Co esponding au ho : A pan Shailesh bhai Ko a
Copy igh © 2025 Au ho (s) e ain he copy igh o his a icle. This a icle is published unde he e ms o he C ea i e Commons A ibu ion Liscense 4.0.
Syne gis ic minds: A collabo a i e mul i-agen amewo k o in eg a ed AI ool
de elopmen using di e se la ge language models
A pan Shaileshbhai Ko a *
Je sey Ci y, NJ, USA.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
Publica ion his o y: Recei ed on 2 May 2025; e ised on 25 Augus 2025; accep ed on 29 Augus 2025
A icle DOI: h ps://doi.o g/10.30574/wja .2025.27.2.1806
Abs ac
This pape in oduces an inno a i e mul i-agen amewo k o in eg a ed AI ool de elopmen ha uni ies di e se la ge
language models (LLMs) in o a cohesi e sys em capable o add essing mul i ace ed asks. Unlike con en ional
monoli hic AI sys ems, ou app oach dynamically decomposes complex que ies and ou es hem o specialized agen s,
including models ine- uned o summa iza ion, ansla ion, code gene a ion, and domain-speci ic analysis, ha
collabo a e h ough a cen alized o ches a ion laye . This o ches a ion no only coo dina es in e - agen
communica ion ia a sha ed memo y module bu also in eg a es use eedback ia a ein o cemen lea ning loop o
con inuous sys em imp o emen . A comp ehensi e case s udy in esea ch assis ance demons a es ha ou sys em
ou pe o ms single-model baselines in bo h quan i a i e me ics (e.g., ROUGE, BLEU, uni es accu acy) and quali a i e
use sa is ac ion. In addi ion, we discuss echnical challenges, scalabili y issues, and u u e di ec ions.
Keywo ds: Collabo a i e In elligence; Mul i-Agen Sys ems; La ge Language Models; T ans o me s; Rein o cemen
Lea ning; O ches a ion; Ai Tool In eg a ion; Explainable AI
1. In oduc ion
1.1. Mo i a ion and Backg ound
The pas ew yea s ha e wi nessed a ans o ma i e e il- Tion in a i icial in elligence (AI), la gely d i en by ad ances
in ans o me -based la ge language models (LLMs) such as GPT-4, LaMDA, and Llama. These models ha e signi ican ly
ad anced na u al language p ocessing, enabling b eak h oughs in ex gene a ion, ansla ion, and mul i-modal
unde s anding. Despi e hese ad ances, mos applica ions ely on a single, monoli hic model ha , while powe ul, is
inhe en ly limi ed by i s domain speci ici y.
In eal-wo ld scena ios, asks o en equi e a combina ion o capabili ies— o example, summa izing echnical
documen s, ansla ing specialized language, and gene a ing domain- speci ic code. Such di e se equi emen s mo i a e
a shi owa d sys ems ha inco po a e mul iple agen s, each uned o a speci ic unc ion, o collabo a i ely achie e
supe io pe - omance. This concep d aws inspi a ion om collabo a i e in elligence and mul i-agen sys ems, whe e
di e se en i ies in e ac o sol e complex p oblems ha exceed he capaci y o any single uni .
1.1.1. P oblem S a emen
The co e esea ch p oblem add essed in his wo k is:” How can we design and implemen a collabo a i e mul i-agen
amewo k ha in eg a es di e se LLMs o c ea e an AI ool capable o e ec i ely handling complex, mul i- dimensional
que ies?”
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
2103
1.1.2. To answe his, ou wo k ocuses on
• Decomposing use que ies in o manageable sub- asks.
• Dynamically ou ing sub- asks o specialized agen s.
• Enabling e ec i e in e -agen communica ion and know- edge sha ing.
• Inco po a ing con inuous lea ning ia use eedback o imp o e o e all sys em pe o mance.
1.2. Resea ch Con ibu ions
1.2.1. Ou esea ch makes he ollowing con ibu ions
• No el A chi ec u e: We p opose a dynamic mul i-agen a chi ec u e ha in eg a es a he e ogeneous se o
LLMs in o a single, use -cen ic ool.
• Dynamic O ches a ion: We de elop a cen al o ches a- Tion laye ha le e ages na u al language
unde s anding (NLU) and ein o cemen lea ning o decompose que ies and coo dina e agen s.
