Qinyuan Wu
qwu [at] mpi-sws [dot] org
Campus E1 5
66125, Saarbruecken, Germany
I am a third-year PhD student at the CS@Max Planck and the Max Planck Institute for Software Systems (MPI-SWS), advised by Krishna Gummadi. I am also fortunate to closely collaborate with and receive guidance from Evimaria Terzi (Boston University), Mariya Toneva (MPI-SWS), and Muhammad Bilal Zafar (Ruhr University Bochum) (Odered by last name alphabet). Before I joined MPI-SWS, I got my bachelor degree in mathematics-physics from University of Electronic Science and Technology of China (UESTC).
I investigate how LLMs internalize and utilize knowledge, aiming to enhance their reliability and interpretability, delving into questions like memorization, latent knowledge estimation and knowledge injection of LLMs.
Beyond this, I am enthusiastic about collaborating on:
- Privacy and security challenges in LLMs – exploring ways to mitigate risks while maintaining model utility.
- The intersection of neuroscience and language models – investigating how insights from the human brain can inform AI research and vice versa.
- Systems for serving LLMs – including Parameter-Efficient Fine-Tuning (PEFT), quantization, and inference optimization methods like KV caching. While not an expert in LLM systems, I find it fascinating to explore how these optimizations influence model behavior.
news
Feb 21, 2025 | Check our new paper to revisit privacy, utility, and efficiency trade-offs when fine-tuning LLMs: ArXiv. |
---|---|
Feb 12, 2025 | Check our new postion paper to argue the importance of episodic memory in long-term LLM agents: ArXiv. |
Oct 24, 2024 | Our paper Towards Reliable Latent Knowledge Estimation in LLMs: Zero-Prompt Many-Shot Based Factual Knowledge Extraction was accepted by WSDM 2025, see you in Hannover! |
Oct 10, 2024 | One benchmark about the episodic memory of LLMs is on Arxiv! You can read it here. |
Jul 27, 2024 | One paper about the memorisation of LLMs is on Arxiv! You can read it here. |
Apr 19, 2024 | One paper about the knowledge estimation of LLMs is on Arxiv! You can read it here. |