AGI-Papers: A Korean-Language Archive of Curated AI Research Reviews

What AGI-Papers Is and Who Maintains It

AGI-Papers is a curated personal archive started by Kiwoong Yeom, a researcher whose LinkedIn profile is linked from the README. The repository's stated goal is tracking papers important to the journey toward AGI (Artificial General Intelligence) in 2026. Each paper entry links to the paper source and includes a Korean-language review that goes beyond the abstract: the reviews assess the paper's significance, identify limitations, and often compare it to related work.

The archive is organized into ten thematic categories, updated continuously as new papers are added. Recent additions in 2026 include work on multi-agent architectures for long-context tasks, dataset distillation through diffusion language models, and efficiency improvements in KV cache compression. The repository also functions as a pre-publication channel: the README states that insights that predate social media posts will sometimes appear here first.

The author maintains a LinkedIn page and the repository provides a point of contact at the LinkedIn profile. There is no issue tracker opened for paper requests or discussion.

Repository Structure: Directories, Pre-README, and Drafts

The repository is organised with a top-level README.md as the main catalogue, supplemented by individual Markdown files for each paper stored in category subdirectories. The directory structure mirrors the ten topic categories: Agents/, Architecture/, Pre_Training/, Post_Training/, Evaluation/, RAG/, On_Device/, Projects/, Trends/, and Drafts/.

Each paper review lives as a numbered Markdown file within its category directory. For example, Agents/132.md contains the full review for paper 132. The README links each entry to its directory file.

Pre-README.md contains the archive of papers reviewed before 2026. Papers reviewed before that year were moved to this separate file to keep the main README focused on current work. Drafts.md lists reviews that are still being written. The PDF/ directory contains lecture slide decks the author has shared publicly: a PDF on Modern Model Architectures Overview and one on Introduction to Post-Training and Beyond.

To access the repository:

git clone https://github.com/gyunggyung/AGI-Papers.git

After cloning, navigate to the README.md for the current year's papers or to Pre-README.md for earlier entries.

The Ten Topic Categories and Recent Coverage

The ten categories cover the full stack of large language model research. Agents covers autonomous agent systems, planning models, and multi-agent frameworks. Architecture covers LLM architecture innovations including Transformer variants, Mamba, and Mixture-of-Experts models. Pre-Training covers training data, scaling laws, and foundation models. Post-Training covers RLHF, DPO, GRPO, and alignment methods. Evaluation covers benchmarks, evaluation methodology, and critiques.

RAG and Knowledge covers retrieval-augmented generation, knowledge graphs, and memory systems. On-Device AI covers local inference, edge computing, and model optimization for constrained hardware. Projects shows the author's own implementations and experiments. Trends and Industry covers AI industry developments and major research news. Recommended Resources links to external study materials.

Recent 2026 entries in the Agents category include reviews of PARSER (a multi-agent architecture for very long contexts using 896K token windows), Frog Nano (a 4B coding agent trained without distillation), and analysis of multi-turn conversation degradation in frontier LLMs. Architecture entries include DeepSeek-V4.1-Flash and 2D-RoPE. The coverage density varies by category: Agents and Architecture receive the most frequent updates.

How to Read and Navigate the Reviews

Each entry in the README follows a consistent format. The paper title appears first in Korean (the reviewer's translated or paraphrased title), followed by a link to the original paper in English. Then comes a one-to-two-sentence Korean summary of the core contribution, formatted as a subtitle.

The full review is in the linked Markdown file within the category directory. These reviews are written in Korean prose and typically run two to ten paragraphs. They describe what the paper does, why the reviewer found it significant, technical details the reviewer noticed, and where the paper fits relative to other work in the same area. The English paper titles and technical terms (model names, benchmark names, algorithm names) are preserved in English within the Korean prose.

For researchers who read Korean fluently, this format provides unusually direct commentary. The reviewer does not hedge: papers are described as boring, interesting, confusing, or significant based on the reviewer's direct reaction.

Navigating by category is the primary workflow. The README anchor links jump to each category's section, and within each category the papers are listed in reverse chronological order of when the review was published.

Lecture Slides and the Pre-Publication Draft System

The PDF/ directory contains two lecture slides that the author has shared from presentations. The first covers Modern Model Architectures Overview (updated March 2026) and the second covers Introduction to Post-Training and Beyond (updated March 2026). These are standalone PDFs that can be downloaded and read independently of the paper reviews.

The Drafts.md file and Drafts/ directory form a draft publication pipeline. Reviews that are still being written appear in Drafts.md before being moved to the main README. This means the repository sometimes contains work in progress that may be incomplete or may change before the final review is posted.

The README's statement that some content will appear here before the author's LinkedIn posts suggests the repository is updated more frequently than the social media presence. For researchers who want the earliest view of the author's analysis on new papers, the repository is the primary channel.

How AGI-Papers Differs from Papers With Code

Papers With Code is an English-language platform that indexes tens of thousands of machine learning papers alongside code implementations, benchmark leaderboards, and structured metadata. It covers the full breadth of ML research with searchable text and automated updates.

AGI-Papers covers a curated subset of papers selected by a single researcher, with Korean-language commentary that reflects personal judgment rather than structured metadata. Papers With Code answers whether a paper has code and how it benchmarks. AGI-Papers answers whether this specific researcher found the paper important, what they noticed that other reviewers might miss, and how it connects to adjacent work they have read.

For a Korean-reading researcher who wants a guide through the current literature on LLM agents and architecture from someone tracking it closely in 2025 and 2026, AGI-Papers provides something Papers With Code does not. For a researcher who needs English-language searchable coverage of any subfield of ML, Papers With Code is the more appropriate resource.

Licence and Maintenance Status

AGI-Papers is licensed under MIT, which permits use, modification, and redistribution with attribution. The repository has no GitHub releases; new content is added directly to the README and the category Markdown files. The last push was on 2026-09-15, confirming the archive is being actively maintained. Paper entries span from 2025 (in the main README) through the current date, with pre-2026 content in Pre-README.md.

The archive has no contribution guidelines and is not structured as a community project. It is one researcher's personal reading record, made public. The MIT licence permits forking and adapting the format, but the commentary is the author's intellectual output and the licence does not grant rights to republish the reviews under a different author's name.

Editorial conclusion

AGI-Papers suits Korean-speaking AI researchers and practitioners who want curated commentary on cutting-edge papers across LLM architecture, multi-agent systems, and on-device AI, written by a single reviewer with a consistent perspective. It is the wrong resource for non-Korean readers who need searchable English-language summaries: the review text is in Korean, and there is no English translation layer. Before bookmarking it, verify that the research areas you care about have recent entries: some categories like Agents and Architecture are updated frequently while others like Evaluation may have fewer additions.

Frequently asked questions

Official sources

1. gyunggyung/AGI-Papers on GitHub
2. Issues
3. License: MIT
4. Project website
5. README