Nils Schwager

I am currently designing and executing TWON's large-scale simulations by implementing novel components and upgrading the existing simulation infrastructure. This involves offloading agent logic to external GPUs via vLLM and increasing simulation speed through parallelisation and code improvements. In addition, I am conducting further research on LLMs as digital human twins by benchmarking lexical and semantic alignment with authentic content in two papers that are currently under review.

As the person responsible for developing and benchmarking human twins for the TWON project, I developed a history-conditioned user modeling approach for LLMs via next-action prediction to align agents, freeing them from researcher bias. Furthermore, I demonstrated that training on synthetic samples, selected based on semantic similarity, with preference optimisation, improves behavioural alignment more effectively than training on authentic data alone. The conditioning and evaluation concepts I developed were also used in the "Don't Trust Generative Agents [...]" paper, which I also co-authored.

Investigated LLM agents as synthetic survey participants and developed an evaluation framework to assess validity against human baselines. Engineered prompting strategies for controlled experimental stimulus generation. Developed and evaluated several automated content analysis pipelines using RAG, fine-tuned and prompt-engineered LLMs.

Contributed to multimodal science communication materials and managed digital platforms for conferences.