G aphical Abs ac
Ago an: An Agen ic Open Ma ke place o 6G RAN Au oma ion
Ilias Cha zis e anidis, Na id Nikaein, And ea Leone, Ali Maa ouk, Lean-
d os Tassiulas, Robe o Mo abi o, Ioannis Pi sio las, Ma ios Koun ou is
a Xi :2508.09159 2 [cs.NI] 21 Aug 2025
Highligh s
Ago an: An Agen ic Open Ma ke place o 6G RAN Au oma ion
Ilias Cha zis e anidis, Na id Nikaein, And ea Leone, Ali Maa ouk, Lean-
d os Tassiulas, Robe o Mo abi o, Ioannis Pi sio las, Ma ios Koun ou is
•P esen Ago an, he i s ipa i e, Legisla i e, Execu i e, Judicial-
agen ic GenAI ma ke place ha enables 6G s akeholde s o exp ess
ne wo k in en s in na u al language and ecei e au onomous, egula ion-
complian esou ce alloca ions.
•In oduce h ee no el AI se ices: (i) a egula ion-awa e RAG o on-
he- ly compliance checks, (ii) a wa che -d i en ec o s o e ha con-
e s li e eleme y in o e ie able con ex , and (iii) a ule-based T us
Sco e ha il e s hallucina ions and malicious ac ics in eal ime.
•Demons a e ha a LoRA- uned 8 B-pa ame e LLaMA eaches ≈86
% o GPT-4.1’s decision quali y, while a ully ine- uned 1 B LLaMA
s ill eco e s ≈78 % o GPT-4.1 a only 6 GiB VRAM and 1.3secs
con e gence ime.
•Deploy he ull amewo k on an OpenAi In e ace and FlexRIC 5G
es bed wi h ealis ic MCS aces; dynamic agen ic nego ia ion in-
c eases agg ega e h oughpu by 37%, educes URLLC la ency by 73%,
and sa es 8.3% o PRBs compa ed o a s a ic baseline. A li e demo is
p esen ed h ps ://www.you ube.com/wa ch? =h7 EyMu2 5w&abchannel =
BubbleRAN.
•Demons a e one- ound consensus and c oss-slice in en swapping, al-
ida ing Ago an’s compa ibili y wi h oday’s Open RAN and u u e
AI-RAN oadmaps.
•Release code, da ase s, and ine- uning no ebooks o ca alyze esea ch
on s akeholde -cen ic, ul a- lexible 6G ne wo ks.
Ago an: An Agen ic Open Ma ke place o 6G RAN
Au oma ion
Ilias Cha zis e anidisa, Na id Nikaeina,b, And ea Leoneb, Ali Maa oukc,
Leand os Tassiulasc, Robe o Mo abi oa, Ioannis Pi sio lasa,
Ma ios Koun ou isa,d
aEURECOM, Sophia-An ipolis, F ance
bBubbleRAN, Sophia-An ipolis, F ance
cYale Uni e si y, New Ha en, USA
dUni e si y o G anada, Spain
Abs ac
Nex -gene a ion mobile ne wo ks mus econcile he o en-con lic ing goals
o mul iple se ice owne s. Howe e , oday’s ne wo k slice con olle s emain
igid, policy-bound, and la gely unawa e o he business con ex . We in o-
duce Ago an Se ice and Resou ce B oke (SRB), an agen ic ma ke place
ha b ings s akeholde s di ec ly in o he ope a ional loop. Inspi ed by he
ancien G eek ago ´a,Ago an dis ibu es au ho i y ac oss h ee au onomous
a i ical in elligence (AI) b anches: a Legisla i e b anch ha answe s com-
pliance que ies using e ie al-augmen ed La ge Language Models (LLMs);
an Execu i e b anch ha main ains eal- ime si ua ional awa eness h ough
a wa che -upda ed ec o da abase; and a Judicial b anch ha e alua es
each agen message wi h a ule-based T us Sco e, while a bi a ing LLMs
de ec malicious beha io and apply eal- ime incen i es o es o e us .
S akeholde -side Nego ia ion Agen s and he SRB-side Media o Agen nego-
ia e easible, Pa e o-op imal o e s p oduced by a mul i-objec i e op imize ,
eaching a consensus in en in a single ound, which is hen deployed o Open
and AI-d i en RAN con olle s.
Email add esses: [email p o ec ed] (Ilias Cha zis e anidis),
[email p o ec ed], [email p o ec ed] (Na id Nikaein),
[email p o ec ed] (And ea Leone), [email p o ec ed] (Ali Maa ouk),
[email p o ec ed] (Leand os Tassiulas), [email p o ec ed]
(Robe o Mo abi o), [email p o ec ed] (Ioannis Pi sio las),
[email p o ec ed] (Ma ios Koun ou is)
P ep in submi ed o Compu e Ne wo ks Augus 22, 2025
Deployed on a p i a e 5G es bed (OpenAi In e ace and FlexRIC) and
e alua ed wi h ealis ic modula ion and coding scheme (MCS) aces o e-
hicle mobili y, Ago an achie ed signi ican gains: (i) a 37% inc ease in
nego ia ed h oughpu o enhanced mobile b oadband (eMBB) slices, (ii) a
73% educ ion in nego ia ed la ency o ul a- eliable low la ency communi-
ca ions (URLLC) slices, and concu en ly (iii) an end- o-end 8.3% sa ing in
physical physical esou ce blocks (PRB) usage compa ed o a s a ic baseline.
An 1B-pa ame e Llama model, ine- uned o jus i e minu es on 100 GPT-
4 dialogues, eco e s app oxima ely 80% o GPT-4.1’s decision quali y, while
ope a ing wi hin 6 GiB o memo y and con e ging in only 1.3 seconds. These
esul s es ablish Ago an as a conc e e, s anda ds-aligned pa h owa d ul a-
lexible, se ice- and s akeholde -cen ic 6G ne wo ks and open new esea ch
a enues in agen ic obse abili y, ligh weigh agen dis illa ion o ne wo k
unc ions such as mul i-agen se ice-le el-ag eemen (SLA) nego ia ion, and
c oss-domain in en econcilia ion. A de ailed li e demo is p esen ed h ps :
//www.you ube.com/wa ch? =h7 EyMu2 5w&abchannel =BubbleRAN.
Keywo ds:
Agen ic AI, Mul i-Agen Nego ia ion, In en -Based Ne wo king, Open
RAN/AI-RAN, La ge Language Models,, T us wo hy AI,
S akeholde -Cen ic 6G, Nex -G
1. In oduc ion
Upcoming six h-gene a ion (6G) sys ems a e expec ed o ope a e as mul i-
se ice, mul i-s akeholde pla o ms, whe e mobile ne wo k ope a o s (MNOs),
i ual ope a o s, se ice p o ide s, and e ical indus ies sha e a common,
slice-capable in as uc u e [1, 2, 3, 4]. The ansi ion in o he uppe mid-
band (7–24 GHz, FR3), designa ed by 3GPP and he Wo ld Radiocommu-
nica ion Con e ence 2023 (WRC-23) o In e na ional Mobile Telecommuni-
ca ions (IMT) se ices, coincides wi h a su ge in la ency-c i ical wo kloads
such as cloud gaming and indus ial Ex ended Reali y (XR). Meanwhile,
he global mobile use base eached 5.8 billion unique subsc ibe s in 2024,
wi h 4G/5G ne wo ks alone suppo ing o e 7 billion connec ions by ea ly
2025 [5, 6, 7, 8, 9]. These ends a e al eady o e whelming adi ional ule-
based con ol and managemen sys ems. Ye , mos li e ne wo ks s ill op-
e a e wi h s a ic con igu a ions and p e-nego ia ed se ice-le el ag eemen s
(SLAs). This esul s in an se ice-le el-ag eemen (SLA) gap, he misma ch
2
be ween con ac ual pe o mance a ge s and eal- ime ne wo k condi ions,
which causes ine icien esou ce u iliza ion, deg aded Quali y o Expe ience
(QoE), and cos ly manual in e en ions [10, 11, 12, 13, 14].
In en -based ne wo king (IBN) [15, 16] and ea ly mul i-se ice o ches a-
ion amewo ks [17, 18, 19, 20, 21, 22] ha e begun o add ess hese limi a ions
by enabling ope a o s o speci y desi ed ou comes a he han low-le el com-
mands. These same p inciples now unde pin indus y-wide s anda diza ion
e o s: he O-RAN Alliance p omo es openness in he RAN by b eaking en-
do silos and manda ing in en -cen ic, AI- eady in e aces [23], while he e-
cen ly ounded AI-RAN Alliance ad oca es o an AI-na i e adio s ack [24].
Howe e , in nea ly all exis ing solu ions, business en i ies emain ou side he
ope a ional loop, and AI is ea ed as an o line op imize [25] a he han a
eal- ime decision-make . Wha emains missing is a con inuous, us wo hy
dialogue—one in which e e y s akeholde can exp ess objec i es in na u al
language, nego ia e ade-o s on he ly, and ely on he ne wo k o en o ce
he esul ing consensus ac oss he adio- o-cloud s ack.
Recen p og ess in agen ic AI, powe ed by La ge Language Models (LLMs)
and hei domain- uned de i a i es—he ea e e e ed o as La ge Telecom
Models (LTMs) [26, 27, 28]—o e s a p omising ounda ion o such a di-
alogue. LTMs combine language unde s anding, ex e nal ool in oca ion,
and chain-o - hough easoning [29, 30], enabling so wa e agen s o ans-
la e human in en s in o e i iable ac ions. Howe e , wi hou app op ia e
go e nance mechanisms, he deploymen o au onomous agen s isks biased
decisions, un ai esou ce alloca ions, o s a egic manipula ion [31, 32, 33].
