Discovering 108 tricks to accelerate grokking | Ziming Liu
kindxiaoming.github.io · 561 words · saved by 2 readers
A simple, whitespace theme for academics. Based on [*folio](https://github.com/bogoli/-folio) design.
Author: Ziming Liu (刘子鸣) Over the past year, AI research agents have made significant progress. They can read papers, modify code, design experiments, and even propose new algorithmic improvements. But these systems share a common problem: they lack research intuition. When faced with a new idea, human researchers often form a preliminary judgment — this direction might work, that one probably won’t; something here looks like a phase transition, something there might just be noise. Most AI agents today do not have this ability. For a new training trick, they typically have to actually…
saved by
related reading
- Learning Beyond Gradientstrinkle23897.github.io
- How to ground your ideas? | Ziming Liukindxiaoming.github.io
- Do Machine Learning Models Memorize or Generalize?pair.withgoogle.com
- Trending Papers - Hugging Facepaperswithcode.com
- Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetsmathai-iclr.github.io
- Explore | alphaXivalphaxiv.org
- [2201.02177] Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetsarxiv.org
- Towards Understanding Grokking: An Effective Theory of Representation Learningarxiv.org
- MAI-Thinking-1: Building a Hill-Climbing Machinemicrosoft.ai
- [2604.21691] There Will Be a Scientific Theory of Deep Learningarxiv.org
- [2510.13786] The Art of Scaling Reinforcement Learning Compute for LLMsarxiv.org
- [2602.05910] Chunky Post-Training: Data Driven Failures of Generalizationarxiv.org