# Paper authors introduce LIBRA, a framework for executable tool libraries [Canonical reading](https://escapemeridian.com/readings/paper-authors-introduce-libra-a-framework-for-executable-tool-51eba475) The authors introduce LIBRA, a way to organize a flat tool catalog into an executable tool library before users submit queries. They argue that retrieving tools one at a time by how closely each matches a request can miss tools needed earlier in a multi-tool workflow. LIBRA uses an LLM to propose contracts describing candidate invocations, reusable outputs, state changes and success conditions; it tests proposed compositions in a sandbox and keeps relations supported by runtime tests. The authors say the library needs no dependency annotations, tool-use trajectories or task-specific training, and can support one-shot and agentic retrieval. Across AppWorld, BFCL and SciToolEval, they report average improvements over the same dense retriever of 7.9 percentage points in Recall and 3.7 in NDCG for one-shot retrieval. Within ToolReAGt, they report gains of 4.5 and 2.9 points, respectively, after ten retrieval iterations. The results suggest a possible way to recover prerequisite tools that similarity ranking alone may omit, while the reported gains concern benchmark retrieval performance. The paper’s comments also note sandbox dependencies and library-construction trade-offs. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Nov 3, 2026 (2026-11-03). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T02:01:29.934Z. First observed: 2026-11-05T05:09:23.114Z. Retrieved: 2026-11-10T01:55:09.873Z. - [LIBRA: Programmatic Induction of Executable Tool Libraries](https://arxiv.org/abs/2611.05445) ## Limits The supplied evidence is the arXiv page text and abstract; it does not give details of benchmark setup, task coverage or uncertainty around the reported gains. The source notes sandbox dependencies and library-construction trade-offs but the supplied text does not describe their specific effects. ## Open question How do LIBRA’s sandbox dependencies and library-construction trade-offs affect its performance across the reported benchmarks? ## Source-stated The authors say LIBRA tests proposed tool compositions in a sandbox and retains relations supported at runtime to build an executable tool library. > LIBRA then tests these compositions in a sandbox and retains the relations supported at runtime - [Supporting source](https://arxiv.org/abs/2611.05445) ## Source-stated The authors report that LIBRA improves one-shot Recall and NDCG over the same dense retriever across three benchmarks. > On AppWorld, BFCL, and SciToolEval, LIBRA improves one-shot Recall and NDCG over the same dense retriever by 7.9 and 3.7 percentage points on average. - [Supporting source](https://arxiv.org/abs/2611.05445) ## Source-stated The authors report gains in Recall and NDCG within ToolReAGt after ten retrieval iterations. > Within ToolReAGt, it yields corresponding gains of 4.5 and 2.9 points after ten retrieval iterations. - [Supporting source](https://arxiv.org/abs/2611.05445) ## Source-stated The abstract identifies sandbox dependencies and library-construction trade-offs as issues discussed alongside the reported results. > Demonstrates consistent performance gains across three benchmarks while discussing sandbox dependencies and library construction trade-offs - [Supporting source](https://arxiv.org/abs/2611.05445) ## Source and contributor context Li, Zekai, Liu, Zhenke, Li, Yonghao, Liu, Junhan, Xiao, Xulong, Zhang, Jiapeng · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Tool use A model's ability to operate other software, such as a search engine, a calculator or a code runner, while it works. - [Definition reference](https://docs.anthropic.com/en/docs/build-with-claude/tool-use/overview) ## Agentic AI AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer. - [Definition reference](https://developers.google.com/machine-learning/glossary#agentic) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 8f16e8aa-53a2-414e-8ca1-a4df40491ef7. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T02:01:29.934Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 2a05f6511800ff92bc793f557f11c0b6d1c2058f4609b98a5ef0e863b81dd694. Source text hash: 70a5bc638ff230274f153c86e7b10ae9a216ebf7051dcdb7636c3f85639b16ef. ## Revision and source-check history ### Version 1 — 2026-11-10T02:01:29.934Z The authors introduce LIBRA, a way to organize a flat tool catalog into an executable tool library before users submit queries. They argue that retrieving tools one at a time by how closely each matches a request can miss tools needed earlier in a multi-tool workflow. LIBRA uses an LLM to propose contracts describing candidate invocations, reusable outputs, state changes and success conditions; it tests proposed compositions in a sandbox and keeps relations supported by runtime tests. The authors say the library needs no dependency annotations, tool-use trajectories or task-specific training, and can support one-shot and agentic retrieval. Across AppWorld, BFCL and SciToolEval, they report average improvements over the same dense retriever of 7.9 percentage points in Recall and 3.7 in NDCG for one-shot retrieval. Within ToolReAGt, they report gains of 4.5 and 2.9 points, respectively, after ten retrieval iterations. The results suggest a possible way to recover prerequisite tools that similarity ranking alone may omit, while the reported gains concern benchmark retrieval performance. The paper’s comments also note sandbox dependencies and library-construction trade-offs. ## Original source arXiv cs.AI · Article Original source date: Nov 3, 2026 (2026-11-03). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T02:01:29.934Z. First observed: 2026-11-05T05:09:23.114Z. Retrieved: 2026-11-10T01:55:09.873Z. - [LIBRA: Programmatic Induction of Executable Tool Libraries](https://arxiv.org/abs/2611.05445) ## Limits The supplied evidence is the arXiv page text and abstract; it does not give details of benchmark setup, task coverage or uncertainty around the reported gains. The source notes sandbox dependencies and library-construction trade-offs but the supplied text does not describe their specific effects. ## Open question How do LIBRA’s sandbox dependencies and library-construction trade-offs affect its performance across the reported benchmarks? ## Source-stated The authors say LIBRA tests proposed tool compositions in a sandbox and retains relations supported at runtime to build an executable tool library. > LIBRA then tests these compositions in a sandbox and retains the relations supported at runtime - [Supporting source](https://arxiv.org/abs/2611.05445) ## Source-stated The authors report that LIBRA improves one-shot Recall and NDCG over the same dense retriever across three benchmarks. > On AppWorld, BFCL, and SciToolEval, LIBRA improves one-shot Recall and NDCG over the same dense retriever by 7.9 and 3.7 percentage points on average. - [Supporting source](https://arxiv.org/abs/2611.05445) ## Source-stated The authors report gains in Recall and NDCG within ToolReAGt after ten retrieval iterations. > Within ToolReAGt, it yields corresponding gains of 4.5 and 2.9 points after ten retrieval iterations. - [Supporting source](https://arxiv.org/abs/2611.05445) ## Source-stated The abstract identifies sandbox dependencies and library-construction trade-offs as issues discussed alongside the reported results. > Demonstrates consistent performance gains across three benchmarks while discussing sandbox dependencies and library construction trade-offs - [Supporting source](https://arxiv.org/abs/2611.05445) ## Source and contributor context Li, Zekai, Liu, Zhenke, Li, Yonghao, Liu, Junhan, Xiao, Xulong, Zhang, Jiapeng · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Tool use A model's ability to operate other software, such as a search engine, a calculator or a code runner, while it works. - [Definition reference](https://docs.anthropic.com/en/docs/build-with-claude/tool-use/overview) ## Agentic AI AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer. - [Definition reference](https://developers.google.com/machine-learning/glossary#agentic) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 8f16e8aa-53a2-414e-8ca1-a4df40491ef7. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T02:01:29.934Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 2a05f6511800ff92bc793f557f11c0b6d1c2058f4609b98a5ef0e863b81dd694. Source text hash: 70a5bc638ff230274f153c86e7b10ae9a216ebf7051dcdb7636c3f85639b16ef. ### uncertain — 2026-11-10T02:01:29.934Z — anthropic/claude-opus-5.5 Publication: 8f16e8aa-53a2-414e-8ca1-a4df40491ef7 attribution — uncertain: Most of the reading correctly attributes LIBRA and its results to the paper's authors. However, claim 4 says "The abstract identifies sandbox dependencies..." That text appears in the arXiv Comments field, not the abstract. The summary gets this right ("The paper's comments also note"), so the claim is internally inconsistent. Smallest repair: change claim 4 to say "The arXiv comments field" instead of "The abstract." > Demonstrates consistent performance gains across three benchmarks while discussing sandbox dependencies and library construction trade-offs summary — pass: The summary leads with the actor and the development. It explains the problem plainly: ranking tools one by one by similarity misses prerequisite tools. It describes the mechanism (LLM-proposed contracts, sandbox testing, keeping relations supported at runtime) and reports the numbers with their conditions. Technical terms are kept exactly: Recall, NDCG, dense retriever, one-shot and agentic retrieval, ToolReAGt. The term library should supply tooltips for Recall, NDCG, dense retriever and ToolReAGt. The question follows from a gap the source names and does not lead the reader. > For multi tool tasks, this recovers the goal matching tool while omitting prerequisite tools required by the workflow. date — pass: originalPublished is 2026-11-03, which matches the v1 submission date. The 5 Nov 2026 date belongs to the later v2 revision and the feed entry, so it is correctly not used. > Submitted on 3 Nov 2026 (v1), last revised 5 Nov 2026 (this version, v2) support — pass: The figures 7.9/3.7 and 4.5/2.9 and their conditions (one-shot, same dense retriever, average over three benchmarks; ten retrieval iterations) match the source. The interpretive sentence is hedged ("suggest a possible way") and explicitly limits the gains to benchmark retrieval performance, so it does not overstate real-world reliability. The claims about no annotations or training are attributed to the authors. No benefits or costs are invented. > the library requires no dependency annotations, tool use trajectories, or task specific training, and supports both one-shot and agentic retrieval media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the supplied candidate and the most direct access to the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/51eba475735c44dc8e866fca2177d9b97a173b68141cd043bccca3e50899d493/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/paper-authors-introduce-libra-a-framework-for-executable-tool-51eba475.json)