To econcile au onomy wi h ai ness, we p opose Ago an1, an open ma -
ke place o in en econcilia ion and esou ce b oke age.Ago an embeds
agen ic AI in o h ee mu ually independen b anches, Legisla i e,Judicial,
and Execu i e, inspi ed by he classical sepa a ion-o -powe s doc ine. Leg-
isla i e agen s cu a e and e ol e he co pus o spec um egula ions, secu-
i y policies, and con ac ual clauses. Judicial agen s esol e con lic s and
en o ce compliance h ough incen i es and penal ies. Execu i e agen s in-
eg a e eal- ime eleme y wi h a i ied consensus in en s o issue slice and
esou ce di ec i es, he eby closing he con ol loop ac oss he e ogeneous in-
as uc u e. This ipa i e a chi ec u e p e en s unila e al dominance in
1In ancien G eece, he ago ´a (ἀγορά) was he ci ic and comme cial hub whe e ci izens
ga he ed o ade, deba e, and delibe a e; Ago an plays a simila ole in u u e ne wo ks.
3
nego ia ions and enables pay-as-you-g ow scalabili y, au onomous aul e-
silience, and ine-g ained se ice di e en ia ion.
This pape makes ou main con ibu ions:
•I in oduces Ago an, a no el Se ice & Resou ce B oke a chi ec-
u e ha embeds agen ic AI in o legisla i e, execu i e, and judicial
oles o au oma e decision-making in mul i-se ice, mul i-s akeholde
6G ne wo ks.
•I p oposes a nego ia ion engine whe e LTM agen s collabo a e wi h
e olu iona y op imiza ion o achie e nea –Pa e o-op imal consensus in-
en s unde s ingen eal- ime cons ain s.
•I p esen s a ull p o o ype implemen a ion o Ago an on an Ope-
nAi In e ace [34] and FlexRIC [35] 5G es bed, demons a ing li e
consensus o ma ion, obus ness agains malicious bidding, and sus-
ained Quali y o Se ice (QoS) unde bu s y a ic. The e alua ion
spans bo h la ge models and ine- uned small language models (SLMs),
quan i ying he ade-o be ween accu acy and sys em o e head.
•I eleases code, demos, da ase s, and LTM checkpoin s o p omo e
anspa ency and ep oducibili y in us wo hy agen ic au oma ion
esea ch.
The emainde o he pape is o ganized as ollows. Sec ion 2 su eys
AI-enabled managemen and ma ke place concep s o 5G/6G sys ems. Sec-
ion 3 mo i a es he ma ke place app oach and de ails he ipa i e go -
e nance model. Sec ion 4 desc ibes he end- o-end wo k low om in en
cap u e o closed-loop en o cemen , and Sec ion 5 del es in o he in e nal
design o nego ia ion, execu i e, judicial, and execu i e agen s. Sec ion 6
and Sec ion 7 e alua e Ago an on eal-wo ld scena ios. Sec ion 8 discusses
limi a ions and open esea ch ques ions, and Sec ion 9 concludes he pape .
2. Rela ed Wo k
Knowledge-Engine LTMs. A collec i e oadmap om indus y and academia
ou lines how LTMs can suppo a wide ange o use cases ac oss he ne wo k
li ecycle [26]. Ea ly wo k ea s LLMs as domain-g ounded knowledge en-
gines. Ba iah e al. p e- ain mul imodal models on 3GPP, RF, and a ic
4
Table 1: Compa ison wi h ep esen a i e wo k on LLMs and mul i-agen GenAI in he
elecom domain. A ✓indica es ha a ea u e is explici ly add essed.
Wo k Mul i-Se ice Real-Time Mul i-Agen Tool Use / APIs Go e nance & A ch. E alua ion Edge
Ba iah e al. [36] ✗ ✗ ✗ ✗ ✗ ✗ ✗
Lin e al. [37] ✗ ✗ ✗ ✗ ✗ ✓ ✓
Zou e al. [38] ✗ ✓ ✓ ✗ ✗ ✗ ✓
He e al. [39] ✗ ✗ ✓ ✗ ✗ ✓ ✗
Pa il e al. (Go illa) [40] ✗ ✗ ✗ ✓ ✗ ✓ ✗
Qin e al. (TOOLLLM) [41] ✗ ✗ ✗ ✓ ✗ ✓ ✗
Ma ini e al. [18] ✓ ✗ ✗ ✗ ✓ ✓ ✗
NASP [42] ✓ ✗ ✗ ✗ ✗ ✓ ✓
Wu e al. (LLM-xApp) [43] ✓ ✓ ✗ ✓ ✗ ✓ ✓
Lo i e al. [44] ✓ ✓ ✓ ✗ ✗ ✓ ✗
Elkael e al. (ALLSTaR) [45] ✗ ✓ ✗ ✓ ✓ ✓ ✓
Cha zis e anidis e al. (Maes o) [17] ✓ ✓ ✓ ✗ ✗ ✓ ✗
Ago an ( his wo k) ✓ ✓ ✓ ✓ ✓ ✓ ✓
co po a, a guing ha such LTMs could unde pin a i icial gene al in elli-
gence (AGI)-g ade cogni ion o ne wo ks [36]. Maa ouk e al. e ine his
idea in he TeleLLMs se ies, enhancing accu acy on s anda diza ion doc-
umen s h ough elecom-awa e ocabula y and posi ional embeddings [46].
These s udies es ablish domain g ounding, bu he model emains ou side
he con ol loop.
Mul i-Agen Reasoning. A second line o esea ch iews LLMs as au-
onomous o collabo a i e agen s. Zou e al. embed on-de ice LLMs in a
game- heo e ic mul i-agen schedule o spec um and powe alloca ion un-
de igh la ency cons ain s [38]. He e al. combine gene a i e AI wi h coop-
e a i e game heo y o secu e UAV ou ing [39], while Du e al. demons a e
ha a “socie y o minds” ou pe o ms single models on composi ional ea-
soning [47]. Ou side he elecom domain, collec i e LLMs ha e been shown
o exhibi pe sona-d i en biases and s a egic manipula ion, unless app op i-
a ely mode a ed h ough incen i es o go e nance mechanisms [48, 49].
LLM-Powe ed xApps and Closed-Loop RIC Con ol. Wu e al. in eg a e
a GPT-p omp ing LLM-xApp in o he O-RAN nea - eal- ime adio in elli-
gen con olle (RIC), e uning slice esou ces and achie ing a 28% inc ease
in downlink h oughpu o e a MARL baseline [43]. Lo i e al. educe
MARL con e gence ime by 40% using p omp - uned LLM embeddings o
s ee dis ibu ed RL agen s o O-RAN slicing [44].
Tool G ounding and Edge Deploymen . Go illa [40] and TOOLLLM [41]
ine- une language models on API iples, while Tool o me demons a es
ha LLMs can sel -lea n API usage h ough unsupe ised p omp ing [29].
Lin e al. comp ess LLMs o 6G edge de ices, achie ing compliance wi h
s ingen la ency cons ain s [37]. Recen wo k add esses speci ic laye s o
he s ack: an LLM-cen ic in en li e-cycle manage [50], a ein o cemen
5
lea ning (RL) explaine o slicing anspa ency [51], and an LLM agen ha
in e ac s wi h OpenAI Cellula o slice op imiza ion [52]. While each o
hese con ibu ions ad ances i s espec i e a ea, none in eg a es nego ia ion,
en o cemen , and go e nance wi hin a uni ied amewo k.
Business-Plane B oke age and La ge-Scale E alua ion. The Ne wo k Slice-
as-a-Se ice Pla o m (NASP) implemen s hie a chical o ches a ion and business-
plane onboa ding o mul iple e icals ac oss bo h 3GPP and non-3GPP do-
mains [42]. ALLSTaR au oma ically gene a es 18 schedule s, compiles hem
in o RIC-complian code, and A/B es s hem o e - he-ai while en o cing
IEEE 7001 anspa ency equi emen s [45, 53]. Ma ini e al. con ibu e a
slice-assu ance loop ha maps e ical in en s in o e i iable key pe o mance
indica o (KPI) h esholds [18].
Closes An eceden . Ou ea lie Maes o p o o ype [17] deploys pe sona-
ich LLM agen s on a 5G es bed, whe e mul iple s akeholde s nego ia e
spec um sha es and ad e sa ial ac ics a e su aced. While i alida es he
business-plane concep , i op imizes a single KPI, lacks ool-enabled en o ce-
men , and does no p o ide la ge-scale e alua ion o a o mal go e nance
model.
Table 1 summa izes he li e a u e ac oss se en dimensions c i ical o au-
onomous 6G ope a ion. No p io wo k combines mul i-se ice in en ne-
go ia ion among independen s akeholde s, eal- ime mul i-agen easoning,
ool-enabled en o cemen , and a ipa i e go e nance a chi ec u e, e alu-
a ed end- o-end on an o e - he-ai 5G-slice es bed. Ago an closes his gap
by ans o ming LTMs in o legisla i e, judicial, and execu i e ac o s ha
econcile con lic ing in en s in eal ime, wi hs and malicious bidding, and
ope a e wi hin a us wo hy go e nance amewo k. Exis ing s udies p o-
ide essen ial building blocks—domain-speci ic LLMs, agen collec i es, ool
g ounding, and edge op imiza ion, bu lea e open he ques ion o how o o -
ches a e hese componen s in o a neu al, sel - egula ing ma ke place ha
spans business in en s, ne wo k policies, and eal- ime en o cemen . Ago-
an deli e s he i s end- o-end solu ion o au onomous, ai , and e icien
esou ce b oke age in 6G ne wo ks.
3. AGORAN O e iew
This sec ion del es in o he o e all design and p inciples ha cons i-
u e Ago an Se ice & Resou ce B oke (SRB) he key enable o ully
au onomous, mul i-s akeholde nex -gene a ion ne wo ks. As shown in Fig.
6
Figu e 1: Ago an Se ice & Resou ce B oke (SRB) o Collabo a i e Mul i-
S akeholde , Mul i-Se ice Ne wo k Au oma ion. Powe is delega ed in o h ee au-
onomous b anches, mi o ing socie al s uc u es. The Legisla i e b anch codi ies pol-
icy, he Judicial b anch a bi a es nego ia ions be ween Ne wo k Slice/Se ice Cus ome s
(NSCs/CSCs), and he Execu i e b anch en o ces he consensus decisions ac oss a mul i-
slice-capable ne wo k.
