Stackin Paper: Building on ML Research

Welcome to Stackin Paper! This is my personal playground for breaking down the latest machine learning research papers, implementing them from scratch, and bridging the gap between intertwined topics; all while documenting my findings along the way.

I am particularly excited by, and aim to focus on, methodologies for enabling continual learning and long-horizon memory management. I feel very strongly towards these areas as I believe them to be fundamental limitations preventing the progression of LLMs beyond the current paradigm.

Objectives

When tasked with explaining my objectives for this endeavor, I fall back to a quote from one of my favorite shows, Bleach, to offer some insight:

"I dislike perfection. If something is perfect, there is nothing above it. There is no room for imagination. No space for intelligence or ability. [...] Be more wonderful than anything that existed before you, but never perfect."

I have squandered much of my time in the pursuit of perfection; always planning, rarely executing. I became so focused on ensuring that my first attempt was successful, that I never actaully got around to making an attempt in the first place.

My hope with this venture is to change that. To embrace imperfection, and the rich trove of experiences that resides within the pitfalls of mistakes. Over the course of this next chapter, I hope to make many, many, many mistakes. More importantly, I hope to use these mistakes as fuel to light the path for my future explorations !

Consequently, through these future explorations, I aim to build a solid foundation for algorithmic and practical rigor; and gain insight on research taste. I have many opinions on what this idea of 'taste' truly entails, but I will leave that perhaps for a future article :)

Research Logs

Date	Paper / Topic	Key Concepts	Action
Feb 2026	Gradient Regularization w/ DR.GRPO	GRPO	View Project
March 2026	Power Sampling	MCMC	View Project
Upcoming	Writing Context into Memory via Gradient Descent	Meta-Learning	Being Imlemented
Upcoming	Residual Attention	Attention Mechanism	Being Imlemented
Upcoming	Self-Distillation Fine-Tuning vs Experiential RL	RL	Being Imlemented

AI Disclosure

I will try to be as transparent as possible regarding my usage of AI where relevant. While it can certainly be tempting to speed up progress and rely entirely on AI to implement these papers, my pursuit of imperfection leaves me seemingly orthogonal to this predisposition.

How can I learn, make mistakes, and be perfectly imperfect without touching upon the work myself ? The extent of my AI usage is to review & refine my conceptual understanding at the start of the project, and verify key implementation details at the end. I will do my best to point out if/where it's been used for other reasons in the articles.

Well, that and also this entire website was coded using AI. I am not a frontend developer :)