{"paper":{"title":"Chain of Evidence: Pixel-Level Visual Attribution for Iterative Retrieval-Augmented Generation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Vision-language models can deliver pixel-level visual evidence chains for iterative retrieval-augmented generation by operating directly on document screenshots.","cross_cats":["cs.AI","cs.CL","cs.IR"],"primary_cat":"cs.CV","authors_text":"Di Liang, Peiyang Liu, Wei Ye, Xi Wang, Ziqiang Cui","submitted_at":"2026-05-02T06:40:42Z","abstract_excerpt":"Iterative Retrieval-Augmented Generation (iRAG) has emerged as a powerful paradigm for answering complex multi-hop questions by progressively retrieving and reasoning over external documents. However, current systems predominantly operate on parsed text, which creates two critical bottlenecks: (1) \\textit{Coarse-grained attribution}, where users are burdened with manually locating evidence within lengthy documents based on vague text-level citations; and (2) \\textit{Visual semantic loss}, where the conversion of visually rich documents (e.g., slides, PDFs with charts) into text discards spatia"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"fine-tuned Qwen3-VL-8B-Instruct achieves robust performance, significantly outperforming text-based baselines in scenarios requiring visual layout understanding, while establishing a retriever-agnostic solution for pixel-level interpretable iRAG.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That vision-language models applied to raw document screenshots can reliably recover spatial logic and layout cues that text conversion discards, without format-specific parsing or additional supervision beyond the fine-tuning described.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CoE applies vision-language models directly to document screenshots to deliver pixel-level bounding-box attribution for evidence in iterative retrieval-augmented generation, outperforming text baselines on visual-layout tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Vision-language models can deliver pixel-level visual evidence chains for iterative retrieval-augmented generation by operating directly on document screenshots.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"aac36e72c896125b8a1411355e904d9260963437b70889c9d8e7cd40bb9a88a5"},"source":{"id":"2605.01284","kind":"arxiv","version":2},"verdict":{"id":"ee448b55-1017-4652-b5e6-87e9ee16626d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-09T14:42:48.086500Z","strongest_claim":"fine-tuned Qwen3-VL-8B-Instruct achieves robust performance, significantly outperforming text-based baselines in scenarios requiring visual layout understanding, while establishing a retriever-agnostic solution for pixel-level interpretable iRAG.","one_line_summary":"CoE applies vision-language models directly to document screenshots to deliver pixel-level bounding-box attribution for evidence in iterative retrieval-augmented generation, outperforming text baselines on visual-layout tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That vision-language models applied to raw document screenshots can reliably recover spatial logic and layout cues that text conversion discards, without format-specific parsing or additional supervision beyond the fine-tuning described.","pith_extraction_headline":"Vision-language models can deliver pixel-level visual evidence chains for iterative retrieval-augmented generation by operating directly on document screenshots."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.01284/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T18:35:45.366112Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T17:27:00.327473Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"650733d44fd774141e603140059a529820898b3fa431f48d1df8b89b55a1e878"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}