Projects & Writing

← Back to main page
Watercolor illustration for the Dr.GRPO post
Technical Write-up

You Could Have Invented Dr.GRPO Yourself

A ground-up walkthrough of modern RL for LLMs: from REINFORCE to Dr.GRPO, showing how each algorithm was forced by exactly one broken thing in the previous one.

Reinforcement Learning PPO GRPO LLMs
Watercolor illustration for the entropy post
Technical Write-up

You Could Have Invented Entropy Yourself

A ground-up derivation that runs from a number-guessing game to the loss function of a language model: halving, bits, entropy, cross-entropy, perplexity, and KL divergence, each step the obvious next move from the last.

Information Theory Entropy Cross-Entropy KL Divergence LLMs
Watercolor illustration for the SDFT Mixture-of-Experts post
Academic Project

Does SDFT Survive the Jump to Mixture-of-Experts?

A theory-first investigation: derive the bound that decides whether self-distillation's trust-region guarantee survives in a Mixture-of-Experts, then measure directly where a real 16B MoE lands on it.

Mixture-of-Experts SDFT Trust Region LLMs

Browse the archive