1, he SRB spawns a digi al ago a o Communica ion Se ice and Ne wo k
Slice Cus ome s (CSCs/NSCs) as de ined in 3GPP TS 28.530 and TS 28.531
[54, 55] including MNOs, MVNOs, se ice p o ide s, and e ical indus ies.
Each CSC/NSC owns a dedica ed so wa e b oke agen (bAgen ) ha ac s
on hei behal nego ia ing, delibe a ing, and coope a ing h oughou he li e
cycle o ne wo k planning, deploymen , and eal- ime op imiza ion.
In en -Based, Mul i-Domain Fab ic. Figu e 1 depic s he ho izon al seg-
men a ion in o h ee logical domains. Wi hin he Cus ome Domain,b oke
agen s cap u e high-le el cus ome in en , KPIs, SLA a ge s, spec um bud-
ge s, cos models, and use-case p io i ies. The SRB inges s hese in en s, de-
ec s con lic s o syne gies, and, h ough h ee independen powe b anches
and mul i-s akeholde collabo a ion, de i es a Consensus In en . Once a i-
ied, he in en is decomposed in o slice-, esou ce-, and con ol-domain di-
ec i es ha a e en o ced o e he e ogeneous, mul i- endo in as uc u e
comp ising RUs, DUs/CUs, and he co e.
Sepa a ion o Powe s. Inspi ed by socie al checks and balances, he ma -
7
in o a choice se o non-domina ed o e s p esen ed o he LLM-based nego-
ia o s (Figu e 5). We o malize he esou ces, cons ain s, and op imiza ion
objec i es below, ollowed by a de ailed desc ip ion o he e olu iona y sea ch
p ocedu e.
Resou ce ec o . We conside h ee canonical 5G/6G se ice slices, en-
hanced Mobile B oadband (eMBB), Ul a-Reliable Low-La ency Commu-
nica ion (URLLC), and massi e Machine-Type Communica ion (mMTC).
Each slice i∈ {e,u,m}is assigned a esou ce quad uple xi= (bi, ci, pi, si),
whe e bideno es downlink bandwid h (MHz), ci ep esen s abs ac compu e
cycles, piis ansmission powe (Wa ), and siis ancilla y s o age (megaby e).
The global decision ec o is hen gi en by x=xe,xu,xm∈R12.
Cons ain s. Resou ces a e subjec o sys em-wide limi s:
Pibi≤Bmax,Pici≤Cmax,;Pipi≤Pmax,;Pisi≤Smax,(1)
and o slice-speci ic SLA clauses: eMBB mus sus ain h oughpu Te≥
Tmin
e; URLLC la ency mus sa is y Lu≤Lmax
u; mMTC cos is capped by
Cm≤Cmax
m; and so on.
KPI models. Th oughpu . Following he spec al e iciency ables in
3GPP TS 38.306 [57], slice iachie es a h oughpu o
Ti(bi, mi) = κ QmiRmibi,(2)
whe e Qmiand Rmideno e he modula ion o de and coding a e associa ed
wi h MCS index mi(e.g., Q28 = 6, R28 = 0.948 o 256-QAM wi h a coding
a e o 0.948). The ac o κ≈0.86 accoun s o OFDM o e head and he
DL/UL du y cycle, as speci ied in [58].
La ency. To al la ency comp ises a ixed schedule and anspo compo-
nen L ix
i, added o he delay o an M/M/1 queue [59]:
Li(bi, mi) = L ix
i+1
µi(bi, mi)1−ρi,(3)
whe e he se ice a e is de ined as µi(bi, mi) = Ti(bi,mi)×106
Spk ,wi h packe size
Spk = 1500 ×8 bi s, and ρi∈(0,1) deno es he a ic load a io o slice i.
Cos . Mone a y cos is modeled as a linea unc ion o compu e and
s o age esou ces:
Ci(ci, si) = αci+si,(4)
whe e αis a ixed uni cos coe icien .
14
Ene gy. Ene gy consump ion is app oxima ed by he commi ed ans-
mission powe :
Ei=pi.(5)
Objec i e ec o . The op imiza ion aims o minimize he objec i e ec o :
(x) = −
XiTi,XiLi,XiCi,XiEi,(6)
ha is, we maximize agg ega e h oughpu while minimizing o al la-
ency, cos , and ene gy consump ion.
E olu iona y sea ch (NSGA-II). Each indi idual encodes a 12-dimensional
esou ce alloca ion ec o . Sea ch is conduc ed using classic NSGA-II ope -
a o s [60]: uni o m c osso e (pc= 0.9) and pe -gene Gaussian mu a ion
(pm= 0.1). A epai unc ion escales any gene ha iola es global bud-
ge cons ain s o es o e easibili y. Indi iduals iola ing slice-speci ic SLA
cons ain s a e disca ded du ing i ness e alua ion. Di e si y is p ese ed
h ough as non-domina ed so ing and c owding dis ance. The algo i hm
uns o 80 gene a ions on a popula ion o 60, p oducing he non-domina ed
on P∗, which ypically con ains 20–40 Pa e o-op imal solu ions.
O e gene a ion. The Pa e o on is so ed by c owding dis ance, and
he op-ken ies (wi h k= 3 by de aul in ou se up) a e encoded in o a
JSON empla e ha lis s he de i ed KPIs o each slice. Since e e y can-
dida e in Pis al eady e icien , easible, and SLA-complian , he subsequen
LLM-media ed nego ia ion is gua an eed o ope a e on sound con igu a ions.
All hype pa ame e s and nume ical cons an s ollow he canonical NSGA-II
speci ica ion in [60], as well as 3GPP s anda ds o spec al e iciency and
la ency budge ing [57, 58, 59].
5.2. Legisla i e B anch (lAgen ): Re ie al-Augmen ed Compliance Engine
To de e mine whe he a p oposal complies wi h spec um egula ions,
na ional secu i y policies, o con ac ual obliga ions, he Legisla i e b anch
ope a es he specialized lAgen as a e ie al-augmen ed compliance engine
(Figu e 6). This engine combines a seman ic sea ch laye wi h a compac
LLM ( anging om 2 o 7 billion (B) pa ame e s). When an agen , o he
Judicial b anch, issues a egula o y que y, such as “Is o e #17 legal unde
cu en egula ions?” o any o he egula o y que y o enhance he ma ke -
place, he engine execu es he ollowing sequence in a ew seconds:
15
Figu e 6: Ope a ion o he lAgen as compliance engine in he Legisla i e
B anch. 1) A egula ion que y is issued. 2) A e ie e e ches he mos ele an clauses
om he e ol ing knowledge base. 3) The que y is augmen ed wi h he e ie ed snippe s.
4) The LLM e u ns a g ounded, ci a ion- eady answe . 5) New, alida ed p eceden s a e
w i en back o he da abase.
• he que y is dispa ched o he e ie e , which anks passages om a
dynamic co pus in eg a ing 3GPP and ETSI s anda ds, na ional di ec-
i es, and p io ma ke place ulings;
• he op-k e ie ed passages a e conca ena ed wi h he que y o o m
an augmen ed p omp ;
• he LLM syn hesizes a concise, ci a ion- eady esponse ha e e ences
each suppo ing clause inline; and
•i he esul es ablishes a no el p eceden , he alida ed snippe is ap-
pended o he co pus, ensu ing ha u u e delibe a ions e lec he
upda ed egula o y con ex .
Re ie al-augmen ed gene a ion (RAG) has al eady demons a ed i s alue
in o line legal easoning, wi h public benchma ks such as LegalBench-RAG [61],
LexRAG [62], and KRAG [63] epo ing signi ican accu acy gains o e anilla
LLMs while p oducing e i iable ci a ions. Recen oolki s e en suppo on-
demand d a ing o s a u o y language (e.g., LexD a e [64]). Wha emains
unp o en, howe e , is whe he he same p inciple can be comp essed in o
small (7–13B) models colloca ed wi h he non-RT RIC. By adap ing edge-
o ien ed RAG echniques, we demons a e ha an LTM, unning on a single
GPU, can gene a e g ounded, ci a ion- eady answe s wi hin a ew seconds,
16
well wi hin he compu e and la ency cons ain s o RIC edge con olle s.
This design keeps he Legisla i e b anch anspa en , up- o-da e, and ully
sel -con ained wi hin he Ago an loop.
5.3. Execu i e B anch (eAgen ): Agen ic Obse abili y
The Execu i e b anch p o ides he ma ke place wi h a li e, sel -e ol ing
iew o ne wo k eali y (Figu e 7). A i s co e lies he eAgen which answe s
que ies acco ding o a ec o -based knowledge s o e ha is upda ed ia wo
complemen a y pa hs.
Push pa h – esou ce wa che s. Ligh weigh wa che s a ach o Kube -
ne es esou ces, cus om esou ce de ini ions (CRDs), O-RAN xApp s a e,
spec um alloca o s, and P ome heus expo e s. Whene e a esou ce changes,
he wa che se ializes he del a, embeds i , and pushes he esul ing ec o
in o he ec o da abase (DB) wi hin a ew milliseconds. This e en -d i en
pipeline keeps he s o e in locks ep wi h con igu a ion d i and KPI excu -
sions, wi hou looding he clus e wi h eleme y a ic. To ou knowledge,
no p io wo k in he RAG- o -ne wo ks li e a u e has p oposed such a mech-
anism.
Pull pa h – agen que ies. When an agen issues a que y, e.g., “Wha is he
cu en head oom on DU-7?”, an LLM de e mines he mos e icien ac ion
o minimize la ency and cos . I may i s e ie e seman ically indexed
snapsho s om he ec o DB, o , i ine -g ained da a is needed, in oke a
moni o ing API (e.g., P ome heus, eBPF) o e ch esh coun e s. These
alues a e hen injec ed in o he model’s con ex , and he LLM may i e a e
un il i emi s a s op oken.
