Sara Kangaslahti
I am a third year CS PhD candidate in the ML foundations group at Harvard University advised by David Alvarez-Melis. I am thankful to be supported by an NSF Graduate Research Fellowship. My research focuses on principled data-centric approaches for understanding and leveraging LLM representations. Recently, I have been working on finding ways to compress and connect models across scales and tasks.
Previously, I completed my Bachelor’s in Computer Science at Caltech, where I worked with Anima Anandkumar and R. Michael Alvarez on scalable tensor-based topic modeling methods.
My email is sarakangaslahti (at) g (dot) harvard (dot) edu. Please feel free to reach out to discuss research!
news
| Aug 21, 2026 | Two of my follow-up works to boomerang distillation are now out! Thinking at the Right Size: Amortized Distillation Across Post-Trained LLMs was accepted at EMNLP Findings and we presented Understanding Layer Patching in Model Size Interpolation at the ICML AdaptFM workshop. |
|---|---|
| Apr 30, 2026 | Paper accepted at ICML 2026! Inverse Depth Scaling From Most Layers Being Similar |
| Jan 26, 2026 | Two of my papers were accepted to ICLR 2026: 🪃 Boomerang Distillation Enables Zero-Shot Model Size Interpolation 🪃 and Hidden Breakthroughs in Language Model Training! |