Research Overview

Research

I build computational models of language — how it is used, how it encodes social meaning, and increasingly, how well today's LLMs actually capture it. My work sits at the intersection of NLP, pragmatics, and formal linguistics, combining machine learning with deep theoretical grounding in how humans communicate.

Currently

Active Work

Research in progress at ZAS Berlin, at the intersection of LLM evaluation and computational pragmatics — plus the datasets, code, and demos that make it usable.

Published · CMCL 2026
LLM Calibration & Motive-Mediated Social Inference
Two evaluation metrics — ESR (Effect Size Ratio) and CDS (Calibration Deviation Score) — quantify how well frontier LLMs (GPT, Claude, Gemini) predict human social evaluations across politeness, precision, and register tasks. A follow-up study tests the motive-mediation hypothesis: LLM social judgments arise from implicit motive attribution rather than direct output patterns, with cross-model calibration differences tracing to motive-to-judgment mapping quality, not to which motives surface in attribution.
LLM evaluation calibration motive mediation social meaning
→ Published in CMCL 2026 proceedings · follow-up under review at EMNLP · code & data
Live Demo · HF Spaces
PRISMA — a Chatbot That Profiles You Back
A Gradio app (GPT-OSS 120B via Groq) that flips the usual chatbot dynamic: while you talk to Prisma, she forms a running impression of you across six social-perception dimensions — competent, likeable, considerate, polite, formal, demanding — updated turn by turn and inspectable at any point in the conversation. A research-facing demo built on the CMCL 2026 / EMNLP 2026 (under review) work on LLM social perception.
LLM demo social perception Gradio real-time inference
→ try it live · code
Public Dataset · Hugging Face
imprecision-bench: A Multimodal Pragmatics Benchmark
475 human time-reference productions across 12 clock states × 2 pragmatic contexts (police witness statement vs. neighbor party chat), paired with clock images, multi-label motive annotations, and a peer-reviewed RSA baseline (r² ≈ 0.97). Tests whether LLMs and VLMs calibrate linguistic precision to social context — benchmarked on GPT-4o mini, Claude Haiku 4.5, and Gemini 2.5 Flash.
benchmark multimodal VLM evaluation pragmatics
→ dataset on HF · code & notebook
In Progress
Probabilistic Pragmatics: Speaker Models & Strategic Requests
Extending Rational Speech Act (RSA) and game-theoretic frameworks to model how speakers make strategic linguistic choices — (im)precision, politeness, register, and the choice between direct and indirect requests — and how these choices convey social meaning beyond literal content.
speaker modeling RSA indirect speech acts register
→ Part of SFB 1412 Register project at ZAS Berlin · presented at Leibniz MMS Days 2025
Across 15+ Years

Research Themes

Recurring questions that have shaped my work across projects, methods, and institutions.

⚖️
Core Theme
LLM Evaluation & Calibration
Do large language models capture social meaning the way humans do? I develop metrics and benchmarks to test LLM calibration on fine-grained pragmatic tasks — politeness, precision, register — using human judgments as ground truth.
ESR · CDS metrics · CMCL 2026
💬
Core Theme
Pragmatics & Social Meaning
How do speakers use language strategically — choosing words not just for content but for social effect? I model politeness, (im)precision, and register as the outcomes of rational communicative choices with social consequences.
Linguistics & Philosophy 2021 · Linguistics Vanguard 2022 · ELM3 2024
🔁
Core Theme
Language Dynamics & Evolution
How do linguistic conventions emerge, spread, and change across populations? I use agent-based simulations and evolutionary game theory to model language change, grammaticalization, and the evolution of semantic ambiguity.
Synthese 2021 · Morphology 2017 · AI & Society 2019
🕸️
Core Theme
Signaling & Social Networks
How do signaling conventions form and propagate through social networks? I study the interplay of local interaction, network structure, and population-level behavior using multi-agent models and network theory.
Games 2022 · AI & Society 2019 · Ph.D. Thesis 2013
🎮
Core Theme
Game Theory & Formal Modeling
Game-theoretic frameworks — signaling games, evolutionary dynamics, reinforcement learning — provide precise tools for modeling the strategic dimensions of communication and the emergence of linguistic conventions.
Experimental Economics 2022 · Games 2019 · Fundamenta Informaticae 2018
🧪
Core Theme
Experimental & Empirical Methods
Formal models need empirical grounding. I design behavioral experiments (LabVanced, Prolific) and apply statistical modeling (Python, R, mixed-effects) to test theoretical predictions against human data.
ELM 2024 · Social Meaning Berlin 2023 · EvoSAL 2019–2021
Toolkit

Methods & Frameworks

The computational tools and theoretical frameworks underlying this work.

// Computational Methods
›LLM evaluation — calibration metrics, prompt engineering, human-model comparison
›Deep learning — PyTorch, HuggingFace Transformers, fine-tuning
›Agent-based modeling — multi-agent simulations, evolutionary dynamics, NetLogo
›Reinforcement learning — fictitious play, replicator dynamics, multi-agent RL
›Statistical modeling — Bayesian inference, mixed-effects models, Python/R
// Theoretical Frameworks
›Game-theoretic pragmatics — signaling games, RSA, strategic communication
›Evolutionary linguistics — population dynamics, language change, iterated learning
›Formal semantics — truth conditions, scalar implicature, register theory
›Network theory — social networks, convention propagation, small-world models
›Experimental pragmatics — behavioral experiments, LabVanced, Prolific
Project History

Past Projects

Funded research projects and institutional affiliations that shaped the current work.

2020 – present DFG · SFB 1412
Modeling Meaning-Driven Register Variation
As Senior Researcher in the collaborative research centre SFB 1412 "Register", I develop probabilistic models of register variation — investigating how speakers adapt their language choices to social context, and what this tells us about the strategic and social dimensions of meaning.
ZAS Berlin, Germany Project site →
2019 – 2021 NAWA · €75K
EvoSAL: The Evolution of Semantic Ambiguity in the Lab
As PI of this NAWA Ulam-funded project, I developed iterated learning models of semantic ambiguity and tested predictions experimentally. The computational models outperformed established theoretical baselines, with results published in Synthese (2021) and Games (2022).
Nicolaus Copernicus University, Toruń, Poland Project site →
2017 – 2019 EU H2020
ODYCCEUS: Opinion Dynamics & Cultural Conflicts
As postdoctoral researcher in this Horizon 2020 project, I designed reinforcement learning models combining cognitive frames and game-theoretic dynamics to predict cooperation in multi-agent interactions — introducing a new RL paradigm that outperformed existing models across five key metrics.
Università Ca' Foscari Venezia, Italy Project site →
2013 – 2017 ERC Advanced
EVOLAEMP: Language Evolution — The Empirical Turn
As postdoctoral researcher in this ERC Advanced Grant project, I developed computational sociolinguistic models using agent-based simulation and network analysis. This period included visiting research stays at Stanford, UC Irvine, Ohio State, and Toulouse.
University of Tübingen, Germany Related publications →