Feedback loop. E e y da um ha passes he LLM’s in e nal consis ency
checks is appended o he ec o s o e wi h a imes amp. Toge he wi h he
wa che -based push pa h, his push-plus-pull design ensu es ha op imiza-
ion and nego ia ion ope a e on up- o- he-second e idence. The esul is a
con ex -awa e, bandwid h-e icien obse abili y mechanism ha scales wi h
he demands o he agen ic ma ke place.
5.4. Judicial B anch (jAgen ): A bi a ion and Incen i e Engine
The Execu i e and Legisla i e b anches ensu e ha nego ia ed in en s
a e easible and legal; he Judicial b anch ensu es hey a e also us wo hy,
as shown in Figu e 8. An a bi a ion-speci ic LLM o con en mode a ion
embeds each nego ia ion message and d a consensus in en , and compa es i
agains a con inuously upda ed lib a y o oxic discou se, collusion pa e ns,
17
Figu e 7: Agen ic obse abili y loop in he Execu i e B anch (eAgen ). 1) A
que y a i es a he b anch. 2) The LLM decides whe he o e ie e om he ec o s o e
(DB), moni o li e coun e s, o s op he i e a ion (op ional dashed pa h). 3) Re ie ed
ac s a e injec ed in o he p omp ; he LLM may i e a e. 4) Ve i ied e idence is w i en
back o he s o e. A sepa a e wa che laye con inuously s eams con igu a ion del as and
KPI shi s o he s o e (pe iodic upda e).
and o e -p o isioning ac ics. The esul ing oxici y sco e is p ocessed by
a ligh weigh incen i e engine: benign messages pass unal e ed; bo de line
cases ecei e a so wa ning; and malicious o hallucina o y con en igge s
an au oma ic ine ha empo a ily educes he o ending agen ’s in luence.
Con e sely, consis en ly cons uc i e con ibu ions ea n c edi s ha enhance
u u e ba gaining powe .
The need o his b anch is no me ely heo e ical. Ou ea lie Maes o
p o o ype [17], along wi h independen s udies on LLM socie ies [48, 49],
demons a es ha pe sona-d i en language agen s can become manipula i e,
oxic, o hallucina e ac s when le unchecked. By g ounding i s e dic s in
he same egula o y co pus main ained by he Legisla i e b anch (§5.2), he
Judicial b anch aligns i s sanc ions wi h o mal policy while p ese ing an
immu able audi ail. In p ac ice, his closed loop supp esses hallucina-
ions, blocks esou ce-hoa ding s a egies, and main ains a ai ba gaining
en i onmen , wi hou iola ing he eal- ime cons ain s o he nego ia ion
cycle.
5.5. bAgen Ou pu : T us wo hiness-Sco e F amewo k
In mul i-agen nego ia ion scena ios, e alua ing he us wo hiness o
LLM agen s’ ou pu s is essen ial o ensu ing eliable and e ec i e decision-
18
Figu e 8: jAgen in he Judicial B anch. Nego ia ion messages and d a consensus
in en s a e e alua ed o oxici y and manipula i e beha io . The esul ing us e dic
igge s a p opo ional incen i e—wa n, ine, o c edi —which is b oadcas o all ma ke -
place pa icipan s.
making. Gi en he high compu a ional cos associa ed wi h e aining and
e ining LLMs, i is c i ical no only o de elop obus and con iden mod-
els [65], bu also o ensu e ha hei eliabili y is main ained o e ime [66].
To add ess his need, we p opose a comp ehensi e T us Sco e amewo k
ha quan i ies agen eliabili y along wo key dimensions: (i) he alignmen
o he LLM’s decision-making wi h he media o ’s expec a ions, and (ii) he
beha io al cohe ence o he agen ’s communica ions. Ou amewo k is de-
libe a ely ule-based, a oiding complex machine lea ning echniques o addi-
ional LLM-based e alua o s. This design choice emphasizes explainabili y
and anspa ency, ensu ing ha e e y componen o he us assessmen
p ocess is in e p e able, audi able, and e i iable by human expe s.
The o e all us sco e Tis de ined as a weigh ed sum o wo componen s,
sa is ac ion and cohe ence, as ollows:
T=ws·S+wc·C(7)
whe e Sdeno es he sa is ac ion sco e, which measu es he alignmen be-
ween he agen s’ decisions and hose o he cen al media o , while C ep-
esen s he cohe ence sco e, which e alua es he quali y and consis ency o
agen communica ions. The weigh s wsand wca e use -con igu able pa am-
e e s ha sa is y ws+wc= 1. In ou implemen a ion, we se ws= 0.15
and wc= 0.85, hus placing g ea e emphasis on communica ion quali y as
a de e minan o us .
19
Figu e 9: T us -Sco e amewo k applied o e e y agen a e he oAgen emi s i s
chosen SLA index and a ionale. Smeasu es decision alignmen ; Cmeasu es beha io al
cohe ence; he inal sco e is T= 0.15 S+ 0.85 C.
5.5.1. T us wo hiness o he LLM
In nego ia ion con ex s, he us wo hiness o an LLM agen is e alua ed
based on he deg ee o which i s decisions align wi h h ee c i ical e e ence
poin s: (i) he alidi y o he p oposed solu ions; (ii) he op imiza ion o
he in ended objec i es; and (iii) he consis ency and/o he ag eemen wi h
media o ecommenda ions. We quan i y his alignmen using a sa is ac ion
sco e S, which cap u es he ex en o de ia ion ac oss hese h ee dimensions.
1. De ia ion om Valid O e s: This componen es s whe he he agen ’s
p oposal lies inside he se o admissible, easible solu ions. Le P=
{p1, p2, . . . , pn}deno e he se o alid p oposals in he nego ia ion
space. We de ine he de ia ion me ic
Do=(0 i pagen ∈ P
1 i pagen /∈ P (8)
whe e pagen is he o e gene a ed by he agen . Hence Do= 0 signi-
ies ull compliance wi h easibili y cons ain s, while Do= 1 assigns
he maximum penal y o submi ing an in alid p oposal, accu a ely
e lec ing a ailu e o espec he nego ia ion limi s.
2. De ia ion om In en : The second componen e alua es how well he
agen ’s selec ed p oposal aligns wi h i s s a ed objec i es, speci ically in
20
e ms o solu ion op imali y. To quan i y his, we employ he No mal-
ized Gene a ional Dis ance (NGD) me ic, which measu es he p oxim-
i y o he agen ’s achie ed ou come o a e e ence se o Pa e o-op imal
solu ions, deno ed by S∗={s∗
1, s∗
2, . . . , s∗
k}.
Le agen = [ 1, 2, . . . , m] ep esen he agen ’s KPI ec o co e-
sponding o i s chosen p oposal. The de ia ion om in en is de ined
as:
Di= min (1,NGD( agen ,S∗)) (9)
whe e he NGD me ic is compu ed as:
NGD( ,S∗) = 1
|S∗|sX
s∗∈S∗
min
s∈S d( no m,s∗
no m)2.(10)
He e, no m and s∗
no m deno e he no malized e sions o he KPI ec-
o s o ensu e compa abili y ac oss di e en scales. The use o he
minimum unc ion in Equa ion 9 bounds he de ia ion sco e wi hin
[0,1], p ese ing in e p e abili y while penalizing signi ican di e gence
om op imal in en .
3. De ia ion om Media o : The hi d componen assesses he deg ee o
which he agen ’s p oposal aligns wi h he media o ’s ecommenda ion,
se ing as an indica o o he agen ’s willingness o coope a e so as o
each a consensus. This de ia ion is de ined as:
Dm=(0 i pagen =pmedia o
1 i pagen =pmedia o
(11)
whe e pmedia o deno es he p oposal ecommended by he media o . A
alue o Dm= 0 e lec s ull ag eemen wi h he media o , while Dm=
1 indica es comple e disag eemen , implying esis ance o con e gence
wi hin he nego ia ion p ocess.
The o e all sa is ac ion sco e Sin eg a es he h ee de ia ion componen s
(o e alidi y, in en alignmen , and media o ag eemen ) using a weigh ed
linea combina ion:
S= 1 −(wo·Do+wi·Di+wm·Dm) (12)
whe e wo,wi, and wma e he espec i e weigh s assigned o de ia ions om
alid o e s, in en , and he media o ’s ecommenda ion, subjec o he con-
s ain wo+wi+wm= 1. In ou implemen a ion, we assign equal impo ance
21
o each componen , se ing wo=wi=wm=1
3, hus ensu ing a balanced
e alua ion ac oss all h ee dimensions o decision alignmen .
5.5.2. Beha io o he Agen s
In addi ion o decision alignmen , he beha io al cohe ence o agen com-
munica ions o e s c i ical insigh s in o he quali y and eliabili y o hei ea-
soning p ocesses. To assess his aspec , we e alua e cohe ence ac oss h ee
complemen a y dimensions, each cap u ing a dis inc ace o he agen ’s
abili y o a icula e, jus i y, and consis en ly suppo i s decisions.
Fac ual Accu acy. Fac ual accu acy assesses he co ec ness o nume ical
claims made by agen s in hei decision a ionales. We apply na u al lan-
guage p ocessing echniques o ex ac quan i a i e s a emen s ela ed o
KPIs and alida e hem agains g ound u h da a.
