{"paper":{"title":"ViperGPT: Visual Inference via Python Execution for Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"ViperGPT uses code generation to create Python programs that combine vision models for answering complex visual queries without training.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Carl Vondrick, D\\'idac Sur\\'is, Sachit Menon","submitted_at":"2023-03-14T17:57:47Z","abstract_excerpt":"Answering visual queries is a complex task that requires both visual processing and reasoning. End-to-end models, the dominant approach for this task, do not explicitly differentiate between the two, limiting interpretability and generalization. Learning modular programs presents a promising alternative, but has proven challenging due to the difficulty of learning both the programs and modules simultaneously. We introduce ViperGPT, a framework that leverages code-generation models to compose vision-and-language models into subroutines to produce a result for any query. ViperGPT utilizes a prov"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"This simple approach requires no further training, and achieves state-of-the-art results across various complex visual tasks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a code-generation model can reliably produce correct, executable compositions of the provided vision modules for arbitrary queries without systematic errors in program logic or API usage.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"ViperGPT generates executable Python code to compose pre-trained vision-and-language modules into programs that answer visual queries, reaching state-of-the-art results with no additional training.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"ViperGPT uses code generation to create Python programs that combine vision models for answering complex visual queries without training.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f15980db7861e4a53f01b42c8d6dfe8448e4e929cb3fe54574e11b2f92be62ff"},"source":{"id":"2303.08128","kind":"arxiv","version":1},"verdict":{"id":"58654800-9a63-4a21-a99b-ddaa12a324a7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T18:09:01.184006Z","strongest_claim":"This simple approach requires no further training, and achieves state-of-the-art results across various complex visual tasks.","one_line_summary":"ViperGPT generates executable Python code to compose pre-trained vision-and-language modules into programs that answer visual queries, reaching state-of-the-art results with no additional training.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a code-generation model can reliably produce correct, executable compositions of the provided vision modules for arbitrary queries without systematic errors in program logic or API usage.","pith_extraction_headline":"ViperGPT uses code generation to create Python programs that combine vision models for answering complex visual queries without training."},"references":{"count":66,"sample":[{"doi":"","year":2022,"title":"Flamingo: a visual language model for few-shot learning","work_id":"6c7c3027-f918-40e9-b9e8-c682e15c12a3","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"Neural module networks","work_id":"5d0a6c6e-61ca-40d3-9377-b7330e86cf81","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"arXiv preprint arXiv:1811.12889 , year=","work_id":"43c4bd82-57c8-4d3e-8e94-66ae1ffab8f3","ref_index":3,"cited_arxiv_id":"1811.12889","is_internal_anchor":true},{"doi":"","year":2019,"title":"The Consciousness Prior, Dec","work_id":"3d734855-ef5a-4524-afd9-bedb436a7293","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Bravo, Sudhanshu Mittal, Simon Ging, and Thomas Brox","work_id":"c6fb2c6a-4031-4584-b576-d3ca0ed8686a","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":66,"snapshot_sha256":"3a390387f7b38ac80d23d44eff6c3526995982f9a376079f94e3bc33d09ec62f","internal_anchors":12},"formal_canon":{"evidence_count":1,"snapshot_sha256":"8e12ce63628226e7721f95aaf10ac650aecb66d8c42e4a5cea1dd417ee4dd854"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}