Michael Bereket

CS PhD @ Stanford

[Michael Bereket @mdbereket]

I will also be presenting these results at the LM4Sci workshop @ COLM 2025! Would love to connect if you are also thinking about language model reasoning and scientific discovery

[Michael Bereket @mdbereket]

Can popular RL methods like GRPO improve LLM reasoning with supervision from verifiable yet stochastic outcomes, like noisy scientific experiments?

Introducing "Uncalibrated Reasoning: GRPO Induces Overconfidence with Stochastic Outcomes"! (1/7)

[Michael Bereket @mdbereket]

Read more ↓↓↓
📄 Paper: https://t.co/RTTkd7VDqj
💻 Code: https://t.co/6GM2euniGm
📓 Colab notebook (run synthetic data experiments in 2 minutes in your browser!): https://t.co/IhyjkmLW9P

(7/7)