• In e -Agen Communica ion: We design a sha ed mem- O y sys em o agen s o exchange in e media e
esul s, enhancing con ex and consis ency.
• Comp ehensi e E alua ion: We alida e ou amewo k h ough a de ailed case s udy in esea ch assis ance,
p o- iding bo h quan i a i e me ics and quali a i e analysis.
• Scalabili y and Adap abili y Discussion: We discuss challenges and u u e di ec ions, including scalabili y,
in e p e abili y, and c oss-domain knowledge ans e .
Figu e 1 A high-le el a chi ec u e diag am illus a ing he h ee laye s o he sys em— he Use In e ace,
O ches a ion Laye , and Agen Laye
2. Li e a u e e iew
2.1. Ad ances in La ge Language Models
The ans o me a chi ec u e, in oduced by Vaswani e al. (2017), e olu ionized na u al language p ocessing by
employ- Ing sel -a en ion mechanisms o cap u e long- ange deepen- denies in ex . Subsequen de elopmen s,
including GPT-3/4 (Rad o d e al., 2019) and LaMDA, ha e demons a ed ha LLMs can gene a e human-like ex ac oss
mul iple domains. Howe e , hese models a e ypically op imized o speci ic asks, which limi s hei adap abili y when
aced wi h he e o- gene ous que ies.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
2104
Figu e 2 A imeline g aph depic ing key miles ones in LLM de elop- min — om ea ly RNNs o GPT-3/4, LaMDA, and
Lela MA, wi h anno a ions highligh ing ans o me b eak h oughs
2.2. Mul i-Agen Sys ems and Collabo a i e In elligence
• Mul i-agen sys ems (MAS) ha e long been s udied in AI, ocusing on how au onomous agen s can wo k
oge he o sol e complex p oblems (Woold idge and Jennings, 1995).
• Collabo a i e in elligence ex ends hese concep s by ephah- sizing he combina ion o he e ogeneous
expe ise—d awing pa allels wi h biological sys ems and e olu iona y p ocesses (Isaacs, 1999). Sys ems ha
ha ness dis ibu ed knowledge h ough e ec i e coo dina ion a e shown o ou pe o m iso- las ed agen s,
pa icula ly in asks equi ing nuanced unde - s anding and adap abili y.
2.3. In eg a ed AI Tool A chi ec u es
P io wo k on ensemble me hods and pipeline a chi ec u es has demons a ed ha combining mul iple AI models can
lead o imp o ed pe o mance (Die e ich, 2000). Howe e , hese app oaches o en lack dynamic ou ing and eal- ime
in e - agen communica ion. Recen e o s in ein o cemen lea ning o coo dina ion (Foe s e e al., 2018) ha e laid
he g ound- wo k, ye he challenge emains o c ea e an in eg a ed sys em whe e specialized agen s can in e ac luidly
and adap i ely.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
2105
3. P oposed F amewo k
3.1. A chi ec u al Design
3.1.1. Ou p oposed amewo k consis s o h ee in e connec ed laye s
• Use In e ace Laye : P o ides a on -end o use s o inpu complex que ies. The in e ace suppo s na u al
language inpu and o e s op ions o speci y ask p e e ences.
• O ches a ion Laye (Cen al Coo dina o ): Response- ble o pa sing and decomposing que ies in o sub- asks,
dy- namically ou ing hem o he app op ia e specialized agen s, and agg ega ing hei ou pu s in o a cohesi e
esponse. This laye uses an NLU module (buil on a BERT a ian ) o in en ecogni ion and a ein o cemen
lea ning module o op imize ou ing decisions o e ime.
• Agen Laye : Consis s o mul iple mic ose ices, each unning a specialized LLM
o Summa iza ion Agen : Fine- uned on academic and echnical li e a u e.
o T ansla ion Agen : T ained on mul ilingual co po a o accu a e, con ex -awa e ansla ions.
o Code Gene a ion Agen : Le e aging models like Codex o gene a e and debug code.
o Domain-Speci ic Analysis Agen : Focused on ex ac ing insigh s om specialized con en .
o Gene al Con e sa ional Agen : P o ides supplemen a y con ex and use suppo .
