{"paper":{"title":"Process Reward Agents for Steering Knowledge-Intensive Reasoning","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Process Reward Agents supply online step-wise rewards from external knowledge to steer reasoning in frozen language models.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Jiwoong Sohn, Kenneth Styppa, Michael Moor, Tomasz Sternal, Torsten Hoefler","submitted_at":"2026-04-10T16:45:44Z","abstract_excerpt":"Reasoning in knowledge-intensive domains remains challenging as intermediate steps are often not locally verifiable: unlike math or code, evaluating step correctness may require synthesizing clues across large external knowledge sources. As a result, subtle errors can propagate through reasoning traces, potentially never to be detected. Prior work has proposed process reward models (PRMs), including retrieval-augmented variants, but these methods operate post hoc, scoring completed trajectories, which prevents their integration into dynamic inference procedures. Here, we introduce Process Rewa"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"PRA consistently outperforms strong baselines, achieving 80.8% accuracy on MedQA with Qwen3-4B, a new state of the art at the 4B scale. Importantly, PRA generalizes to unseen frozen policy models ranging from 0.5B to 8B parameters, improving their accuracy by up to 25.7% without any policy model updates.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That retrieval-augmented process rewards can be computed reliably and cheaply at every generation step from external knowledge sources without introducing undetected errors or prohibitive latency that would negate the search benefits.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Process Reward Agents enable online step-wise guidance for frozen AI models in medical reasoning, raising accuracy to 80.8% on MedQA and up to 25.7% gains across 0.5B-8B models without policy updates.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Process Reward Agents supply online step-wise rewards from external knowledge to steer reasoning in frozen language models.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"bf2bc304333b3d191247118c6c200e1e5447a6369ea4569439463b0c149ac691"},"source":{"id":"2604.09482","kind":"arxiv","version":2},"verdict":{"id":"5aff9e65-e160-4734-9b40-00819bb5c476","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T17:57:44.951995Z","strongest_claim":"PRA consistently outperforms strong baselines, achieving 80.8% accuracy on MedQA with Qwen3-4B, a new state of the art at the 4B scale. Importantly, PRA generalizes to unseen frozen policy models ranging from 0.5B to 8B parameters, improving their accuracy by up to 25.7% without any policy model updates.","one_line_summary":"Process Reward Agents enable online step-wise guidance for frozen AI models in medical reasoning, raising accuracy to 80.8% on MedQA and up to 25.7% gains across 0.5B-8B models without policy updates.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That retrieval-augmented process rewards can be computed reliably and cheaply at every generation step from external knowledge sources without introducing undetected errors or prohibitive latency that would negate the search benefits.","pith_extraction_headline":"Process Reward Agents supply online step-wise rewards from external knowledge to steer reasoning in frozen language models."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.09482/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":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}