The question: cooperation between two LLM agents is measurable. What happens in groups, and on networks?

What exists: the dyadic games pipeline, behavioral phenotypes, pre-registered predictions, and first triadic simulations already run

What you would build: n-player and networked games (public goods, coordination, threshold games), varying interaction structure rather than only payoffs; comparison against evolutionary game theory and the human experimental literature

Also possible: run the human online experiment that validates our pre-registered predictions in game regions no human study has covered so far

Skills you would come out with: experimental design, large-scale inference on HPC, network science, evolutionary game theory

Measure: extend the embedding-based recovery of value structure to more instruments, more languages, more model families

Compare: do value representations converge across models, or diverge? Convergence would imply hard limits on cultural differentiation through training data alone

Intervene: locate where value-relevant information is encoded (probing), steer it, and measure what changes downstream in tasks like hate speech classification, incivility detection, or political annotation

Where it sits: the core of DEMOS-AI, with Paula Szewach at BSC; the audit question (“whose values are the reference?”) has a legal deadline attached to it through the EU AI Act

Skills: psychometrics, interpretability, multilingual NLP, research on annotation and benchmark validity

(Especially this project could also involve an in-person visit at the Barcelona Supercomputing Center)

The instrument: models fine-tuned on the traces of a specific population and period, then administered a questionnaire, as a complement to survey data

The open questions are the interesting ones: when do synthetic respondents fail, which groups do they systematically misrepresent, how do we quantify uncertainty, what would a validity theory for this look like?

The applied side: LLM agents grounded in empirical findings through retrieval, used to generate hypotheses and pre-test survey and experiment designs before they cost money

Where it sits: funded through the AI4SOCIALPLUS use case I co-lead, with benchmarks against representative survey programmes

Skills: finetuning and evaluation, survey methodology, validation design, a healthy scepticism about silicon samples

(Especially this project could also involve an in-person visit at the University of Konstanz)

The data: a decade of threaded news forum interactions with explicit positive and negative votes, already published and open

What exists: a pipeline that separates antagonism from alignment in signed networks, instead of collapsing polarization into one number

The new angle: measure the content flowing through those structures with LLM-based instruments, while testing whether those instruments carry political bias of their own (ACL 2025)

Why it matters here: polarization, opinion dynamics and socioeconomic segregation are already central to this department. I would add the measurement layer and the signed structure

Skills: network science, NLP, temporal and causal analysis on observational data

Regular one-on-one, and an open door in between. Generally, fewer longer meetings. I read drafts properly and I do my best to answer fast

First year shape: one self-contained project that gives you a working pipeline, a public repository and a paper draft, before we go after the ambitious thing

Open by default: pre-registrations where we make predictions, code on GitHub with a DOI, data in a repository (Dataverse for example).

I write with you, not for you.

I will tell you honestly when something is not working. All kinds of efforts of mine have been rejected, many times over the years, and you will hear about how that was handled

Networks matter to me: conferences (IC2S2, *ACL and similar), external stays, internships. Andrea Cera Palatsi went from a master’s thesis with me at BSC to a PhD at the CSH Digital Innovation School. If you do good work, I am happy to recommend you or talk to people about you.

You want to combine a social science question with AI systems (very broadly) and care about measurement rather than benchmarks

Useful to have: R, Python, willingness to learn, some statistics

Not required: prior NLP research, deep learning from scratch, a finished idea

Probably not a fit: small application of your pet methodology with no social science question behind it

Do a bit of reading, bring a half-formed idea, that is enough to start