Parameter Exploration for RLVR via Variational Learning
Explore in parameter space, not just token space: sampling policies from a posterior (3PO) beats GRPO on math and code at near-identical FLOPs.
tap a tile to plant or clear a bush
drag the snack to a new spot
tap the cat to pet it (at own risk)
PhD student · INSAIT & UKP Lab · Sofia
Vatsal Venkatkrishna
Hi! I'm a second-year PhD student at INSAIT and UKP Lab, advised by Prof. Iryna Gurevych. I work on LLM post-training: using reinforcement learning to align and improve language models, lately with verifiable rewards (RLVR). Before that, I did an integrated B.Tech + M.Tech at IIT Kharagpur, with research internships at ANU and UBC along the way.
Outside of work I'm usually at the gym, on the tennis court, or at a board-game café. I also have a major soft spot for Sofia's street cats, though the feeling is rarely mutual (exhibit A: the cat up top).
Papers
Aletheia: What Makes RLVR For Code Verifiers Tick?
A controlled, execution-grounded testbed for training code verifiers with RLVR. The best recipe depends on scale, and cheaper recipes can match the full pipeline.
Beyond Borders: Investigating Cross-Jurisdiction Transfer in Legal Case Summarization
Summarizing legal cases from jurisdictions with no reference summaries, using unlabeled target data and silver summaries.
Multi-step Automated Generation of Parameter Docstrings in Python: An Exploratory Study
A hybrid pipeline of fine-tuned code models and rules for generating structured Python parameter docstrings.
News
- Preprint Parameter Exploration for RLVR via Variational Learning is out on arXiv
- Aletheia: What Makes RLVR For Code Verifiers Tick? accepted at TMLR
- Started my PhD at INSAIT 🇧🇬
- Beyond Borders accepted at NAACL 2024
- Multi-step Automated Generation of Parameter Docstrings in Python accepted at ICSE 2024 (posters)
- Research intern at the Australian National University 🇦🇺
- Research intern at the University of British Columbia 🇨🇦