{"paper":{"title":"ART: Automatic multi-step reasoning and tool-use for large language models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"ART lets large language models automatically generate multi-step reasoning programs that call external tools.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Bhargavi Paranjape, Hannaneh Hajishirzi, Luke Zettlemoyer, Marco Tulio Ribeiro, Sameer Singh, Scott Lundberg","submitted_at":"2023-03-16T01:04:45Z","abstract_excerpt":"Large language models (LLMs) can perform complex reasoning in few- and zero-shot settings by generating intermediate chain of thought (CoT) reasoning steps. Further, each reasoning step can rely on external tools to support computation beyond the core LLM capabilities (e.g. search/running code). Prior work on CoT prompting and tool use typically requires hand-crafting task-specific demonstrations and carefully scripted interleaving of model generations with tool use. We introduce Automatic Reasoning and Tool-use (ART), a framework that uses frozen LLMs to automatically generate intermediate re"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"ART achieves a substantial improvement over few-shot prompting and automatic CoT on unseen tasks in the BigBench and MMLU benchmarks, and matches performance of hand-crafted CoT prompts on a majority of these tasks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a fixed task library contains sufficiently diverse and high-quality demonstrations so that nearest-neighbor selection reliably supplies useful programs for entirely new, unseen tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"ART automatically generates multi-step reasoning programs with tool integration for LLMs, yielding substantial gains over few-shot and auto-CoT prompting on BigBench and MMLU while matching hand-crafted CoT on most tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"ART lets large language models automatically generate multi-step reasoning programs that call external tools.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"28d7690fa1db8d61d0b9b4bc61bdeadc255b91c797f197c58f2bfa9570b91f50"},"source":{"id":"2303.09014","kind":"arxiv","version":1},"verdict":{"id":"ad0922ee-78ed-4279-a5e7-512d55e1bcfc","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T18:58:48.398414Z","strongest_claim":"ART achieves a substantial improvement over few-shot prompting and automatic CoT on unseen tasks in the BigBench and MMLU benchmarks, and matches performance of hand-crafted CoT prompts on a majority of these tasks.","one_line_summary":"ART automatically generates multi-step reasoning programs with tool integration for LLMs, yielding substantial gains over few-shot and auto-CoT prompting on BigBench and MMLU while matching hand-crafted CoT on most tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a fixed task library contains sufficiently diverse and high-quality demonstrations so that nearest-neighbor selection reliably supplies useful programs for entirely new, unseen tasks.","pith_extraction_headline":"ART lets large language models automatically generate multi-step reasoning programs that call external tools."},"references":{"count":267,"sample":[{"doi":"10.18653/v1/2022.acl-long.579","year":2022,"title":"Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. https://doi.org/10.18653/v1/2022.acl-long.579 I nternet-augmented dialogue generation . In Proceedings of the 60th Annual Meeting of the Associat","work_id":"4b63a530-9990-4f21-a11b-29062ce01cc7","ref_index":15,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.18653/v1/2021.naacl-main.168","year":2021,"title":"Are NLP Models really able to Solve Simple Math Word Problems?","work_id":"01c9a44a-5ebf-4209-a628-5f1fd19d2103","ref_index":22,"cited_arxiv_id":"2103.07191","is_internal_anchor":true},{"doi":"","year":null,"title":"Chain-of-thought prompting elicits reasoning in large language models","work_id":"f4edf7fa-4b0c-45e2-86ab-96e1d16f24bb","ref_index":33,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Finetuned Language Models Are Zero-Shot Learners","work_id":"7ed6cdaa-ed67-4db4-aceb-b7e1b0e6e7c4","ref_index":37,"cited_arxiv_id":"2109.01652","is_internal_anchor":true},{"doi":"","year":null,"title":"Chain of Thought Prompting Elicits Reasoning in Large Language Models , author=. ArXiv , year=","work_id":"75eacb63-2829-4757-8923-44fbe8ccdd56","ref_index":38,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":267,"snapshot_sha256":"da9fa28cbe2d3794c22ddf8a4347b1d0c40e7c8f03a53f49ec8b01fd2f9f98f7","internal_anchors":38},"formal_canon":{"evidence_count":2,"snapshot_sha256":"8066222b8c40f1850eb92d5a7d0a32d7eeb36653bd62ca4232f5f123348f7d8e"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}