<?xml version='1.0' encoding='utf-8'?>
<rss version="2.0"><channel><title>Dhruvi Paprunia writing</title><link>https://dhruvipaprunia.com/blog/</link><description>Paper breakdowns and research writing by Dhruvi Paprunia</description><language>en</language><item><title>A million tokens can still lose the right document</title><link>https://dhruvipaprunia.com/blog/a-million-tokens-can-still-lose-the-right-document/</link><guid>https://dhruvipaprunia.com/blog/a-million-tokens-can-still-lose-the-right-document/</guid><description>Why seeing every document is not the same as finding the right one.</description></item><item><title>Your model found the right answer. It might never learn to find it again.</title><link>https://dhruvipaprunia.com/blog/a-rare-correct-answer-can-still-get-lost-in-training/</link><guid>https://dhruvipaprunia.com/blog/a-rare-correct-answer-can-still-get-lost-in-training/</guid><description>Why GRPO, training that compares rewards across attempts, can help one try more than many.</description></item><item><title>Can one reward teach an assistant everything?</title><link>https://dhruvipaprunia.com/blog/can-one-reward-teach-an-assistant-everything/</link><guid>https://dhruvipaprunia.com/blog/can-one-reward-teach-an-assistant-everything/</guid><description>A reading of MAHALO, where accuracy, human values and tutoring pull an assistant in different directions.</description></item><item><title>A coin flip made this model better at math. Or did it?</title><link>https://dhruvipaprunia.com/blog/coin-flip-math/</link><guid>https://dhruvipaprunia.com/blog/coin-flip-math/</guid><description>Notes on </description></item><item><title>The polite sentence that broke my model</title><link>https://dhruvipaprunia.com/blog/grpo-tool-use/</link><guid>https://dhruvipaprunia.com/blog/grpo-tool-use/</guid><description>How we taught small language models to call tools with GRPO, and why the strictest rule worked best</description></item><item><title>Never Give Up: why easy problems eat the RL budget</title><link>https://dhruvipaprunia.com/blog/never-give-up/</link><guid>https://dhruvipaprunia.com/blog/never-give-up/</guid><description>A deep dive into the Matthew Effect, and how adaptive sampling gives hard problems another chance.</description></item><item><title>The one percent we stopped looking at</title><link>https://dhruvipaprunia.com/blog/one-percent-we-stopped-looking-at/</link><guid>https://dhruvipaprunia.com/blog/one-percent-we-stopped-looking-at/</guid><description>Muon, Never Give Up, and the cases an average can hide.</description></item><item><title>The agent did everything right. The website still said no.</title><link>https://dhruvipaprunia.com/blog/the-agent-did-everything-right-the-website-still-said-no/</link><guid>https://dhruvipaprunia.com/blog/the-agent-did-everything-right-the-website-still-said-no/</guid><description>What a verifier can learn when the assistant took the right steps but the website failed.</description></item><item><title>The model weights are open. Can we see how the model learned?</title><link>https://dhruvipaprunia.com/blog/the-model-weights-are-open-can-we-see-how-the-model-learned/</link><guid>https://dhruvipaprunia.com/blog/the-model-weights-are-open-can-we-see-how-the-model-learned/</guid><description>Why open model training matters, and how to test the steps that shape an assistant.</description></item></channel></rss>