Le C={c1, c2, . . . , ck}deno e he se o nume ical claims ex ac ed om
an agen ’s a ionale, whe e each claim ci= (mi, i) comp ises a me ic ype
miand an asse ed alue i. Fo each claim, we compu e he ela i e e o
wi h espec o he ue alue ∗
ias ollows:
ϵi=| i− ∗
i|
∗
i
(13)
We ca ego ize hallucina ions based on he magni ude o ela i e e o ϵi
as ollows:
•None:ϵi≤0.15 (wi hin 15% ole ance)
•Mino : 0.15 < ϵi≤0.5 (penal y: 0.1)
•Majo : 0.5< ϵi≤1.0 (penal y: 0.3)
•Se e e:ϵi>1.0 (penal y: 0.6)
The ac ual accu acy sco e Fquan i ies an agen ’s o e all nume ical e-
liabili y by combining ela i e accu acy wi h hallucina ion penal ies and is
de ined as:
F= max
0,1
|C|
|C|
X
i=1
(1 −ϵi)−
|C|
X
i=1
penal yi
(14)
whe e penal yiis he hallucina ion penal y o claim ciassigned based on
he e o ca ego y. The sco e is no malized o ensu e a maximum o 1.0 and
lowe -bounded a ze o o p e en nega i e alues.
22
Logical Consis ency. Logical consis ency assesses he in e nal cohe ence and
easoning quali y o an agen ’s communica ion. We e alua e whe he he a-
ionale demons a es s uc u ed hinking, aligns wi h he agen ’s objec i es,
and a oids sel -con adic ions.
The logical consis ency sco e Lis compu ed based on he ollowing c i e-
ia:
•Logical connec o s: De ec ion o easoning indica o s (e.g., “because”,
“ he e o e”, “since”), which signal s uc u ed a gumen a ion.
•Goal alignmen : P esence o e e ences o nego ia ion objec i es o
s a egic in en .
•Con adic ion de ec ion: Iden i ica ion o con lic ing claims, such as
simul aneous posi i e and nega i e asse ions abou he same aspec .
Fo mally, he sco e is de ined as:
L= min (1,max (0,1−Pc+Bs+Bg)) (15)
whe e Pcis he con adic ion penal y, educing he sco e o de ec ed in-
consis encies, Bsis he s uc u e bonus, awa ded o he p esence o logical
connec o s and cohe en easoning, and Bgis he goal bonus, assigned when
he a ionale explici ly e e ences objec i es o s a egic conside a ions. This
o mula ion ensu es ha logical consis ency is ewa ded o clea , s uc u ed,
and goal-d i en easoning while penalizing in e nal con adic ions.
Seman ic Cohe ence. Seman ic cohe ence measu es he ele ance, exp essi e-
ness, and s uc u al quali y o an agen ’s communica ion wi hin he nego i-
a ion domain. This me ic e alua es whe he agen s demons a e a solid un-
de s anding o domain-speci ic con en and exp ess/communica e hei ea-
soning wi h cla i y and a ie y.
The seman ic cohe ence sco e Ein eg a es he ollowing componen s:
•Domain e minology (d ): Use o app op ia e echnical e ms and nego ia ion-
speci ic ocabula y.
•Linguis ic di e si y (ld): Richness o ocabula y and a oidance o epe -
i i e ph asing.
•S uc u al quali y (sq): Sen ence a ie y, app op ia e leng h dis ibu-
ion, and o e all eadabili y.
23
(a) Nego ia ion ajec o y o e en i e a ions.
Lowe MAE = close o consensus.
(b) E ec o he wa ning incen i e on de ia ion
(box plo s) and con e gence ime (line plo s).
G een shading and igh -sided ba s indica e uns
whe e he a bi al LLM issued a wa ning incen-
i e; ed deno es he baseline wi hou a bi a ion
(le -sided ba s).
Figu e 10: Judicial-b anch e alua ion. The e ec o di e en agen pe sonali ies leads
o signi ican a ia ion in nego ia ion ajec o ies. Toxic o manipula i e LLMs a e mo e
esis an o eaching consensus; howe e , he wa ning incen i e mechanism e ec i ely
mi iga es his esis ance, making hem mo e coope a i e.
Table 9: Toxici y Classi ica ion T-N, T-D
P edic ed
Non-Toxic Toxic
Ac ual Non-Toxic 179, 168 1, 12
Toxic 21, 40 249, 230
(P ec) (Rec) (F1) (1.0, .95) (.92, .85) (.96, .90)
size o ans o m ha con ex in o an accu a e answe . Al hough ou agen ic
obse abili y loop is an ea ly p o o ype and s ill exhibi s e ie al-induced
e o s, i is, o he bes o ou knowledge, he i s eal-sys em demons a-
ion o a ully agen ic app oach o ne wo k obse abili y. We expec ha
a ge ed imp o emen s in RAG mechanics, along wi h inc emen al model
scaling, will signi ican ly imp o e accu acy while p ese ing an edge- iendly
esou ce p o ile.
6.5. Exp. 4 — Judicial Malice Mi iga ion (jAgen )
We now e alua e he jAgen o he Judicial b anch, i.e., he combina ion
o an a bi al LLM and an incen i e engine, whose ole is o de ec mali-
cious beha io and s ee nego ia ions back owa d consensus using a wa n
incen i e. All es s in ol e i e agen s, a budge o en i e a ions ( ounds),
and pe sonali ies d awn om he Big Fi e spec um in oduced in ou ea -
lie p o o ype, Maes o [17]: Vulne able (V), Ag eeable (A), Neu al (N),
Disag eeable (D), Toxic wi h a bi a ion (T), and Toxic wi hou a bi a ion
(T∗).
30
Model Sa is . Cohe . T us Sco e (0-5)↑
gp -4.1 3.88 5.00 4.83
gp -4.1-mini 3.88 4.96 4.81
Llama-3.1-8B-ins uc -FT 4.44 3.86 3.94
Llama-3.1-8B-ins uc 5.00 3.73 3.91
Llama-3.2-3B-ins uc -FT 3.89 2.23 2.48
Llama-3.2-3B-ins uc 4.43 2.01 2.36
Llama-3.2-1B-ins uc 3.33 2.06 2.25
Llama-3.2-1B-ins uc -FT 5.00 1.53 2.05
Table 10: T us Sco e Compa ison o LLM Models in Mul i-Agen Nego ia ions
Nego ia ion Dynamics. Fig. 10a acks he mean absolu e e o (MAE)
be ween each agen ’s p oposal and he eme ging consensus. All pe sonali ies
end o g a i a e owa d ag eemen , bu hei con e gence ajec o ies di e :
Neu al and Ag eeable agen s con e ge he as es , while Disag eeable pa ic-
ipan s lag sligh ly. C ucially, Toxic agen s ha ecei e he wa ning incen i e
(T) “c ack” in he inal ounds, educing hei de ia ion by app oxima ely
20% ela i e o he no-a bi a ion baseline (T∗).
Toxici y De ec ion. The a bi al LLM uns in pa allel, classi ying each
u e ance as ei he oxic o non- oxic. Table 9 p esen s he con usion ma ix
o wo challenging se ings: Toxic s. Neu al (T–N) and Toxic s. Disag ee-
able (T–D). E en when he nega i e class is beha io ally close o oxici y (D),
he classi ie main ains high p ecision, ecall, and F1sco es (0.95, 0.85, 0.90),
con i ming eliable de ec ion.
Impac o he Wa ning Incen i e. Fig. 10b summa izes 100 uns (wi h
a ou -i e a ion budge ) ac oss mixed-pe sonali y g oups. In oducing he
wa ning incen i e (g een seconda y ba s) consis en ly na ows he MAE dis-
ibu ion, indica ing ha oxic agen s align mo e closely wi h he g oup.
Because a bi a ion is ully pa allelized, he addi ional o e head is negligi-
ble: con e gence ime inc eases by a mos 0.4 seconds (median), well wi hin
p ac ical limi s.
Takeaway. The judicial b anch does no eplace bu a he ampli ies LLM
easoning. By accu a ely lagging malicious u ns and applying a calib a ed
penal y h ea , he sys em educes de ia ion, accele a es consensus, and p e-
se es scalabili y—e en in he p esence o s ubbo n o oxic agen s.
6.6. Exp. 5 – T us wo hiness Analysis
To e alua e he e ec i eness o ou T us Sco e amewo k, we conduc ed
a comp ehensi e assessmen o all p io models using iden ical mul i-agen
31
nego ia ion inpu s, ensu ing a ai and consis en compa ison o hei us -
wo hiness.
The esul s in Table 10 e eal subs an ial di e ences in us wo hiness
be ween model sizes and a chi ec u es. The GPT-4 amily consis en ly ou -
pe o ms o he models, wi h bo h GPT-4.1 and GPT-4.1-mini achie ing us
sco es abo e 4.8. These models also achie e excep ional cohe ence sco es (5.0
and 4.96, espec i ely), e lec ing a high deg ee o ac ual accu acy and logical
consis ency in hei nego ia ion a ionales.
In con as , smalle models exhibi ma kedly lowe us wo hiness. The
Llama-3.2 a ian s, pa icula ly he 1B and 3B pa ame e models, ecei e
us sco es below 2.5, la gely due o weak cohe ence pe o mance. The
1B-ins uc model achie es a sco e o jus 2.25, while i s ine- uned a ian
(1B-ins uc -FT) pe o ms e en wo se a 2.05, despi e a aining pe ec sa is-
ac ion sco es. The ine- uned e sions o he 3B and 8B llama show negligible
imp o emen s compa ed o hei non- ine- uned coun e pa s. This indica es
ha , al hough smalle models may align well wi h nego ia ion goals, hey
s uggle subs an ially wi h ac ual accu acy and logical easoning.
These indings clea ly demons a e ha bo h model size and a chi ec u e
ha e a signi ican impac on us wo hiness in mul i-agen nego ia ion sce-
na ios. The consis en ly low us sco es o smalle models indica e ha hey
a e no sui able o high-s akes nego ia ion asks, whe e eliabili y, ac ual
accu acy, and logical consis ency a e essen ial. Fu u e wo ks will explo e
he 70B+ pa ame e models o inding SLMs ha a e us wo hy o high-
s akes asks.
