{"paper":{"title":"Towards On-Policy Data Evolution for Visual-Native Multimodal Deep Search Agents","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"On-policy data evolution from agent rollouts boosts multimodal deep search performance from 24.9% to 39% on average.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Chenxin Li, Guanting Dong, Hangyu Guo, Hongru Wang, Junting Lu, Shijue Huang, Shuang Chen, Xinyu Geng, Yi R. Fung, Zhaochen Su, Zhenyu Li","submitted_at":"2026-05-11T16:49:36Z","abstract_excerpt":"Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average, surpassing Gemini-2.5 Pro in standard agent-workflow setting (37.9%). At 30B, ODE raises the average score from 30.6% to 41.5%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That rollouts from the current policy produce training data that accurately identifies and fills the precise capability gaps without introducing self-reinforcing biases or training instability.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A new image-bank harness and closed-loop on-policy data evolution method raises multimodal agent performance on visual search benchmarks from 24.9% to 39.0% for an 8B model and from 30.6% to 41.5% for a 30B model.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"On-policy data evolution from agent rollouts boosts multimodal deep search performance from 24.9% to 39% on average.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5bb6ee0af1ca64bfb9cf6447b9a1bea750562ba4153f266f382f115c32e5b2a0"},"source":{"id":"2605.10832","kind":"arxiv","version":2},"verdict":{"id":"7a3df81e-29a5-422c-ad95-61721d7662e6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T04:10:59.872308Z","strongest_claim":"Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average, surpassing Gemini-2.5 Pro in standard agent-workflow setting (37.9%). At 30B, ODE raises the average score from 30.6% to 41.5%.","one_line_summary":"A new image-bank harness and closed-loop on-policy data evolution method raises multimodal agent performance on visual search benchmarks from 24.9% to 39.0% for an 8B model and from 30.6% to 41.5% for a 30B model.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That rollouts from the current policy produce training data that accurately identifies and fills the precise capability gaps without introducing self-reinforcing biases or training instability.","pith_extraction_headline":"On-policy data evolution from agent rollouts boosts multimodal deep search performance from 24.9% to 39% on average."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.10832/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T05:22:00.342835Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T14:34:47.442439Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T10:31:17.587837Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T08:57:25.750037Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"d5d3cb52ab2d139c70b8feedfeceeb355eb45c3f7582d09b4f05e0f3292d6e6a"},"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"}