Scholar
Sadhika Malladi
Google Scholar ID: 9HCmTcwAAAAJ
Princeton University
Citations & Impact
Citations
1,529
H-index
15
i10-index
19
Publications
20
Co-authors
9
list available
Contact
No contact links provided.
Publications
Resume (English only)
Academic Achievements
- Named a 2025 Siebel Scholar
- Oral presentation at NeurIPS 2023: Proposed MeZO, a memory-efficient zeroth-order optimizer for LLM fine-tuning, achieving up to 12x memory savings and 50% fewer GPU-hours
- ICLR 2021: Formalized fine-tuning dynamics via kernel theory, explaining the efficacy of parameter-efficient methods like LoRA and motivating MeZO
- ICML 2024: Developed LESS, a data selection algorithm that outperforms full-dataset tuning using only 5% of data
- NeurIPS 2021/2022 and ICLR 2024: Series of works using SDEs to model optimization dynamics of SGD, Momentum, RMSProp, and Adam, enabling efficient large-batch training
- NeurIPS 2024: Showed that preference learning algorithms rarely increase likelihood of preferred responses
- ICLR 2025: Demonstrated that standard preference learning can cause 'unintentional unalignment' in LLMs and proposed a theory-motivated data filtering solution