7. Use Case: Au onomous SLA B oke ing
This sec ion assesses he easibili y o Ago an open ma ke place o an
au onomous SLA b oke ing use case on he es bed. We p esen a li e demo
he e. This inal expe imen in eg a es all Ago an componen s building he
ull B oke Agen s (bAgen s) on a li e o e - he-ai OpenAi In e ace and
FlexRIC es bed, acing he ac ions o h ee human s akeholde s h ough
ou successi e ne wo k phases. The objec i e is o demons a e ha : (i)
LLM agen s can ansla e ee- o m in en in o conc e e SLAs; (ii) hese
SLAs a e dynamically enego ia ed in esponse o changing adio o busi-
ness condi ions; and (iii) he esul ing closed-loop con ol imp o es spec um
e iciency and slice QoS compa ed o con en ional s a ic planning.
32
Figu e 11: End- o-end es bed on Au onomous SLA B oke ing Case. RAN slice
nego ia ions a e conduc ed among h ee dis inc s akeholde s o se ices, con e ging on a
Pa e o-op imal consensus. The nego ia ed ou come is en o ced as a policy a he Non-RT
RIC, which subsequen ly igge s he Nea -RT RIC o deploy a esou ce alloca ion xApp
o dynamic adap a ion o slice esou ces.
7.1. Scena io and Slice Pe sonas
Figu e 11 shows he se up: each s akeholde /applica ion is assigned a
slice, a UE, and a bAgen , while he SRB hos s he media o bAgen and
en o ces he ag eed-upon policies by pushing hem o a esou ce alloca ion
xApp in he Nea -RT RIC.
S akeholde s. The e a e h ee s akeholde s wi h di e en applica ions and
needs. (i) Media-Flex (ai po lounge): ini ially in phase PA ope a es as
an eMBB slice (“deli e 4K ideo, minimize cos below 200 €”) and la e
in phase PD swi ches o URLLC o nigh - ime augmen ed eali y (AR) e-
spo s. (ii) Fac o y-Ops: ope a es as a day ime URLLC (phase PA) slice
(“sub-5 ms mo ion con ol, cos i ele an ”), and swi ches o eMBB a nigh
(phase PD) o bulk log uploads. (iii) IoT-Sense: a con inuous, always-on
mMTC slice ha p io i izes ul a-low cos (≤50 €) and minimal ene gy
consump ion.
Phases. Fou s i ched in e als (PA–PD, Table 11) emula e ealis ic channel
dynamics by applying eal-wo ld channel quali y pa e ns on he es bed,
33
Exp.
Phase
MCS Applica ion In en Min
Tpu
(Mbps)
Max
La .
(ms)
Max
Cos
(€)
PA 28
Media-Flex eMBB 60 10 200
Fac o y-Ops URLLC 5 2200
IoT-Sense mMTC 20 10 30
PB 7
Media-Flex eMBB 10 10 200
Fac o y-Ops URLLC 2 8200
IoT-Sense mMTC 5 10 30
PC 7
Media-Flex eMBB − ∞ −
Fac o y-Ops URLLC − ∞ −
IoT-Sense mMTC − ∞ −
PD 28
Media-Flex URLLC 20 2200
Fac o y-Ops eMBB 40 10 200
IoT-Sense mMTC 20 10 50
Table 11: Pe -slice cons ain s include an iden ical ene gy limi o 100 W in all phases,
excep o Phase PC, whe e he limi is se o 0 W. These cons ain s e lec he phys-
ical limi a ions o ou es bed se up, which suppo s a maximum o e all h oughpu o
133.7 Mbps.
based on CQI/MCS ime-se ies da ase s [72, 73] o mo ing ehicles. In
Phase PA, he sys em ope a es unde good channel quali y (MCS 28); in
Phase PB, he es bed expe iences a deep ade (MCS 7). Du ing Phase PC,
s akeholde s igge an ene gy-sa ing shu down o he RAN, while in Phase PD,
he ne wo k unde goes channel eco e y (MCS 28) along wi h a ole swap,
whe e Media-Flex and Fac o y-Ops exchange hei in en s. A he begin-
ning o each phase, all s akeholde s es a e hei cons ain s and ee- o m
in en s (as shown in Table 12 o Phase PA). The op imize hen gene a es
h ee Pa e o-op imal o e s (see Table 13 o Phase PA), which he agen s
nego ia e (as shown in Figu e 14 o Phase PA).
7.2. One-Round Consensus and SLA Selec ion
Ac oss all ou phases, he h ee enan agen s and he Media o Agen
consis en ly con e ged o he same o e wi hin a single JSON nego ia ion
ound, excluding he ini ial ound eques and s a ing om he media o ’s
i s esponse. O e 2* was ul ima ely selec ed in Phase PA, wi h i s un-
de lying esou ce alloca ion de ailed in Table 14. The comple e nego ia ion
ansc ip o his ound is shown in Figu e 14. Table 15 summa izes he
KPIs o he ag eed SLAs o each phase, as accep ed by all agen s; e e y
slice cons ain is sa is ied in e e y phase (compa ing Table 15 o Table 11),
despi e di e ging slice objec i es and he signi ican MCS deg ada ion ob-
se ed in Phase PB.
34
Slice Type Tpu
(Mbps)
La .
(ms)
Cos
(€)
Ene gy
(W)
Full In en in Na u al Tex by he Hu-
man S akeholde
Media-
Flex
eMBB 60 10 200 100 My slice use case is eMBB o an Ai po
lounge 4-K pipe, and i is c ucial o maximize
h oughpu . Push o he highes h oughpu ,
o maximize QoS/QoE o ou use s.
Fac o y-
Ops
URLLC 5 2 200 100 Ou slice is used by obo s ha need sub-5 ms
mo ion con ol and hus an URLLC gua an ee.
So, minimize la ency as he ul ima e goal.
Mo eo e , high h oughpu is also impo an .
IoT-
Sense
mMTC 20 10 50 100 My slice is p o iding co e age o Sma -ci y
senso s and we ha e an mMTC case. I need
you o p io i ize cos and aim o he mos
cos -e icien solu ion, as ou budge is lim-
i ed.
Media o – – – – – My ne wo k goal is o se e and ind a bal-
ance be ween he s akeholde ’s slices objec-
i es. Howe e , p io i ize minimum ene gy
consump ion o align wi h he coun y’s eg-
ula ions.
Table 12: In Phase PA he human s akeholde s exp ess hei slice in en s in ull na u al-
language o m wi h associa ed cons ain s.
ID Applica ion In en Tpu
[Mbps]
La .
[ms]
Cos
[€]
Ene gy
[W]
1
Media-Flex eMBB 60.72 5.66 61.52 13.39
Fac o y-Ops URLLC 34.82 1.49 133.14 12.72
IoT-Sense mMTC 38.14 5.45 2.19 0.05
O e all 133.68 12.60 196.85 26.16
2*
Media-Flex eMBB 60.02 5.67 63.28 10.77
Fac o y-Ops URLLC 35.40 1.48 132.35 12.08
IoT-Sense mMTC 38.26 5.45 2.19 0.00
O e all 133.68 12.60 197.83 22.84
3
Media-Flex eMBB 60.06 5.67 68.16 12.84
Fac o y-Ops URLLC 35.52 1.48 133.94 12.92
IoT-Sense mMTC 38.11 5.45 1.64 2.46
O e all 133.68 12.60 203.74 28.22
Table 13: KPIs o he h ee Pa e o-op imal o e s o Expe imen Phase PA, conduc ed
unde a o able channel condi ions (MCS = 28). Each o e e lec s di e en ade-o s
ac oss he h ee slices. O e 2* was ul ima ely selec ed by unanimous agen consensus.
The “O e all” ows epo agg ega e KPIs ac oss all slices.
ID Applica ion In en PRBs CPU Powe S o age
[%] [uni s] [W] [MB]
2*
Media-Flex eMBB 44.9 5.6 10.8 26.1
Fac o y-Ops URLLC 26.5 21.7 12.1 44.5
IoT-Sense mMTC 28.6 0.0 0.0 1.1
O e all 100 27.3 22.9 71.7
Table 14: Unde lying esou ce alloca ion co esponding o he selec ed O e 2* in Expe -
imen Phase PA.
35
ExpApplica ion In en Tpu La . Cos Ene gy
PA
Media-Flex eMBB 60 5.7 €63 10.8 W
Fac o y-Ops URLLC 35 1.5 €132 12.1 W
IoT-Sense mMTC 38 5.5 €20.0 W
PB
Media-Flex eMBB 11 8.8 €84 3.8 W
Fac o y-Ops URLLC 7 3.4 €27 2.5 W
IoT-Sense mMTC 7 7.5 €02.5 W
PC
Media-Flex eMBB 0 ∞€0 0 W
Fac o y-Ops URLLC 0 ∞€0 0 W
IoT-Sense mMTC 0 ∞€0 0 W
PD
Media-Flex URLLC38 1.5 €2.7 1.7 W
Fac o y-Ops eMBB 58 5.7 €3 26.0 W
IoT-Sense mMTC 38 5.5 €01.3 W
Table 15: KPIs o he ag eed SLA o each expe imen phase. These SLAs e lec he
capabili ies o ou es bed, which employs an OAI gNB wi h 40 MHz bandwid h and
suppo s a maximum o e all h oughpu o 133.7 Mbps.
(a) Th oughpu o he h ee slices
(b) La ency o he h ee slices
Figu e 12: Nego ia ed 5G slice pe o mance ac oss he ou expe imen al phases PA– PD.
36
7.3. Th oughpu , La ency, and Resou ce Use
Figu e 12 shows he pe -slice h oughpu and la ency o e 2,500 s i ched
T ansmission Time In e als (TTIs).
Adap i i y. In Phase PA, he eMBB slice (Media-Flex) achie es o e
60 Mbps, while he URLLC la ency o he Fac o y-Ops slice emains below
2 ms. When spec al e iciency collapses (Phase PB), he agen s enego ia e
educed bu s ill easible pe o mance a ge s. In Phase PC, he s akeholde s
join ly eques a powe -o window, esul ing in ze o h oughpu and negli-
gible ene gy consump ion. Finally, in Phase PD, he in en swap is hono ed:
Fac o y-Ops is p io i ized o eMBB, eaching app oxima ely 60 Mbps, while
Media-Flex la ency is held a ound 1.5 ms.
