Yifei Chen | University of Tuebingen

About Me

Currently seeking career opportunities in data-driven product analyst.

From October, I will be a gradute in M.A. Computational Linguistics at the Univeristy of Tuebingen.

I also hold these degrees: MSc in Sociology at University of Oxford (UK), BA in Linguistics and Sociology at the University of Manchester (UK).

I have worked full-time at ByteDance as a marketing specialist at Game Publishing Business Sector, with a focus on partnership developments, growth and performance attribution. Previously also an assistant strategic consultant at Roland Berger and A.T.Kearney, and a TV producer at Mango TV (consider it as the Netflix in China).

Research Interests

LLM-Based Agent Simulations: AI(LLM) NPC, Conversational AI Companions.

Professional Interests

Data Analytic, Performance Marketings: using data to improve user experience and drive product growth.

Hobby

Love music: game and film original background music (e.g. Zelda, Kirby, Ori series)
I pratice Kendo, which is a Japanese martial art.

Projects

Three Detectives in a Room: Investigating Character Consistency in Multi-Agent LLM Dialogue (Ongoing)
Github Access
Abstract As LLMs increasingly simulate fictional characters, their role-playing fluency is impressive but often inconsistent, harming dialogue believability. While prior work focuses on single-agent simulation or architectures, multi-agent character consistency remains understudied—especially from a linguistic perspective. This thesis examines how well LLM agents maintain distinct characters during collaborative tasks, addressing a key gap in understanding multi-agent interactions. By analysing linguistic consistency in collective problem-solving, it explores whether current models can sustain coherent character identities beyond isolated exchanges.

A Gradient Adjust Approach to Machine Unlearning (Accepted. SemEval 2025 and ACL 2025)
Github Access
Abstract LLMs risk leaking sensitive data, making machine unlearning crucial. We propose gradient ascent forgetting with KL-divergence retention for a 1B-parameter model. While effective at unlearning, utility preservation remains challenging. Experiments reveal a key trade-off: stronger forgetting reduces performance. This highlights practical difficulties in machine unlearning, as seen in SemEval 2025’s Task 4.

Experiencing Co-Creativity: The Practice and Evaluation of Writing Reality Show Scripts with ChatGPT-3.5 by Professionals (Coursework. Graded 1.0)
Github Access
Abstract This project is a first attempt on the application of large language models (ChatGPT-3.5) into co-creating variety show scripts, practising and evaluating by professionals. Currently, the development of language models has brought profound effects on the content creation industry but the realm of variety show production does not seem to be explored. Based on quantitative questionnaires and qualitative interviews, this project highlights the advantages and disadvantages of ChatGPT-3.5 in casting and script writing, envisioning a fine-tuned co-writing system in the future.

Emotion Recognition of Pedagogical Agents in Education (Coursework. Graded 1.0. Waiting for publication.)
Github Access
Abstract This study explores how a virtual instructor’s emotions (happy, content, frustrated, sad) affect learning in educational videos. Participants (N=300) accurately recognized the agent’s emotions, with happy rated highest for happiness, sad for sadness, etc. Surprisingly, learners exposed to frustrated/sad agents performed better than those with happy/content agents, though positive emotions were rated as more engaging. Findings suggest emotional expressions influence learning differently than perceived engagement, highlighting their complex role in education.

The Relations Between Personality and Language Use (Coursework. Graded 1.0)
Github Access
Abstract The study extract data from all series of Harry Potter films to analyse and investigate if language use truly reflect character personalities.