Yingjin Song 宋盈瑾 Department of Information and Computing Sciences Utrecht University Email: y.song5 AT uu.nl

About Me

I’m a Ph.D. student of Natural Language Processing Group at Utrecht University, advised by Prof. Albert Gatt and Dr. Denis Paperno.

I work on Vision-Language Models, Multimodal Learning, Interpretability and Efficient Reasoning. I am now pariticularly passionate about understanding and improving VLMs on visual reasoning (e.g., temporal, spatial) tasks through the lens of mechanistic interpretability. I aim to uncover how these models represent and process multimodal inputs, diagnose their systematic failures, and develop interpretable methods that align model behavior with human reasoning.

I obtained my M.Sc. degree in Computer Science at The University of Melbourne (advised by Dr. Daniel Beck) in 2021 and my B.Sc. degree in Information Technology at The Australian National University in 2019.

🔔 I am actively seeking internship opportunities in 2026/2027, feel free to reach out if you think I am a suitable fit!

Publications

• Who Is Left of Whom? Tracing Spatial Evidence and Role Binding in Relative-Position Reasoning
  Yingjin Song, Denis Paperno, and Albert Gatt.
  Actionable Interpretability Workshop at COLM2026
  TL;DR:We identify complementary contributions of object-location information and query-role representations to spatial reasoning across visual and textual inputs.
  

• Burn After Reading: Do Multimodal Large Language Models Truly Capture Order of Events in Image Sequences?
  Yingjin Song, Yupei Du, Denis Paperno, and Albert Gatt.
  ACL 2025 Findings
  TL;DR:We proposed a multimodal benchmark for multi-event temporal grounding and reasoning in visual story image sequences.
  

• Disentangling the Roles of Representation and Selection in Data Pruning
  Yupei Du, Yingjin Song, Hugh Mee Wong, Daniil Ignatev, Albert Gatt, and Dong Nguyen.
  ACL 2025 Main
  TL;DR:We disentangled and systematically studied the influence of data representation and selection algorithm in data pruning.
  

• Context-aware Visual Storytelling with Visual Prefix Tuning and Contrastive Learning
  Yingjin Song, Denis Paperno, and Albert Gatt.
  INLG 2024 Oral
  TL;DR:We introduced a context‑aware visual storytelling framework that enhances frozen vision encoder and LLM with prefix tuning.
  

• Modeling Emotion Dynamics in Song Lyrics with State Space Models
  Yingjin Song and Daniel Beck.
  TACL 2023 (Oral Presentation at EACL 2023)
  TL;DR:We proposed a method that combines verse-level emotion prediction with a State Space Model and EM algorithm to reconstruct continuous emotion dynamics in song lyrics.
  

Academic Services

• Reviewer: ACL Rolling Review (2024 - 2026), COLING (2024 - 2025), INLG (2024 - 2025), CoNLL (2025), ACM MM (2025)

Teaching

• Transformers: Applications in Language and Communication: 2025-2027

• Methods in Artificial Intelligence Research: 2022 - 2026

• Responsible Data Science: 2024 - 2025

• Computational Thinking and Programming with Python: 2023

© Yingjin Song | Last update: Sep 2026