llm-paper-daily: A Daily-Updated Feed of LLM and Agent Research Papers

What llm-paper-daily Contains and Who Uses It

llm-paper-daily is a GitHub repository that tracks new research papers on LLMs and AI agents. Each day's entry appears in the repository README and in a monthly table under `summary/`, listing the paper title, the authoring institution, an arXiv PDF link, a link to a related GitHub repository when one is available, and a Chinese-language summary of the paper's contribution and findings.

The intended audience is Chinese-speaking researchers and engineers who follow LLM and agent research and want a single place to scan new work each morning without manually checking arXiv or conference proceedings. The English-language summaries live under `summary_en/` for readers who need them.

The repository itself does not provide a search interface, filtering by topic, or citation tracking. It is a manually curated or automated feed rather than a research database.

How Each Daily Entry Is Structured

Each entry in the repository follows a consistent format. The README displays the most recent batch, and the full history is organized by month in the `summary/` directory. A representative entry from September 2026 shows the format: paper title in English, institution affiliation, a Chinese summary paragraph discussing the key problem, the proposed method, and the main result, and a badge linking to the arXiv PDF.

The summary text is substantive, typically two to four sentences covering what problem the paper addresses, what technique it introduces, and what the result demonstrates. For example, the September 15 entry on FlashVector describes the paper's scope (hierarchical model serving stack optimization via an LLM-driven agent framework), the validation context (Unity production environment), and the outcome (significant performance improvement, demonstrating cross-level automated optimization).

When a paper's code is publicly available, the entry includes a badge linking to the associated GitHub repository. Not every paper has an associated repository, and the feed does not track whether code is later released.

Subscribing Locally with the paper-subscribe Skill

The repository includes a skill in the `skill/` directory (`paper-subscribe`) that handles local digest delivery. Instead of requiring the user to write a subscription script manually, the README provides a setup prompt to send to a local AI agent (OpenClaw, Codex, or Claude Code). The agent reads `SUBSCRIBE.md` and configures the local setup, including a configuration file, a preview digest, and a scheduled task. The README shows the following setup prompt to send to your agent:

请帮我配置 llm-paper-daily 的本地订阅。订阅仓库是 https://github.com/xianshang33/llm-paper-daily ，请阅读仓库根目录的 SUBSCRIBE.md，按文档创建本地配置、预览 digest、安装定时任务，并在完成后告诉我配置文件位置、运行时间、语言、每次推送数量和验证结果。

The skill only reads `feed-papers.json` from the public repository. It does not run paper collection or summary generation locally on the subscriber's machine. The `feed-papers.json` file serves as the distribution endpoint; subscribers read from it rather than from the full repository structure.

The `SUBSCRIBE.md` file in the repository root documents the setup parameters. These include the number of papers pushed per digest, the delivery schedule, and the language preference for the output. Checking `SUBSCRIBE.md` directly before setup is necessary because the README does not duplicate all the configuration options.

Repository Organization and Navigation

The top-level structure contains four directories relevant to readers: `summary/` for the monthly paper tables in Chinese, `summary_en/` for English equivalents, `data/` for any supporting data files, and `skill/` for the subscription skill. The `feed-papers.json` file at the root is the machine-readable feed.

The README and `README_en.md` are the primary entry points. The monthly tables use a consistent date-indexed format, so navigating to a specific month or paper requires either reading the table in-order or using GitHub's file search to find a paper by title.

There is no web interface, no search functionality, and no tagging system for filtering by topic such as alignment, efficiency, or multi-agent systems. Readers who want to find all papers on a specific topic from past months need to manually scan the monthly tables or clone the repository and use command-line search tools.

Limitations: Language, Recency, and Coverage Gaps

The primary summaries are written in Chinese. Readers who rely on the `summary_en/` directory for English equivalents should verify that the English entries match the Chinese ones in coverage, as the README does not describe how English translations are generated or maintained.

The feed covers LLM and agent papers. It does not cover adjacent areas such as computer vision models, reinforcement learning without LLM components, or software engineering research that does not involve language models. Coverage within LLM and agent research is not guaranteed to be exhaustive: the README does not describe the paper selection criteria or how papers are identified and added.

The last push was on 2026-09-16. At that point the most recent papers listed were from September 15, 2026, indicating a roughly one-day lag between paper posting and repository update.

Alternative Paper Tracking Tools

Papers With Code at paperswithcode.com is a widely used paper tracking site that links research papers to their code implementations, supports search and filtering by task and method, and covers a broader range of machine learning topics beyond LLMs. Unlike llm-paper-daily, it is a web service rather than a Git repository, so it does not support local digest subscriptions or offline browsing.

HuggingFace Daily Papers provides a curated daily selection of prominent ML papers with brief highlights. It is also a web service with no local subscription option. llm-paper-daily's distinguishing characteristic is that it is a Git repository with a machine-readable JSON feed and a skill for automating local delivery, which makes it more directly usable in agent-based workflows that already consume GitHub data.

Editorial conclusion

llm-paper-daily suits researchers and engineers who want a structured daily feed of LLM and agent papers without building their own paper-tracking pipeline. The summaries are written in Chinese, so readers who need English should use `summary_en/` or translate manually. Before setting up the local subscription, read `SUBSCRIBE.md` and verify that the digest parameters match your preferred daily volume.

Frequently asked questions

Official sources

1. Issues
2. README
3. xianshang33/llm-paper-daily on GitHub