{"paper":{"title":"JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A single model trained on synthetic multimodal data creates a visual-programmatic interface that generates code from text, images, or both.","cross_cats":["cs.CL","cs.CV","cs.SE"],"primary_cat":"cs.AI","authors_text":"Ben Kao, Fei Yuan, Jingyang Gong, Kai Chen, Lei Li, Qiaosheng Chen, Qipeng Guo, Qiushi Sun, Yang Liu","submitted_at":"2025-10-27T17:13:49Z","abstract_excerpt":"The scope of neural code intelligence is rapidly expanding beyond text-based source code to encompass the rich visual outputs that programs generate. This visual dimension is critical for advanced applications like flexible content generation and precise, program-driven editing of visualizations. However, progress has been impeded by the scarcity of high-quality multimodal code data, a bottleneck stemming from challenges in synthesis and quality assessment. To address these challenges, we make contributions from both a data and modeling perspective. We first introduce a complete synthesis tool"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our unified model is a departure from existing approaches that build specialized models for isolated tasks. Extensive experiments on both text-centric and vision-centric coding tasks demonstrate the superior performance of the JanusCoder series, with our 7B to 14B scale models approaching or even exceeding the performance of commercial models.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The synthesized JanusCode-800K dataset is assumed to be high-quality and representative enough to train models that generalize to real-world visual-programmatic tasks, without detailed validation against human-created or external benchmarks mentioned in the abstract.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"JanusCoder is a unified visual-programmatic model for code generation from text and/or visual inputs, powered by the new JanusCode-800K multimodal dataset.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A single model trained on synthetic multimodal data creates a visual-programmatic interface that generates code from text, images, or both.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"29a80baed31fb0214fe50da7f0ef3e9fdf4cca53c19cf753c97d18dacc294c5d"},"source":{"id":"2510.23538","kind":"arxiv","version":3},"verdict":{"id":"6ad96179-8f73-4a78-b883-516d27c54acd","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-18T04:04:02.016344Z","strongest_claim":"Our unified model is a departure from existing approaches that build specialized models for isolated tasks. Extensive experiments on both text-centric and vision-centric coding tasks demonstrate the superior performance of the JanusCoder series, with our 7B to 14B scale models approaching or even exceeding the performance of commercial models.","one_line_summary":"JanusCoder is a unified visual-programmatic model for code generation from text and/or visual inputs, powered by the new JanusCode-800K multimodal dataset.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The synthesized JanusCode-800K dataset is assumed to be high-quality and representative enough to train models that generalize to real-world visual-programmatic tasks, without detailed validation against human-created or external benchmarks mentioned in the abstract.","pith_extraction_headline":"A single model trained on synthetic multimodal data creates a visual-programmatic interface that generates code from text, images, or both."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2510.23538/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"e1489c221a298059dc0375fe0ba90abb7eaee4d6c4ac565e62edd6cbe469f9cb"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}