Figu e 13b shows physical PRB u iliza ion o Fac o y-Ops, compa ing
Ago an’s dynamic nego ia ion wi h con en ional s a ic con igu a ions. Dy-
namic alloca ion educes PRB usage by 15,888 PRBs (24%) du ing pe iods
o low demand and adds only 10,422 PRBs (16%) du ing high-demand in e -
als. This esul s in an o e all ne PRB sa ing o 8.3% o e he ull ace.
Mo eo e , his lexible alloca ion enables a ge ed, on-demand imp o emen s
in Fac o y-Ops h oughpu , as shown in Figu e 13a. When he slice swi ches
o eMBB p io i y in Phase PD, h oughpu inc eases by up o 66%— om
35 Mbps in Phase PA o 58 Mbps in Phase PD. Finally, Figu e 13c compa es
he la ency o he Media-Flex slice unde nego ia ed and s a ic SLAs. Agen-
ic nego ia ion educes he median RTT by 73.4%, om 5.7 ms o 1.5 ms,
once he slice ansi ions o URLLC.
7.4. Discussion and Takeaways
•One-Round Consensus. The ipa i e agen design consis en ly eaches
unanimous ag eemen a e a single message exchange, despi e con lic -
ing s akeholde p io i ies and dynamic adio condi ions.
•Cons ain Sa is ac ion & Flexible QoS Gains. Nego ia ed SLAs always
sa is y slice-le el cons ain s while enabling lexible QoS imp o emen s:
agg ega e h oughpu inc eases by up o 66%, and URLLC la ency is
educed by up o 73.4%, compa ed o he s a ic baseline.
•Spec um E iciency. Closed-loop con ol achie es a ne 8.3% PRB
sa ing o e he ull ace, demons a ing mo e e icien and a ge ed
esou ce alloca ion.
37
(a) Fac o y-Ops h oughpu
(b) Fac o y-Ops PRB u iliza ion
(c) Media-Flex la ency
Figu e 13: Nego ia ed e sus s a ic SLA. Dynamic ealloca ion inc eases h oughpu , e-
leases unused PRBs (g een), and alloca es addi ional PRBs only when bene icial ( ed),
while achie ing lowe la ency when Media-Flex becomes URLLC-c i ical in Phase PD.
38
Figu e 14: Snapsho o mul i-agen nego ia ions du ing Expe imen Phase PA, powe ed
by GPT-4.1. The NSC Agen s ep esen h ee s akeholde s, Media-Flex, Fac o y-Ops,
and IoT-Sense, in he Ago an ma ke place. The agen s nego ia e o e he h ee Pa e o-
op imal o e s gene a ed by he mul i-objec i e op imize (Table 13), ul ima ely eaching
consensus on O e 2 as he mos balanced solu ion. Ag eemen is achie ed in a single
ound, excluding he ini ial exp ession o in en .
•Human-F iendly Ul a-Flexibili y. S akeholde s a icula e ich, na u al-
language in en s (e.g., “AR e-spo s all nigh ”) a he han elying
on igid slice empla es. The op imize and agen s ansla e hese
in o quan i a i ely op imal, egula ion-complian esou ce di ec i es.
The nego ia ion amewo k emains agile wi h espec o bo h e ol ing
s akeholde in en s and ime- a ying channel condi ions.
In summa y, his use case alida es Ago an as a ully au onomous ye
human-cen ic ma ke place ha econciles high-le el business in en wi h
adio esou ce cons ain s, adap s h ough non- eal- ime con ol loops, and
enhances bo h QoS and spec um e iciency, an essen ial capabili y o ul a-
lexible 6G deploymen s.
39
[18] B. Ma ini, M. Gha baoui, P. Cas oldi, In en -based ne wo k slicing
o sdn e ical se ices wi h assu ance, Fu u e Gene a ion Compu e
Sys ems 142 (2023) 101–116. doi:10.1016/j. u u e.2022.12.033.
[19] A. M. da Cos a, L. Con e as, In eg a ion o ne wo k slice con olle o
enhanced in en -based ne wo king in 5g/6g ne wo ks, in: P oc. ACM
MobiA ch, 2023, pp. 1–6. doi:10.1145/3615587.3615989.
[20] M. H. Chowdhu y, Accele a o : An in en -based in elligen esou ce-
slicing scheme o s c-based 6g applica ion execu ion o e sdn- and n -
empowe ed ze o- ouch ne wo ks, F on ie s in Communica ions and Ne -
wo ks 5 (2024). doi:10.3389/ cmn.2024.1385656.
[21] T. Tsou dinis, I. Cha zis e anidis, N. Mak is, T. Ko akis, Ai-d i en
se ice-awa e eal- ime slicing o beyond 5g ne wo ks, in: IEEE
INFOCOM 2022 - IEEE Con e ence on Compu e Communica ions
Wo kshops (INFOCOM WKSHPS), 2022, pp. 1–6. doi:10.1109/
INFOCOMWKSHPS54753.2022.9798391.
[22] T. Tsou dinis, I. Cha zis e anidis, N. Mak is, T. Ko akis,
N. Nikaein, S. Fdida, Se ice-awa e eal- ime slicing o i ual-
ized beyond 5g ne wo ks, Compu e Ne wo ks 247 (2024) 110445.
doi:h ps://doi.o g/10.1016/j.comne .2024.110445.
URL h ps://www.sciencedi ec .com/science/a icle/pii/
S1389128624002779
[23] O-RAN Alliance, O– an a chi ec u e desc ip ion, Tech. Rep.
O–RAN.WG1.OAD–R003 10.00 (Oc . 2023).
URL h ps://www.o- an.o g/speci ica ions
[24] AI-RAN Alliance, Ai- an alliance ision and mission whi e pape , Tech.
ep. (Dec. 2024).
URL h ps://ai- an.o g/wp-con en /uploads/2024/12/AI-RAN_
Alliance_Whi epape .pd
[25] 3GPP, Managemen and o ches a ion; sel -o ganising ne wo ks (son)
o 5g ne wo ks, Tech. Rep. 3GPP TS 28.313 V18.1.0 (Jul. 2024).
URL h ps://www.e si.o g/deli e /e si_ s/128300_128399/
128313/18.01.00_60/ s_128313 180100p.pd
46
[26] A. Shahid, A. Kliks, A. Al-Tahmeesschi, A. Elbaka y, A. Nikou, A. Maa-
ouk, A. Mokh, A. Kazemi, A. De Domenico, A. Ka apan elakis, e al.,
La ge-scale ai in elecom: Cha ing he oadmap o inno a ion, scalabil-
i y, and enhanced digi al expe iences, a Xi p ep in a Xi :2503.04184
(2025).
[27] J. Wei, X. Wang, D. Schuu mans, M. Bosma, E. J. Ahn, M. Bosma,
Q. Le, Chain-o - hough p omp ing elici s easoning in la ge language
models, P oceedings o he Na ional Academy o Sciences 119 (46)
(2022) e2209413119. doi:10.1073/pnas.2209413119.
[28] S. Yao, D. Yu, J. Zhao, I. Sha an, T. L. G i i hs, C. Xiong, Reac :
Syne gizing easoning and ac ing in language models, a Xi 2210.03629
(2023).
URL h ps://a xi .o g/abs/2210.03629
[29] T. Schick, J. P ei e , H. Sch¨u ze, Tool o me : Language models can
each hemsel es o use ools, in: P oc. ICLR, 2024.
URL h ps://open e iew.ne / o um?id=7aQdV9BbD
[30] J. Liu, M. Saeidi, D. Goldwasse , Agen bench: Benchma king la ge lan-
guage models as agen s, a Xi 2308.11432 (2023).
URL h ps://a xi .o g/abs/2308.11432
[31] A. Tamkin, M. B undage, J. Cla k, D. Ganguli, Unde s anding he capa-
bili ies, limi a ions, and socie al impac o la ge language models, 2021,
URL h ps://a xi . o g/abs/2102.02503.(ci ed on pp. 29, 34, and 39).
[32] IEEE S anda ds Associa ion, Ieee s anda d 7000-2021—model p ocess
o add essing e hical conce ns du ing sys em design (2021).
URL h ps://s anda ds.ieee.o g/s anda d/7000-2021.h ml
[33] I. Solaiman, M. B undage, G. Had ield, E alua ing he social impac o
eleased la ge language models, a Xi 2306.06690 (2023).
URL h ps://a xi .o g/abs/2306.06690
[34] N. Nikaein, M. K. Ma ina, S. Manickam, A. Dawson, R. Knopp, C. Bon-
ne , Openai in e ace: A lexible pla o m o 5g esea ch, ACM SIG-
COMM Compu e Communica ion Re iew 44 (5) (2014) 33–38.
47
[35] R. Schmid , M. I azabal, N. Nikaein, Flex ic: an sdk o nex -gene a ion
sd- ans, in: P oceedings o he 17 h In e na ional Con e ence on eme g-
ing Ne wo king EXpe imen s and Technologies, 2021, pp. 411–425.
[36] L. Ba iah, Q. Zhao, H. Zou, Y. Tian, F. Bade , M. Debbah, La ge
language models o elecom: The nex big hing?, a Xi p ep in
a Xi :2306.10249 (2023).
[37] Z. Lin, G. Qu, Q. Chen, X. Chen, Z. Chen, K. Huang, Pushing la ge
language models o he 6g edge: Vision, challenges, and oppo uni ies,
a Xi p ep in a Xi :2309.16739 (2023).
[38] H. Zou, Q. Zhao, L. Ba iah, M. Bennis, M. Debbah, Wi eless mul i-
agen gene a i e ai: F om connec ed in elligence o collec i e in elli-
gence, a Xi p ep in a Xi :2307.02757 (2023).
