Di Wu

Github Pages template for academic personal websites, forked from mmistakes/minimal-mistakes

Project maintained by moore3930 Hosted on GitHub Pages — Theme by mattgraham

I pressed the shutter at random and captured this scene.

Hi there! I’m Di Wu 吴迪, a Phd candidate (starting in Sept 2022) under the supervision of Christof Monz at Language Technology Lab, University of Amsterdam.

I appreciate designing models or approaches that are driven by intuition after capturing deep understandings of specific fields, equipped with modern neural architectures, and grounded in real-world scenarios. The research that attracted me most is simple, insightful, and far-reaching creations or findings, such as Word2Vec.

I mainly focus on Machine Translation, having strong interests in tokenization, quality estimation, and the knowledge transfer mechanism. Beyond translation, I am drawn to fundamental problems in NLP/ML that appear counterintuitive or conceptually puzzling.

Here are some problems I get interested in now. If you’re willing to chat about them, leave me a message. I’m always open to collaborations, or any kind of chat.

Selected Papers

Calibrating Translation Decoding with Quality Estimation on LLMs
- We unify quality optimization and estimation in translation, achieving both SOTA translation performance and accurate quality estimation within a single model.
- Di Wu, Yibin Lei, Christof Monz; arXiv25, [PDF]
Two Simple Experiments on Whether Human-Like Reasoning Helps Translation
- We use two simple experiments to question the trend of using CoT for translation.
- Di Wu, Seth Aycock, Christof Monz; EMNLP25, [PDF]
Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book?
- We show no evidence that long-context LLMs can make use of grammatical explanations for translation, challenging the setting of MT from One Book.
- Seth Aycock, David Stap, Di Wu, Christof Monz, Khalil Sima’an; ICLR25, [PDF]
Representational Isomorphism and Alignment of Multilingual Large Language Models
- We show that the representation of LLMs shares a high degree of isomorphism across languages, providing foundations for zero- or few-shot learning for cross-lingual tasks.
- Di Wu, Yibin Lei, Andrew Yates, Christof Monz; EMNLP24 Findings, [PDF]
How Far can 100 Samples Go? Unlocking Zero-Shot Translation with Tiny Multi-Parallel Data
- Pragmatic views on zero-shot MT: (1) Its potential is underestimated, tiny data brings big gains; (2) The off-target issue is overestimated, one single example can resolve it.
- Di Wu, Shaomu Tan, Yan Meng, David Stap, Christof Monz; ACL24 Findings, [PDF]
Beyond Shared Vocabulary: Increasing Representational Word Similarities across Languages for Multilingual Machine Translation
- We reparameterize the embedding table to overcome the barriers of writing scripts across languages, encouraging positive knowledge transfer.
- Di Wu, Christof Monz; EMNLP23, [PDF]

You may know me better from this personal page.