": [ 2509.854, 3838.521 ], "outcome": [ 2512.032, 3839.745 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.1, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 15.981 } }, "9.484": { "0": { "verified": 1929, "n": 1929, "ack": [ 0.005, 0.006 ], "outcome": [ 1.382, 2.812 ], "decision_share_ack": 0.5, "decision_share_outcome": 0.0, "attainment": 84.3, "attainment_ci": [ 82.8, 85.9 ], "rho": 0.0 }, "0.1": { "verified": 1929, "n": 1929, "ack": [ 1.441, 2.674 ], "outcome": [ 2.954, 4.655 ], "decision_share_ack": 7.0, "decision_share_outcome": 3.4, "attainment": 18.6, "attainment_ci": [ 9.3, 29.0 ], "rho": 0.955 }, "0.5": { "verified": 1929, "n": 1929, "ack": [ 627.802, 970.569 ], "outcome": [ 628.943, 972.304 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.1, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 4.749 }, "2": { "verified": 1929, "n": 1929, "ack": [ 2981.299, 4625.779 ], "outcome": [ 2982.439, 4627.48 ], "decision_share_ack": 0.1, "decision_share_outcome": 0.1, "attainment": 0.0, "attainment_ci": [ 0.0, 0.0 ], "rho": 18.977 } } } } }, "checks": { "contracts": [ 45, 48 ], "routes": [ 4, 120 ], "gates": [ { "condition": "Independent quotas", "self": [ 29, 120 ], "gate": [ 84, 120 ], "gain": [ 45.8, 37.5, 54.2 ] }, { "condition": "Shared quotas", "self": [ 14, 120 ], "gate": [ 42, 120 ], "gain": [ 23.3, 15.8, 30.8 ] }, { "condition": "Shared quotas + fault-domain separation", "self": [ 6, 120 ], "gate": [ 45, 120 ], "gain": [ 32.5, 25.0, 40.0 ] } ] }, "lower_agreement": [ "Input load", "Queue / waiting", "Stability evidence", "Network round trip included", "Final-state check reported" ] } }