LLMAgentPapers: A Structured Reading List for LLM Agent Research

What This Repository Does and Who It Serves

LLMAgentPapers is a curated reading list for researchers and practitioners who want to understand large language model agents. It solves the problem of scattered literature: papers on agents appear across NLP, robotics, and multi-agent systems, and finding them requires chasing citations or keyword searches. The repository groups papers into categories such as Agent, Multiple Agents, Application, and Framework, which gives a newcomer a map of the field. It is not a tutorial or a codebase. It is a bibliography with links to arXiv abstracts. The intended audience is someone who needs a quick orientation before diving into primary sources, not someone looking for implementation guidance.

The Classification Scheme: From Single Agent to Multi-Agent

The repository organizes papers into a hierarchy that reflects the research questions in the field. The top level separates Overview papers from those on the Agent itself, Multiple Agents, Application, and Framework. Under Agent, it splits into Personality, Memory, Planning, Tool use, and RL training. That division is practical: it mirrors the common decomposition of an agent into components. Under Multiple Agents, the list distinguishes Task-Oriented Communication from Casual/Open Conversations, and further splits task-oriented into Collaborative Exchanges and Adversarial Interactions. This taxonomy is not arbitrary. It helps a reader see that multi-agent research is not a single topic. The categories also reveal a bias toward social and behavioral aspects, such as Theory of Mind and role-play, which may not cover all engineering concerns like latency or cost.

How to Use the List: Commands and Navigation

The repository is a standard GitHub project, so getting it running is trivial: clone it with git clone https://github.com/zjunlp/LLMAgentPapers.git, or simply browse the README on the web. There are no configuration keys or dependencies. The README contains a table of contents with anchor links to sections. Each paper entry lists authors, a link to arXiv, and a date. For example, the entry for A Survey on Large Language Model based Autonomous Agents shows the arXiv ID 2308.11432. To read a paper, click the abs link. The contribution section explains how to add papers, but the exact instructions are not shown in the provided material. The practical use is to open the README, scan the categories, and follow links to papers that match your current question.

Coverage Strengths: Surveys and Recent Directions

The Overview section is the strongest part of the list. It includes major surveys from 2023 to 2025, such as the survey by Wang et al. on autonomous agents and the survey on human-agent systems by Zou et al. from May 2025. It also includes a 2026 preprint on harness engineering, which is a less common topic. This coverage gives a reader a broad sense of the field's evolution. The inclusion of a survey on agentic reinforcement learning from September 2025 shows that the list tracks newer subfields. However, the list is not exhaustive. It does not include every important paper, and the selection criteria are not stated. A reader cannot know why a particular paper is included or omitted. That is a limitation for someone who wants a comprehensive review rather than a personal selection.

Maintenance and Currency: A Real Concern

The repository's last push was September 5, 2025, according to the metadata. Yet the README includes a paper dated March 2026. That inconsistency suggests the README was edited at some point, but the last push date may reflect a later commit that did not change content, or the metadata is misleading. Either way, the list is not actively maintained on a regular schedule. The News section has an entry from March 2024 about the authors' own paper, KnowAgent, and an entry from June 2023 about creating the repository. There is no news from 2025 or 2026. For a research field that moves quickly, a six-month gap means missing important work. A reader should check the arXiv dates and search for newer surveys after 2025. The license is MIT, which permits reuse, but the paper links themselves are external and subject to their own copyright.

Limitations and When It Is the Wrong Tool

This list is a curated index, not a living database. It does not include code, benchmarks, or comparisons of agent frameworks. If you need to choose a specific agent library like LangChain or AutoGPT, this repository will not help. It also does not provide any analysis or commentary on the papers. Each entry is just a citation. For a practitioner who wants to evaluate a tool, a paper list is the wrong starting point. The taxonomy can also be misleading. For example, the Personality category includes a paper on Theory of Mind, which is a controversial claim, and a paper on toxicity in persona-assigned models. Those are not the same kind of work. A reader might assume the category is coherent when it is not. The repository is best for literature discovery, not for understanding the technical details of agent implementations.

Alternatives and How They Differ

A direct alternative is the Awesome-LLM-Agents list or similar curated GitHub repositories, but the provided material does not name one. However, a general alternative is a searchable academic database like arXiv's own listing or Semantic Scholar. Those tools let you search by keyword, filter by date, and see citations. The difference is that they are not curated. They return everything, which can be overwhelming. LLMAgentPapers offers a human-curated subset, which saves time if the curator's taste matches yours. Another alternative is the survey papers themselves. The Overview section includes several surveys that synthesize the field. Reading one survey gives you a structured narrative, whereas this list gives you raw citations. If you want context and comparison, choose a survey. If you want a quick list of primary papers, use this repository.

Maintenance and Upgrade Cost

The maintenance cost for a reader is low: there is no software to update or configure. The cost is in the time spent verifying that the papers are still relevant and that the list is current. The repository has not had a recent release, and the last push is from September 2025. The README shows a paper from March 2026, which is confusing. A reader should not assume the list is up to date. The license is MIT, so you can fork the repository and maintain your own version. That is a concrete option if you want to add papers or update the categories. However, forking means you take on the curation burden. The repository does not provide any automated way to track new papers, so you would need to check arXiv manually. The upgrade cost is the cost of doing your own literature search.

Editorial conclusion

Adopt LLMAgentPapers if you are a researcher or engineer who needs a structured entry point into LLM agent literature and can tolerate a list that is not exhaustive. Do not rely on it as a complete or up-to-date source, because the last push was September 2025 and the most recent paper entry is from March 2026, suggesting sporadic maintenance. Before using it, verify that each paper still matches your interest area, cross-check with arXiv for the latest versions, and consider supplementing with a more actively maintained list or a database like Semantic Scholar. The repository is a useful filter, but you must apply your own relevance check to every entry.

Official sources

1. Issues
2. README
3. zjunlp/LLMAgentPapers on GitHub