Blog

Sporadic posts about machine learning, math, and things I find interesting.

• Mar 27, 2026

  Length-controlled Baselines in Reinforcement Learning with Language Models

  We propose a modification to the GRPO reward baseline that might lead to more stable training of reasoning models, with initial empirical evidence.

  RL
• Mar 26, 2026

  From Decision Theories to Multi-Agent Reinforcement Learning

  Notes on rational decision-making in predictor environments and how it relates to RL.

  AlignmentRL
• Mar 24, 2026

  Newcomb's Paradox, Free Will and Superintellignce

  What could Newcomb's Paradox on near-perfect prediction of human behavior mean for alignment?

  Alignment