huskydoge/Awesome-Loop-Models

What it solves

Most neural networks process input in one straight pass: layers stack up, and each layer is used exactly once. This project focuses on a different family of models — loop models — where the same learned layer, block, or operator is reused many times, either along depth inside a single forward pass or across time by feeding the model's own output back in. That reuse brings benefits like parameter efficiency, better reasoning, adaptive compute, and easier scaling at test time, but these models are scattered across many papers and hard to track. This project collects, organizes, and structures the research on loop models so people can find relevant work, understand the design choices, and follow the field's progress.

How it works

The repository is a curated, structured list of loop-model papers and blogs, maintained by huuskydoge and contributors. The full catalog lives in an interactive web browser (linked in the README), where you can search, browse by chapter, filter by section and loop design, and read summaries, venues, code links, and daily briefings.

Every entry is generated from a YAML file (papers/*.yaml or blogs/*.yaml) by a Python build script (scripts/build.py). The catalog is organized as a topic map: it has five major parts (Foundations, What Looping Buys, Building/Training/Deploying, Understanding/Applications, Open Problems) split into chapters and sections. Each paper is tagged with controlled values for design axes like loop_topology, sharing, depth_control, and for claims (claims, comparison). A strict scope definition is included: only models that truly reuse a shared component are in scope; agent loops, repeated full-model calls, and simple sequence recurrence are excluded. The README includes a “Start here” list of must-read core references, a “Latest additions” section, a “Most cited” table, and a list of long-form blog posts.

Who it's for

Researchers, engineers, and students working on or interested in looped, recurrent-depth, weight-tied, deep-equilibrium, and similar neural architectures. It's useful for anyone who wants to find papers, understand design trade-offs, track new publications, or get an overview of the field's open problems.

Highlights

• Interactive browser — search, filter, and read TL;DRs, venue info, code links, and daily briefings all in one place.
• Structured topic map — papers are organized into 32 sections across 10 chapters, making it easy to navigate by theme.
• 354 papers and 8 blogs as of the last update, with a daily watch adding new arXiv papers.
• Controlled metadata — each paper records design axes (loop topology, sharing, depth control) and claimed benefits, with no free-form tags.
• Clear scope definition — explicitly excludes agent loops, repeated full-model calls, and ordinary sequence recurrence, so the list stays focused.
• Open contribution — anyone can submit a paper or blog via pull request, using a structured YAML template and a build script.