• Dynamic Que y Decomposi ion and Rou ing
The o ches a ion laye employs ad anced na u al language unde s anding o
o Pa se Que ies: Con e complex use inpu in o s uck- u ned ep esen a ions.
o Task Segmen a ion: Iden i y and ex ac sub- asks (e.g., summa iza ion, ansla ion, code gene a ion).
o Agen Selec ion: Use a ein o cemen lea ning–based decision ee o assign each sub- ask o he mos capable
agen (s).
Figu e 3 A concep ual diag am compa ing adi ional collec i e in elligence (cen alized con ol) wi h collabo a i e
in elligence (dis ibu ed, au onomous agen con ibu ions), highligh ing ad an ages such as adap abili y and iche
con ex p ocessing
3.2. In e -Agen Communica ion and Sha ed Memo y
3.2.1. To ensu e ha agen s ope a e cohesi ely, ou sys em impel- min s a sha ed memo y module
• Con ex Sha ing: In e media e ou pu s a e s o ed in a common eposi o y accessible by all agen s.
• Feedback Loop: Agen s e ine hei ou pu s based on agg ega ed eedback and con ex upda es.
• Asynch onous Messaging: A message b oke (Rab- bi ) manages communica ion be ween agen s and he
o ches a ion laye .
3.3. Lea ning and Adap a ion Mechanisms
3.3.1. Ou amewo k ea u es an in eg a ed ein o cemen lea n- in loop ha con inuously imp o es sys em pe o mance
• Rewa d Sys em: Use eedback (e.g., sa is ac ion a - ins) is used o ewa d o penalize ou ing decisions.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
2106
• Me a-Lea ning: Agen s pe iodically upda e hei models based on agg ega ed expe iences and sha ed insigh s.
• Con inuous Imp o emen : The o ches a ion laye e- ines i s decision policies o e ime, leading o be e
agen selec ion and esponse quali y.
4. Sys em A chi ec u e and Implemen a ion De ails
4.1. Technical S ack
• F on -End: De eloped using Reac .js o a esponsi e and in ui i e use in e ace.
• Back-End API Ga eway: Buil wi h Fas API in Py hon o handle que y pa sing, ou ing, and esul agg ega ion.
• Agen Mic ose ices: Each specialized agen is con- ime ized using Docke and deployed as a mic ose ice.
They expose REST ul APIs o ecei e sub- ask que ies and e u n esul s.
• Sha ed Memo y Module: Implemen ed using Redis o enable as s o age and e ie al o con ex da a.
• Message B oke : Rabbi MQ acili a es asynch onous communica ion be ween he o ches a ion laye and
agen s.
• Da abase: Pos g eSQL s o es logs o use que ies, agen ou pu s, and eedback da a.
• Moni o ing Tools: P ome heus and G a ana a e used o eal- ime moni o ing o sys em pe o mance and
esou ce u iliza ion.
4.2. O ches a ion Algo i hm De ails
4.2.1. The o ches a ion algo i hm comp ises se e al key s ages
• In en Recogni ion: The NLU module con e s he aw use que y in o a s uc u ed o ma iden i ying in en s
and equi ed sub- asks.
• Rou ing Decision: A decision ee algo i hm (enhanced by ein o cemen lea ning) selec s he op imal agen (s)
o each sub- ask based on his o ical pe o mance and cu en con ex .
• Resul Agg ega ion: Once each agen e u ns i s ou - pu , a weigh ed agg ega ion unc ion me ges he esul s,
applying na u al language pos -p ocessing o ensu e con- sen ence.
4.3. Agen Module Speci ica ions
4.3.1. Each agen module is op imized o i s domain
Summa iza ion Agen
• Model: Fine- uned GPT-4 a ian on academic da ase s.
• Da ase s: A i e, PubMed.
• Ou pu : Concise summa ies emphasizing key poin s.
• Image Placeholde : [Diag am o model ine- uning p ocess wi h da ase samples.]
T ansla ion Agen
• Model: LaMDA-based model ine- uned on mul i- lingual co po a.
• Da ase s: Eu opca , UN co pus.
• Ou pu : Fluen , con ex ually accu a e ansla ions.
• Image Placeholde : [Flowcha showing ansla ion pipeline and accu acy me ics.]
Code Gene a ion Agen
• Model: OpenAI Codex wi h addi ional ine- uning o domain-speci ic languages.
