{"paper":{"title":"OASES: Outcome-Aligned Search-Evaluation Co-Training for Agentic Search","license":"http://creativecommons.org/licenses/by/4.0/","headline":"OASES improves agentic search by co-training policies with outcome-aligned evaluators for better process rewards.","cross_cats":["cs.CL","cs.IR"],"primary_cat":"cs.AI","authors_text":"Erhan Zhang, Jiaxin Mao, Wei Yang, Xiaochi Wei, Yan Gao, Yao Hu, Yiqun Chen, Yi Wu, Zechun Niu","submitted_at":"2026-04-04T10:23:46Z","abstract_excerpt":"Agentic search enables language models to solve knowledge-intensive tasks by adaptively acquiring external evidence over multiple steps. Reinforcement learning with verifiable rewards (RLVR) has emerged as a widely adopted training paradigm for search agents, yet outcome-only rewards are sparse and provide limited credit assignment for intermediate search actions. Existing process-reward methods therefore seek to densify supervision through proxy signals, external evaluators, or likelihood-based information gain. However, proxy rewards can deviate from the final outcome objective, while fixed "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Experiments on five multi-hop QA benchmarks show that OASES consistently outperforms strong RL baselines, with further analyses confirming the benefits of outcome-aligned process rewards and search-evaluation co-training.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That co-training the evaluator on the evolving policy will produce reliable, non-stale process rewards without introducing new biases or instability that undermine the final outcome alignment.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"OASES co-trains search policies and evaluators to generate outcome-aligned process rewards, outperforming standard RL baselines on five multi-hop QA benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"OASES improves agentic search by co-training policies with outcome-aligned evaluators for better process rewards.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"55f67e31d805cec2b98399c31057da4981cf0b50f6bb63a6d9bd6005e6302495"},"source":{"id":"2604.03675","kind":"arxiv","version":3},"verdict":{"id":"b6697de3-c9bc-4f17-89f1-3de54d8a77fb","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T17:12:43.396904Z","strongest_claim":"Experiments on five multi-hop QA benchmarks show that OASES consistently outperforms strong RL baselines, with further analyses confirming the benefits of outcome-aligned process rewards and search-evaluation co-training.","one_line_summary":"OASES co-trains search policies and evaluators to generate outcome-aligned process rewards, outperforming standard RL baselines on five multi-hop QA benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That co-training the evaluator on the evolving policy will produce reliable, non-stale process rewards without introducing new biases or instability that undermine the final outcome alignment.","pith_extraction_headline":"OASES improves agentic search by co-training policies with outcome-aligned evaluators for better process rewards."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.03675/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":1,"snapshot_sha256":"08e6c632355f317b811f748568968728fb6ae2cef81d28a9178b43a5bb8f8421"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}