LLM-Tool-Survey: A Paper Index for Tool Learning, Not a Tool Library

What the Repository Actually Ships

The name suggests a tool. The contents are a bibliography. The README opens by describing the project as "the collection of papers related to tool learning with LLMs," organized according to the authors' survey paper, "Tool Learning with Large Language Models: A Survey." The primary language field is unknown and no license is stated, which is consistent with a repository whose payload is Markdown and one framework diagram at assets/Framework.png. There is no package manifest, no install target, and no release. The only versioned artifact referenced in the material is the arXiv preprint at arxiv.org/abs/2405.17935 and the journal version in Frontiers in Computer Science, volume 19, number 8, pages 198343, year 2025. If you arrive expecting a library that wraps tool calls for an agent, you will leave with a reading list instead. That is not a defect in the work. It is a mismatch in expectation, and the README does not spend much effort correcting it.

The Four-Stage Taxonomy That Organizes the List

The structure is the substance here. Papers are grouped under Why Tool Learning and How Tool Learning, and the How branch splits into four stages: task planning, tool selection, tool calling, and response generation. Benchmarks and evaluation sit in their own section, followed by challenges and future directions. That four-stage split is a claim about how tool-augmented systems decompose, and it is more useful than a flat list because it tells you where a given paper sits in your own pipeline. If you are debugging why an agent picks the wrong API, the Tool Selection group is the relevant pile. If your agent picks correctly but formats the call badly, look at Tool Calling. The Why branch is further divided by benefit: knowledge acquisition (with search engine, database and knowledge graph, and weather or map subgroups) and expertise enhancement (mathematical tools and others). The README states the survey reviews the why from six specific aspects, though the repository listing itself only enumerates the benefit categories rather than naming all six.

Reading the Entries: What a Listing Contains

Each entry follows a consistent format: a bolded title, a venue and year, and a link. Under Knowledge Acquisition, the search engine group includes Internet-Augmented Dialogue Generation (ACL 2022), WebGPT (Preprint 2021), REPLUG (Preprint 2023), Toolformer (NeurIPS 2023), ART (Preprint 2023), ToolCoder (Preprint 2023), and CRITIC (ICLR 2024). The database and knowledge graph group lists Lamda, Gorilla (NeurIPS 2024), ToolkenGPT (NeurIPS 2023), ToolQA (NeurIPS 2023), a finite-state decoding paper (NeurIPS 2023), and Middleware for LLMs (EMNLP 2024). The weather or map group has three entries. Note the venue labels: several entries are marked Preprint with a year, which means the list does not track whether those preprints later appeared at a venue or were revised. There are no abstracts, no per-paper notes, and no indication of which papers the survey authors consider foundational versus incremental. The list is an index, not an annotated bibliography, and the difference matters when you are triaging forty papers on a deadline.

The Survey Itself and How to Cite It

The repository exists to serve one paper. The README announces that the survey was accepted by Frontiers of Computer Science and that the latest version has been released. The citation block gives a complete BibTeX entry: author list Qu, Changle; Dai, Sunhao; Wei, Xiaochi; Cai, Hengyi; Wang, Shuaiqiang; Yin, Dawei; Xu, Jun; Wen, Ji-Rong, with the journal, volume, number, pages, year, and publisher fields filled in. That is the most operationally useful block in the whole repository. If you are writing a related-work section, you can paste it directly. The README also notes that two Chinese-language writeups, one described as brief and one as comprehensive, were produced by PaperAgent and 旺知识, with links to both. The survey abstract, quoted in the README, frames the motivation as fragmentation: the existing literature "remains fragmented and lacks systematic organization, posing barriers to entry for newcomers." That is the problem statement the repository is meant to answer.

Contribution Path and the Absence of a Licence

The contribution model is the standard one for a curated list. The README states: "Please feel free to open an issue or make a pull request!" There is no CONTRIBUTING file described, no stated inclusion criteria, and no template for what a submission should contain. In practice that means acceptance depends on the maintainers' judgement, and the README gives no way to predict it. More consequential for anyone planning to reuse the content: the material reviewed here states no license. A repository of paper titles and links has a thin copyright surface, but the framework diagram at assets/Framework.png and any prose lifted from the survey are not covered by an open license as far as this material shows. If you intend to republish the taxonomy diagram or the section structure, you need to establish the terms first. Nothing here should be read as legal advice; the point is that the terms are not stated, and you should not assume permissive reuse.

What It Cannot Do, and What to Use Instead

The failure mode is straightforward: this repository cannot tell you whether a given tool-calling approach works in your stack. It has no code, no evaluation harness, no latency figures, and no per-paper notes on implementation cost. If you need to compare how ToolkenGPT and Gorilla actually handle a large API surface, the README gives you the paper links and nothing else; you will read the papers. Contrast this with a framework such as LangChain, which ships runnable abstractions for tool definition, selection, and invocation, and lets you execute a chain today. The difference in approach is total: LangChain is an implementation surface you build on, while LLM-Tool-Survey is a map of the research that motivated surfaces like it. A second, closer alternative is the Awesome-list pattern of annotated paper collections, which typically add one-line summaries per entry. This repository does not. The absence of annotations is the single biggest gap, because a title and a venue rarely tell you whether a paper is worth your afternoon.

Maintenance Cost and the Freshness Question

The repository metadata shows a last push of 2026-09-09 and no releases. For a bibliography, a push is the only meaningful maintenance signal, and it does not reveal what changed. The venue labels raise a specific concern: entries marked Preprint with years like 2021, 2022, and 2023 may have since been published, retitled, or superseded, and the README gives no indication that the list is periodically reconciled against venue records. Tool learning moves quickly, and the survey's own framing is that the field is advancing rapidly while the literature stays disorganized. A static list reintroduces the problem it set out to solve, just at a smaller scale. The mitigation is cheap if you are already in the habit: check the arXiv identifier for each paper you plan to cite, and treat the venue label in this repository as a snapshot rather than a current fact. There is no stated cadence for updates, so you cannot tell how stale a given line is.

Editorial conclusion

Adopt this repository if you are scoping a tool-calling system and want a stage-by-stage reading list before you commit to an architecture, or if you are writing a related paper and need the citation in BibTeX form. Do not adopt it if you need installable code, a benchmark harness, or a maintained leaderboard. Before you rely on it, verify that the paper list under Tool Selection, Tool Calling, and Response Generation reflects work published after the survey's submission, and check whether the repository has a licence file at all, since none is stated in the material reviewed here.

Official sources

1. Issues
2. Project website
3. quchangle1/LLM-Tool-Survey on GitHub
4. README