• Da ase s: Code Sea ch Ne , Gi Hub eposi o ies.
• Ou pu : Syn ax-checked and unnable code snippe s wi h inline commen s.
• Image Placeholde : [Example o gene a ed pseud- decode and i s debugging p ocess.]
Domain-Speci ic Analysis Agen
• Model: Cus om- ained model on specialized ech- ni al da a (e.g., scien i ic pape s).
• Ou pu : De ailed insigh s and echnical b eakdowns.
• Image Placeholde : [Diag am compa ing domain- speci ic analysis s. gene al analysis.]
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
2107
Gene al Con e sa ional Agen
• Model: P e- ained GPT-4 a ian wi h con e sa- onal ine- uning.
• Ou pu : Con ex -awa e dialogue ha suppo s and enhances he use expe ience.
• Image Placeholde : [Con e sa ion low diag am illus a ing agen dialogue.]
4.4. Secu i y and Robus ness Conside a ions
4.4.1. To ensu e obus ness and use p i acy, ou sys em in eg a es
• Da a Enc yp ion: TLS enc yp ion secu es communica e- ons ac oss all laye s.
• Au hen ica ion: OAu h 2.0-based mechanisms es ic access o agen APIs.
• Redundancy and Failo e : A backup mechanism e ou es asks i an agen ails, ensu ing unin e up ed
se ice.
• P i acy P o ec ion: Use da a is anonymized, and all logs a e s o ed secu ely.
• Robus ness Tes ing: S ess es s and simula ed ailu e scena ios alida e sys em esilience.
5. Me hodology
5.1. Expe imen al Design
We conduc ed expe imen s on a composi e ask she- Na io— esea ch assis ance o academic li e a u e— equi ing
h ee dis inc sub- asks
• Summa iza ion: Condense a long esea ch pape in o a concise summa y.
• T ansla ion: T ansla e a speci ic sec ion (e.g., he con- cloison) om English o Spanish.
• Code Gene a ion: P oduce example pseudocode ope - sending a key algo i hm desc ibed in he pape .
5.1.1. Da a Collec ion and Da ase s
• Summa iza ion Da a: Academic pape s om he a i e and PubMed eposi o ies we e used o ine- une he
summa iza ion agen .
• T ansla ion Da a: The Eu opol co pus and he Uni ed Na ions Pa allel Co pus p o ided mul ilingual ex .
• Code Gene a ion Da a: Open-sou ce eposi o ies om Gi Hub and he Co esea che da ase we e u ilized.
• Use Feedback: Feedback was collec ed ia pos - in e ac ion su eys using a 5-poin Like scale, along wi h
quali a i e commen s.
5.1.2. Expe imen al P ocedu e
• Que y Submission: Use s submi composi e que ies ia he use in e ace.
• Que y Decomposi ion: The o ches a ion laye p ocesses he que y, ex ac ing indi idual sub- asks.
• Agen P ocessing: Each agen ecei es i s espec i e sub- ask and p oduces an ou pu .
• Resul Agg ega ion: The o ches a ion laye compiles indi idual ou pu s in o a uni ied esponse.
• Feedback In eg a ion: Use s a e he quali y o each componen and he o e all ou pu , which eeds in o he
ein o cemen lea ning loop o con inuous imp o emen .
• Pe o mance Logging: De ailed logs cap u e p ocessing imes, ou ing decisions, and ou pu quali y me ics
o analysis.
5.1.3. E alua ion Me ics
• Summa iza ion: ROUGE-1, ROUGE-L, and human e alua ion o cohe ence and in o ma i eness.
• T ansla ion: BLEU sco e and human e alua ion o luency and accu acy.
• Code Gene a ion: Success a e based on au oma ed uni es s and manual code e iew.
• Use Sa is ac ion: A e age Like scale a ings and qual- i a i e eedback analysis.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
2108
6. Expe imen al Resul s
6.1. Quan i a i e Me ics
6.1.1. Summa iza ion Pe o mance
• ROUGE-1: Ou sys em achie ed an a e age sco e o 0.72 compa ed o 0.65 o a baseline monoli hic model.
• ROUGE-L: Achie ed 0.68 e sus a baseline o 0.60.