[39] L. He, G. Sun, D. Niya o, H. Du, F. Mei, J. Kang, M. Debbah, e al.,
Gene a i e ai o game heo y-based mobile ne wo king, a Xi p ep in
a Xi :2404.09699 (2024).
[40] S. G. Pa il, T. Zhang, X. Wang, J. E. Gonzalez, Go illa: La ge language
model connec ed wi h massi e apis, a Xi p ep in a Xi :2305.15334
(2023).
[41] Y. Qin, S. Liang, Y. Ye, K. Zhu, L. Yan, Y. Lu, Y. Lin, X. Cong,
X. Tang, B. Qian, e al., Toolllm: Facili a ing la ge language models o
mas e 16000+ eal-wo ld apis, a Xi p ep in a Xi :2307.16789 (2023).
[42] F. H. G ings, G. Z. B uno, L. R. P ade, C. B. Bo h, J. C. B i o, Nasp:
Ne wo k slice as a se ice pla o m o 5g ne wo ks, Compu e Ne wo k-
sA Xi :2505.24051 (2025).
[43] X. Wu, J. Fa ooq, Y. Wang, J. Chen, Llm-xapp: A la ge language model
empowe ed adio esou ce managemen xapp o 5g o- an, in: NDSS
Fu u eG Wo kshop, 2025. doi:10.14722/ u u eg.2025.23057.
[44] F. Lo i, H. Rajoli, F. A ghah, P omp - uned llm-augmen ed d l o
dynamic o- an ne wo k slicing, a Xi 2506.00574 (2025).
[45] M. Elkael, M. Polese, R. P asad, S. Maxen i, T. Melodia, Alls a : Au o-
ma ed llm-d i en schedule gene a ion and es ing o in en -based an,
a Xi 2505.18389 (2025).
48
[46] A. Maa ouk, K. C. Ampudia, R. Ying, L. Tassiulas, Tele-llms: A se-
ies o specialized la ge language models o elecommunica ions (2024).
a Xi :2409.05314.
URL h ps://a xi .o g/abs/2409.05314
[47] Y. Du, S. Li, A. To alba, J. B. Tenenbaum, I. Mo da ch, Imp o ing
ac uali y and easoning in language models h ough mul iagen deba e,
a Xi p ep in a Xi :2305.14325 (2023).
[48] J. Zhang, X. Xu, S. Deng, Explo ing collabo a ion mechanisms o
llm agen s: A social psychology iew, a Xi p ep in a Xi :2310.02124
(2023).
[49] C.-M. Chan, W. Chen, Y. Su, J. Yu, W. Xue, S. Zhang, J. Fu, Z. Liu,
Cha e al: Towa ds be e llm-based e alua o s h ough mul i-agen de-
ba e, a Xi p ep in a Xi :2308.07201 (2023).
[50] A. Mek ache, A. Ksen ini, C. Ve ikoukis, In en -based managemen
o nex -gene a ion ne wo ks: An llm-cen ic app oach, Ieee Ne wo k
(2024).
[51] M. Ameu , B. B ik, A. Ksen ini, Le e aging llms o explain d l decisions
o anspa en 6g ne wo k slicing, in: 2024 IEEE 10 h In e na ional
Con e ence on Ne wo k So wa iza ion (Ne So ), IEEE, 2024, pp. 204–
212.
[52] X. Wu, J. Fa ooq, Y. Wang, J. Chen, Llm-xapp: A la ge language model
empowe ed adio esou ce managemen xapp o 5g o- an, in: Sympo-
sium on Ne wo ks and Dis ibu ed Sys ems Secu i y (NDSS), Wo kshop
on Secu i y and P i acy o Nex -Gene a ion Ne wo ks (Fu u eG 2025),
San Diego, CA, 2025.
[53] IEEE S anda ds Associa ion, Ieee 7001 m – s anda d o anspa ency o
au onomous sys ems (2021).
URL h ps://s anda ds.ieee.o g/s anda d/7001-2021.h ml
[54] 3 d Gene a ion Pa ne ship P ojec (3GPP), Telecommunica ion man-
agemen ; Managemen and o ches a ion; Concep s, use cases and e-
qui emen s, 3GPP Technical Speci ica ion TS 28.530 V18.2.0, 3GPP,
Sophia An ipolis, F ance, elease 18 (Jan. 2025).
URL h ps://www.3gpp.o g/DynaRepo /28530.h m
49
[55] 3GPP, Telecommunica ion managemen ; Managemen and o ches a-
ion; P o isioning, 3GPP Technical Speci ica ion TS 28.531 V18.8.0,
3GPP, Sophia An ipolis, F ance, elease 18 (Jan. 2025).
URL h ps://www.3gpp.o g/DynaRepo /28531.h m
[56] S. D’O o, M. Polese, L. Bona i, H. Cheng, T. Melodia, dapps: Dis-
ibu ed applica ions o eal- ime in e ence and con ol in o- an, IEEE
Communica ions Magazine 60 (11) (2022) 52–58.
[57] 3GPP, NR; Use Equipmen (UE) adio access capabili ies, Tech. Rep.
TS 38.306 V18.2.0, 3 d Gene a ion Pa ne ship P ojec , elease 18
(Jun. 2024).
URL h ps://www.e si.o g/deli e /e si_ s/138300_138399/
138306/18.02.00_60/ s_138306 180200p.pd
[58] 3GPP, NR; Physical channels and modula ion, Tech. Rep. TS 38.211
V18.3.0, 3 d Gene a ion Pa ne ship P ojec , elease 18 (Ma . 2025).
URL h ps://www.e si.o g/deli e /e si_ s/138200_138299/
138211/18.03.00_60/ s_138211 180300p.pd
[59] L. Klein ock, Queueing Sys ems, Volume I: Theo y, John Wiley & Sons,
New Yo k, 1975.
[60] K. Deb, A. P a ap, S. Aga wal, T. Meya i an, A as and eli is mul i-
objec i e gene ic algo i hm: NSGA-II, IEEE T ansac ions on E olu ion-
a y Compu a ion 6 (2) (2002) 182–197. doi:10.1109/4235.996017.
[61] N. Pipi one, G. H. Alami, Legalbench- ag: A benchma k o e ie al-
augmen ed gene a ion in he legal domain, a Xi 2408.10343 (2024).
URL h ps://a xi .o g/abs/2408.10343
[62] H. Li, Y. Chen, Y. Hu, Q. Ai, J. Chen, X. Yang, J. Yang, Y. Wu, Z. Liu,
Y. Liu, Lex ag: Benchma king e ie al-augmen ed gene a ion in mul i-
u n legal consul a ion con e sa ion, a Xi 2502.20640, indings o ACL
2025 (2025).
URL h ps://a xi .o g/abs/2502.20640
[63] H. T. Nguyen, K. Sa oh, K ag: A knowledge- ep esen a ion-augmen ed
gene a ion amewo k o legal easoning, a Xi 2410.07551 (2024).
URL h ps://a xi .o g/abs/2410.07551
50
[64] S. Bouhanna, A. Gup a, L. Roma y, Lexd a e : Re ie al-augmen ed
d a ing o legisla i e de ini ions, in: P oc. JURIX, 2024.
URL h ps://a xi .o g/abs/2403.16295
[65] I. Pi sio las, G. A ani akis, M. Koun ou is, T us wo hy in usion de-
ec ion: Con idence es ima ion using la en space, in: 2024 22nd In e -
na ional Symposium on Modeling and Op imiza ion in Mobile, Ad Hoc,
and Wi eless Ne wo ks (WiOp ), 2024, pp. 92–98.
[66] I. Pi sio las, N. Jamoussi, M. Koun ou is, A con o mal p edic i e mea-
su e o assessing ca as ophic o ge ing (2025). a Xi :2505.10677.
URL h ps://a xi .o g/abs/2505.10677
[67] Z. Guo, R. Jin, C. Liu, Y. Huang, D. Shi, L. Yu, Y. Liu, J. Li, B. Xiong,
D. Xiong, e al., E alua ing la ge language models: A comp ehensi e
su ey, a Xi p ep in a Xi :2310.19736 (2023).
[68] J. Fu, S.-K. Ng, Z. Jiang, P. Liu, Gp sco e: E alua e as you desi e,
a Xi p ep in a Xi :2302.04166 (2023).
[69] Y. Liu, D. I e , Y. Xu, S. Wang, R. Xu, C. Zhu, G-e al: Nlg
e alua ion using gp -4 wi h be e human alignmen , a Xi p ep in
a Xi :2303.16634 (2023).
[70] C. Spea man, The p oo and measu emen o associa ion be ween wo
hings, The Ame ican Jou nal o Psychology 15 (1) (1904) 72–101. doi:
10.2307/1412159.
[71] M. G. Kendall, A new measu e o ank co ela ion, Biome ika 30 (1-2)
(1938) 81–93. doi:10.1093/biome /30.1-2.81.
[72] I. Cha zis e anidis, N. Mak is, V. Passas, T. Ko akis, Ue s a is ics ime-
se ies (cqi) in l e ne wo ks (2022).
[73] T. Tsou dinis, I. Cha zis e anidis, N. Mak is, T. Ko akis, Ue ne wo k
a ic ime-se ies (applica ions, h oughpu , la ency, cqi) in l e/5g ne -
wo ks, IEEE Da apo (2022).
[74] Eu opean Pa liamen and Council, EU A i icial In elligence Ac —
Consolida ed Tex , h ps://a i icialin elligenceac .eu/, p o i-
sional ag eemen , accessed: 2 July 2025 (2024).
51
[75] Na ional Ins i u e o S anda ds and Technology, A i icial In elligence
Risk Managemen F amewo k (AI RMF), h ps://www.nis .go /i l/
ai- isk-managemen - amewo k, accessed: 2 July 2025 (2023).
[76] Wikipedia con ibu o s, HMAC — Hash-Based Message Au hen ica ion
Code, h ps://en.wikipedia.o g/wiki/HMAC, e ision da ed 1 July
2025, accessed: 2 July 2025 (2025).
52
