This year, HKU IDS has been honoured to sponsor the Conference on Parsimony and Learning (CPAL) after serving as its inaugural local host in 2024, providing to our researchers and students a novel, yet cozy arena for addressing the parsimonious, low dimensional structures that prevail in machine learning, signal processing, optimization, and beyond. As the platinum sponsor of CPAL 2025, representatives of our Institute had a prestigious chance of interacting with subject experts in Stanford University, California, at the 4-day conference between March 24 – 27, 2025. The conference attracted over 150 of participants from the academia to share insights on the latest developments in computation learning.
This year, HKU IDS has been honoured to sponsor the Conference on Parsimony and Learning (CPAL) after serving as its inaugural local host in 2024, providing to our researchers and students a novel, yet cozy arena for addressing the parsimonious, low dimensional structures that prevail in machine learning, signal processing, optimization, and beyond.
As the platinum sponsor of CPAL 2025, representatives of our Institute had a prestigious chance of interacting with subject experts in Stanford University, California, at the 4-day conference between March 24 – 27, 2025. The conference attracted over 150 of participants from the academia to share insights on the latest developments in computation learning.
Poster presentation on foundation model post-training featuring Tianzhe: SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training
Authors: Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Sergey Levine, Yi Ma
Poster presentation on foundation model post-training featuring Tianzhe: SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training
Authors: Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Sergey Levine, Yi Ma
Mr. Tianzhe ZHU
First-Year PhD Candidate, HKU IDS
Primary Supervisor:
Director & Professor, Chair of Artificial Intelligence, HKU IDS / CDS
Q
Congratulations on having your paper featured in ECCV 2024! It seems that there were some other activities for you to experience at the conference in addition to the academic parts. Can you tell me more?
A
The cool California breeze welcomed me to CPAL at Stanford. I’m proud to have maintained my perfect attendance record (2/2, after last year’s event at HKU). So excited!
As usual, CPAL offers a cozy environment for researchers to connect. As always, CPAL provided an intimate setting for researchers to connect. Hosted in the Simonyi Center on Stanford’s beautiful campus, the conference featured a thoughtfully structured single-stream program of paper presentations, keynote talks, and engaging panel discussions. Unlike the overwhelming mega machine learning conferences, CPAL allowed participants to attend every session and easily interact with speakers. I treasured the opportunity to learn from greatest theoreticians, like Yuandong Tian, whose insightful analysis of Transformer training dynamics was fascinating. Between sessions, the IDS-sponsored coffee breaks (with excellent refreshments) offered valuable networking time. I especially enjoyed savoring snacks while listening to engaging talks by leading researchers like Jason Lee and Atlas Wang.
Q
How could it benefit yourself as a junior data science researcher by engaging in the conversations with other scholars at an international platform like conferences or symposiums??
A
Attending conferences is the bread and butter of researchers, I would say. “Come to see your pals” is CPAL’s fitting slogan, which perfectly matched my own motivation for attending: reconnecting with colleagues and making new connections, with photography as a pleasant bonus. It was wonderful to reunite with my Bay Area collaborators—Simon, Dhruv, Sam, and Ziyang—and to meet promising new scholars from undergraduate to postdoctoral levels. These diverse conversations about current research interests proved, as always, both enjoyable and inspiring.
Mr. Tianzhe ZHU
First-Year PhD Candidate, HKU IDS
Primary Supervisor:
Director & Professor, Chair of Artificial Intelligence, HKU IDS / CDS
Q
How did you feel about attending CPAL again? I know you were also present in its first occurrence last year when IDS hosted the event!
A
The cool California breeze welcomed me to CPAL at Stanford. I’m proud to have maintained my perfect attendance record (2/2, after last year’s event at HKU). So excited!
As usual, CPAL offers a cozy environment for researchers to connect. As always, CPAL provided an intimate setting for researchers to connect. Hosted in the Simonyi Center on Stanford’s beautiful campus, the conference featured a thoughtfully structured single-stream program of paper presentations, keynote talks, and engaging panel discussions. Unlike the overwhelming mega machine learning conferences, CPAL allowed participants to attend every session and easily interact with speakers. I treasured the opportunity to learn from greatest theoreticians, like Yuandong Tian, whose insightful analysis of Transformer training dynamics was fascinating. Between sessions, the IDS-sponsored coffee breaks (with excellent refreshments) offered valuable networking time. I especially enjoyed savoring snacks while listening to engaging talks by leading researchers like Jason Lee and Atlas Wang.
Q
How could it benefit yourself as a junior data science researcher by engaging in the conversations with other scholars at an international platform like conferences or symposiums??
A
Attending conferences is the bread and butter of researchers, I would say. “Come to see your pals” is CPAL’s fitting slogan, which perfectly matched my own motivation for attending: reconnecting with colleagues and making new connections, with photography as a pleasant bonus. It was wonderful to reunite with my Bay Area collaborators—Simon, Dhruv, Sam, and Ziyang—and to meet promising new scholars from undergraduate to postdoctoral levels. These diverse conversations about current research interests proved, as always, both enjoyable and inspiring.
Q
Many of the attendees at CPAL 2024 shared with us how they enjoyed the conference itself as a close-knit community for foundational machine learning/AI researchers. Was it the same this year? What was the best thing about it – was it your poster presentation with your accepted paper?
A
Attending CPAL functioned as a perfect closed-loop system—sharing ideas through my poster presentation while gathering valuable insights from fellow researchers. I thoroughly enjoyed it a lot to present my works to the audience on the stage. The most valuable takeaway was hearing diverse perspectives on academia’s role compared to industry in this era of large models. Some researchers candidly acknowledged limitations in their work, while others firmly advocated for academia’s crucial contributions to understanding and benchmarking artificial intelligence, showcasing their dedicated efforts in this direction. These thoughtful perspectives have significantly helped shape both my research direction and career planning.
Q
Many of the attendees at CPAL 2024 shared with us how they enjoyed the conference itself as a close-knit community for foundational machine learning/AI researchers. Was it the same this year? What was the best thing about it – was it your poster presentation with your accepted paper?
A
Attending CPAL functioned as a perfect closed-loop system—sharing ideas through my poster presentation while gathering valuable insights from fellow researchers. I thoroughly enjoyed it a lot to present my works to the audience on the stage. The most valuable takeaway was hearing diverse perspectives on academia’s role compared to industry in this era of large models. Some researchers candidly acknowledged limitations in their work, while others firmly advocated for academia’s crucial contributions to understanding and benchmarking artificial intelligence, showcasing their dedicated efforts in this direction. These thoughtful perspectives have significantly helped shape both my research direction and career planning.
