LLMAgentPapers: A Topic-Organized Reading List for LLM Agent Research

What LLMAgentPapers Contains and Who It Serves

LLMAgentPapers is a repository maintained by the ZJUNLP research group that collects and organizes academic papers on large language model agents. The README describes it as a list of must-read papers. It was created in June 2023 to track multi-agent research, and the README reflects updates through at least March 2026.

The repository serves researchers who are new to the LLM agent field and want a curated starting point, as well as experienced researchers who want a single reference for tracking the major papers across subtopics. The content is a single README.md file with no separate data files or build process. Using it means reading the README on GitHub or cloning the repository and opening the file locally.

The README also links to two related curated lists from the same group: Prompt4ReasoningPapers, covering reasoning with language model prompting, and KnowledgeEditingPapers, covering knowledge editing for large language models.

How the Repository Is Organized

The table of contents reveals a two-level taxonomy. The top-level categories are Overview, Agent, Multiple Agents, Application, Framework, Others, and Resources.

The Agent section is subdivided into five capability categories: Personality, Memory, Planning, Tool Use, and RL training. The Multiple Agents section separates task-oriented communication from casual or open conversations, and within task-oriented communication distinguishes between Collaborative Exchanges and Adversarial Interactions. The Resources section includes Benchmarks, Types of Tools, and a Tool List.

Each paper entry in the list follows a consistent format: title in bold, author list in italics, a link to the abstract on arXiv or another source, and the publication year and month. Where a paper has associated code, a separate code link is included. This format makes it straightforward to assess the recency of a paper and navigate to the original source.

Overview Section: Survey Papers That Define the Field

The Overview section contains ten survey papers that provide broad coverage of the LLM agent landscape. The earliest is a 2023 survey on interactive NLP (arXiv 2305.13246). Two 2023 surveys directly address LLM-based autonomous agents: one from Renmin University covering agent architecture components (arXiv 2308.11432) and one covering agent potential and design principles (arXiv 2309.07864).

From 2024, the survey list includes a paper on code as a medium for LLM agency (arXiv 2401.00812), a survey on multimodal agentic interaction (arXiv 2401.03568), a paper on personal LLM agents covering capability, efficiency, and security (arXiv 2401.05459), and a survey on neural code intelligence (arXiv 2403.14734). A 2025 paper covers LLM-based human-agent systems (arXiv 2505.00753) and another covers agentic reinforcement learning for LLMs (arXiv 2509.02547). A 2026 paper on harness engineering for language agents is also listed (preprints.org/manuscript/202603.1756).

The inclusion of papers from 2026 confirms that the list is being updated beyond the initial 2023 scope.

Single-Agent Capability Categories and the KnowAgent Paper

The Agent section is organized by the capabilities a single agent requires: Personality, Memory, Planning, Tool Use, and RL training.

The Personality category begins with a Theory of Mind paper (arXiv 2302.02083) from 2023 and includes papers on toxicity in persona-assigned language models, repeated game behavior, role-playing, and expert-instructed prompting. These papers address how assigning a persona or character affects model outputs.

The Planning category is where the ZJUNLP group's own work appears. The README news section highlights a 2024 paper, KnowAgent: Knowledge-Augmented Planning for LLM-Based Agents (arXiv 2403.03101), which the group released in March 2024. This is the only paper in the README news section, distinguishing it as a notable contribution from the repository's maintainers.

The truncated README does not show the full content of the Memory, Planning, Tool Use, and RL training categories, so the total paper count within each subcategory is not available here.

Multi-Agent and Application Categories

The Multiple Agents section divides multi-agent research into two interaction patterns. Task-oriented communication covers scenarios where agents coordinate toward a shared goal, with collaborative exchanges grouped separately from adversarial interactions. The adversarial subcategory distinguishes the cooperative from the competitive dynamics in multi-agent LLM setups.

The Casual and Open Conversations subcategory covers research on agents interacting without a fixed task objective, such as social simulation or open-ended dialogue.

The Application and Framework sections are listed in the table of contents but their contents are in the portion of the README not shown. Based on the structure, Application likely covers domain-specific deployments such as code generation agents, web browsing agents, and scientific research agents. Framework likely covers tool or infrastructure papers that describe how to build agent systems, though this cannot be confirmed from the visible content.

Limitations: What a Curated List Cannot Provide

The repository consists of a single README.md file with no search functionality, no tagging system beyond the sections, and no way to filter by publication year, venue, or citation count. Navigating to a specific subfield requires scrolling to the relevant section header.

New papers are added manually, which means the list reflects the attention and bandwidth of the ZJUNLP group rather than a complete index of the field. A paper may be published and widely cited for weeks or months before it appears here. The README does not document the criteria for inclusion beyond describing the list as "must-read."

The licence is listed as unknown in the repository metadata, though the README displays an MIT badge. Teams who want to republish or adapt the list should verify the actual licence by reading any LICENSE file in the repository or contacting the maintainers.

The repository has no GitHub releases and contains only two files: README.md and a .DS_Store file. All value is in the README.

Comparison with Papers With Code: Curated Selection versus Live Database

Papers With Code is a widely used resource that tracks machine learning papers alongside their associated code implementations, benchmarks, and results tables. It covers LLM and agent research extensively and is updated continuously as papers appear on arXiv.

The difference in approach is significant. Papers With Code is a database with search and filtering by task, dataset, method, and benchmark. It indexes tens of thousands of papers automatically and links evaluation numbers to leaderboards. LLMAgentPapers is a manually curated selection of papers deemed essential for understanding the LLM agent field, organized by concept rather than by task or method.

For a researcher who wants to find every paper on a specific benchmark or see which methods have the highest score on a specific task, Papers With Code is more functional. For a researcher who wants a structured reading plan for the LLM agent field and wants to start with the papers the ZJUNLP group considers foundational, LLMAgentPapers provides a more focused and opinionated path.

Editorial conclusion

Researchers and engineers entering the LLM agent field will find LLMAgentPapers a direct starting point for building a reading list across the core subtopics. The list covers overview surveys, single-agent capabilities, multi-agent dynamics, and frameworks. Its limitation is inherent to any manually curated repository: coverage depends on what the maintainers choose to include, and a recent paper may not appear for weeks or months. The last push was on 2026-09-12. Researchers who need a continuously updated index should combine this list with Papers With Code searches on the agent tag.

Frequently asked questions

Official sources

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