OS
LLM
KG
Au ho s
OS co po a
(pape s, da a, code)
LLM-assis ed
ex ac ion
(nodes + edges)
Answe wi h OS
sou ce ancho s
Cu a ion-Loop
(Human-in- he-Loop)
KG Upda e
OUR APPROACH
A CONTRADICTION RESISTS RESOLUTION, YET INVITES COMPLEMENTARITY.
2
CORE QUESTION
is ounded on and ex ends he guiding hough s
o RCR o p omo e esponsible conduc o esea ch, sha e eliable da a,
minimise was e o esou ces, and os e inno a ion.
Open Science
In con as , gene al-pu pose op imize o scale ia p obabilis ic
aining on as (non-)scien i ic da a— as o deploy, bu wi h limi ed
p o enance and highe hallucina ion isk.
LLMs
AI-d i en esea ch is shaped by wo
di e en logics:
B idging Open
Science and La ge
Language Models
CORE PROBLEM
Coupling wo logics: Open Science ac s as a p o enance-
awa e KG ga ekeepe ha s ee s LLMs; con e sely, LLMs
li OS co po a in o s uc u ed, e sioned knowledge
g aphs (no base-model ine- uning).
Enhancing Resea ch
Accu acy h ough
Knowledge G aphs
How can he wo di e en
logics—epis emic go e nance
(sou ce i s ) s. p obabilis ic
comp ession a scale (scale
i s )—coexis and be u ilised
esponsibly in esea ch?
Wilde Nicolaus: [email p o ec ed]
Ma ie Ala i: [email p o ec ed]
Nicolaus Wilde
Ma ie Ala i
Julia P iess-Buchhei
USE (Resea che s using LLMs)
PRODUCTION (Resea che s p oducing KGs)
Awa eness
Comple eness
P e en ing
plagia ism
RCR
O iginali y
Explainabili y
Reliabili y
Rep oducibili y
T us in
Science
Opennes
E hics
Hones y
P e en ing
edundancy
Accoun abili y
Validi y
Fai ness
Fac i i y
Responsibili y
Recip oci y
P e en ing ab ica ion
P e en ing
alsi ica ion
T anspa ency
Sha ing
In e p e abili y
Sa e y
T aceabili y
Equi y
Reusabili y
In eg i y
Quali y
Con iden iali y
FAIR-R
Da a con iden iali y
Da a p o ec ion
Consis ecy
Da a
quali y
1
OS egula es he sea ch space (p o enance, FAIR-R, RCR), LLMs ill i wi h gene a i e elas ici y.
OS co po a
KG laye
(on ology + ins ances)
Ve sioning
(Log changes)
Con ex
selec ion
ia he KG
LLM gene a es
answe om
con ex
Bene i s
Risks &
Go e nance
Co e age bias (KG)
License‑awa e
e ie al
G aph/p omp
injec ion de enses
Upda e d i /
e o s ( ollback)
Mul ilingual
equi y
Fas e o ien a ion
wi h e i iable
sou ces
Lowe
hallucina ion ia
p o enance ga ing
Clea e c edi
No base-model
e aining o ine-
uning equi ed
T aceable p ocess
Reusable (wo ks
ac oss domains)
Li e a u e: (1) Ala i, M., Wilde , N., & P iess-Buchhei , J. (2024). P omo ing Open Science in imes o A i icial In elligence: Do we g asp he in e play? Wo ld Con e ence o Resea ch
In eg i y (WCRI), A hens, G eece. Zenodo. h ps://doi.o g/10.5281/zenodo.11562117
(2) Kho ashadizadeh, H., Ama a, F. Z., Ezzabady, M. e al. (2024). Resea ch ends o he in e play be ween la ge language models and knowledge g aphs. 10.48550/a Xi .2406.08223.
(3) Ve huls , S e aan and Zahu anec, And ew and Cha e z, Hannah, Mo ing Towa d he FAIR-R p inciples: Ad ancing AI-Ready Da a (Ma ch 04, 2025). A ailable a SSRN:
h ps://a xi .o g/abs/2405.04333
18
Awa eness
Comple eness
P e en ing
plagia ism
RCR
O iginali y
Explainabili y
Reliabili y
Rep oducibili y
T us in
Science
Openness
E hics
Hones y
P e en ing
edundancy
Accoun abili y
Validi y
Fai ness
Fac i i y
Responsibili y
Recip oci y
P e en ing ab ica ion
P e en ing
alsi ica ion
T anspa ency
Sha ing
In e p e abili y
Sa e y
T aceabili y
Equi y
Reusabili y
In eg i y
Quali y
Con iden iali y
FAIR-R
Da a con iden iali y
Da a p o ec ion
Consis ency
Da a
quali y