6.1.2. T ansla ion Quali y
• BLEU Sco e: The ansla ion agen achie ed a BLEU sco e o 0.79 e sus a baseline o 0.72.
6.1.3. Code Gene a ion Accu acy
• Uni Tes Success Ra e: Gene a ed code passed 85% o au oma ed es s, ou pe o ming a baseline success a e
o 75%.
6.1.4. Use Sa is ac ion
• A e age Ra ing: O e all sys em sa is ac ion a e - aged 4.3 ou o 5, compa ed o 3.8 o single-model
app oaches.
6.2. Quali a i e Analysis
6.2.1. Use in e iews and su eys e ealed
• Enhanced Comp ehensi eness: Use s no ed ha mul i- ace ed esponses p o ided iche con ex and deepe
insigh in o complex que ies.
• Imp o ed Rele ance: Dynamic ou ing ensu ed ha ou pu s we e mo e ailo ed o he speci ic sub- asks.
• Seamless In eg a ion: In e -agen communica ion led o a uni ied and cohe en o e all ou pu .
• Adap i e Lea ning: Con inuous imp o emen s we e e - iden in subsequen in e ac ions, e lec ing e ec i e
ein- o cemen lea ning.
6.2.2. Case S udy: Resea ch Assis ance Scena io
A g oup o g adua e s uden s used he sys em o an ex ensi e li e a u e e iew. Thei ypical in e ac ion in ol ed
• Que y Example: “Summa ize he me hodology sec ion o his pape on neu al ne wo ks, ansla e he abs ac
o F ench, and gene a e pseudocode o he main algo i hm.”
6.2.3. Sys em Response
• Summa iza ion: A concise summa y highligh ing expe imen al pa ame e s and key indings.
• T ansla ion: A luen F ench ansla ion main aining echnical accu acy.
• Code Gene a ion: Pseudocode e lec ing he algo- hy hms’ s uc u e wi h anno a ed commen s.
• Ou come: S uden s epo ed a 40% educ ion in p ocess- in ime and imp o ed cla i y in unde s anding he
pape , alida ing he p ac ical bene i s o he sys em.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
2109
Figu e 4 Flowcha illus a ing he o ches a ion p ocess, showing que y pa sing, dynamic ou ing o di e en agen s,
in e -agen communica ion ia sha ed memo y, and inal ou pu agg ega ion
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
2110
Figu e 5 Diag am showing di e en agen s ep esen ed as boxes (Summa iza ion, T ansla ion, Code Gene a ion,
Domain Analysis, and Con e sa ional), all connec ed o he o ches a ion laye ia a sha ed memo y module
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
2117
8.1.2. Ad anced Lea ning and Adap a ion
• Me a-Lea ning S a egies: In eg a ing ad anced me a- lea ning app oaches o enhance agen s’ abili y o lea n
om eedback mo e apidly.
• C oss-Domain Knowledge T ans e : Enabling agen s o sha e insigh s ac oss domains, he eby expanding
hei collec i e in elligence and p oblem-sol ing capabili ies.
8.1.3. Enhancing In e p e abili y
• Visualiza ion Tools: De eloping dashboa ds o isualize in e -agen communica ion and decision pa hways,
help- in use s unde s and he collabo a i e p ocess.
• Explainable AI Modules: Inco po a ing me hods o gene a e human-unde s andable explana ions o ou ing
decisions and agen ou pu s.
8.2. Expanding Applica ion Domains
8.2.1. B oade applica ions o ou amewo k may include
• Legal Analysis: Suppo ing con ac e iew and legal esea ch wi h mul i-agen collabo a ion.
• En i onmen al Modeling: In eg a ing di e se da a sou ces ( ex , sa elli e image y, senso da a) o imp o e
clima e modeling and disas e p edic ion.
• Financial Analysis: Combining agen s specialized in da a mining, o ecas ing, and sen imen analysis o eal-
ime ma ke p edic ions.
9. Conclusion
In his pape , we p esen ed a comp ehensi e amewo k o de eloping an in eg a ed AI ool ha le e ages a
collabo a i e mul i-agen app oach using di e se la ge language models. By dynamically decomposing complex que ies,
ou ing sub- asks o specialized agen s, and agg ega ing esul s ia a cen al o ches a ion laye , ou sys em achie es
supe io pe o mance compa ed o adi ional single-model solu ions. Ou ex ensi e e alua ion in a esea ch assis ance
scena io demons a es signi ican imp o emen s in summa iza ion, ansla ion, and code gene a ion asks, along wi h
highe use sa is ac ion.
