Papers, rebuilt
so you can explore them.

00 — an embedding space, mid-training interactive
0%

A raw embedding space untangling itself as a model trains. Drag the slider.

Down to the index

The index

Every page opens with the same question: what is this method actually doing?

arXiv 2026 01

Recirculation

Michael C. Mozer, Shoaib Ahmed Siddiqui, Danny Sawyer, Sunny Sanyal, Rosanne Liu Michael C. Mozer et al.

A training-free, inference-time modification that leaks a small amount of activation from a deep layer of a pretrained transformer back down to a shallow one at every step, giving the model a genuine recurrent state without touching a single weight.

walkthrough architectural efficacyinference-time techniques
arXiv 2026 02

Unifying Depth and Width Pruning for LLMs via Binary Knapsack Optimization

Palaash Goel, Ayan Sengupta, Akshay Nambi, Tanmoy Chakraborty Palaash Goel et al.

A two-stage dual-axis pruning framework that first solves a 0/1 knapsack optimization objective over coarse-granularity components, and follows it with a width pruning stage to fill the residual compression budget.

walkthrough LLM compressionefficiencystructured pruning

More papers land here as I write them. If there is one you want pulled apart, say so.

Send a suggestion