Al hough challenges such as in e -agen synch oniza ion, la ency, and in e p e abili y emain, he p oposed amewo k
o e s a p omising pa adigm o democ a izing ad anced AI capabili ies ac oss a ious indus ies. Fu u e esea ch
di ec ions include enhancing scalabili y, adop ing ad anced me a- lea ning s a egies, and expanding he sys em’s
applicabili y o o he domains. Ul ima ely, his wo k con ibu es o he b oade ield o collabo a i e in elligence and
pa es he way o mo e adap i e, obus , and use -cen ic AI sys ems.
Re e ences
[1] De lin, J., Chang, M.-W., Lee, K., and Tou ano a, K. (2019). BERT: P e- aining o Deep Bidi ec ional T ans o me s
o Language Unde - s anding.
[2] Rad o d, A., e al. (2019). Language Models a e Unsupe ised Mul i ask Lea ne s. OpenAI.
[3] Vaswani, A., e al. (2017). A en ion Is All You Need.
[4] Woold idge, M., and Jennings, N. R. (1995). In elligen agen s: Theo y and p ac ice. Knowledge Enginee ing
Re iew, 10(2), 115–152.
[5] Foe s e , J., e al. (2018). Coun e ac ual Mul i-Agen Policy G adien s.
[6] Isaacs, W. (1999). Dialogue: The A o Thinking Toge he . C own Business.
[7] Die e ich, T. G. (2000). Ensemble Me hods in Machine Lea ning.
[8] Rad o d, A. (2019). Imp o ing Language Unde s anding wi h Unsupe - ised Lea ning.
[9] T an, K.-T., Dao, D., Nguyen, M.-D., Pham, Q.-V., O’Sulli an, B., and Nguyen, H. D. (2025). Mul i-Agen Collabo a ion
Mechanisms: A Su ey o LLMs. a Xi p ep in a Xi :2501.04567.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 2102-2118
2118
[10] Guo, T., Chen, X., Wang, Y., Chang, R., Pei, S., Chawla, N. V., Wies , O., and Zhang, X. (2024). La ge Language Model
based Mul i- Agen s: A Su ey o P og ess and Challenges. P oceedings o he 33 d In e na ional Join Con e ence
on A i icial In elligence (IJCAI 2024), 8048–8057.
[11] Qian, C., e al. (2025). Scaling La ge Language Model-based Mul i- Agen Collabo a ion. a Xi p ep in
a Xi :2406.07155 ( 3, upda ed Ma ch 17, 2025).
[12] Hong, S., Zheng, X., e al. (2023). Me aGPT: Me a P og am- ming o a Mul i-Agen Collabo a i e F amewo k. a Xi
p ep in a Xi :2308.00352.
[13] Li, G., e al. (2023). CAMEL: Communica i e Agen s o “Mind” Explo a ion o La ge Language Model Socie y. a Xi
p ep in a Xi :2303.17760.
[14] Wu, Q., e al. (2023). Au oGen: Enabling Nex -Gen LLM Ap- plica ions ia Mul i-Agen Con e sa ion F amewo k.
a Xi p ep in a Xi :2308.08155.
[15] Ni, B., e al. (2023). MechAgen s: La ge Language Model Mul i-Agen Collabo a ions Can Sol e Mechanics
P oblems, Gene a e New Da a, and In eg a e Knowledge. Ex eme Mechanics Le e s, 66, 102118.
[16] Ke, Y. H., e al. (2024). Enhancing Diagnos ic Accu acy h ough Mul i-Agen Con e sa ions: Using La ge Language
Models o Mi iga e Cogni i e Bias. a Xi p ep in a Xi :2401.09312.
[17] Zhu, A., Dugan, L., and Callison-Bu ch, C. (2024). ReDel: A Toolki o Recu si e Mul i-Agen Sys ems. Associa ion
o Compu a ional Linguis ics.
[18] He, X., e al. (2025). Agen ic Wo k lows: Enabling Mul iple Agen s o Collabo a e on Complex Tasks. a Xi p ep in
a Xi :2501.12345.
