{"paper":{"title":"Policy Improvement Reinforcement Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Reinforcement learning for language models can be made self-correcting by directly maximizing cumulative verified policy improvement across iterations.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Deqing Wang, Haoyi Zhou, Huaiyang Wang, Jianxin Li, Xiaojie Li, Yaodong Yang, Yikun Ban, Zixuan Huang","submitted_at":"2026-04-01T13:10:20Z","abstract_excerpt":"Reinforcement Learning with Verifiable Rewards (RLVR) has become a central post-training paradigm for improving the reasoning capabilities of large language models. Yet existing methods share a common blind spot: they optimize policies based on instantaneous group-level or batch-level statistics without ever verifying whether the resulting update actually improved the model. This open-loop design -- updating in isolation at each step, guided only by within-group (batch) reward signals -- means optimization can drift or collapse with no mechanism to detect and correct these failures. We argue t"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"we prove this temporal objective is perfectly aligned with maximizing final task performance... PIPO performs ascent on the PIRL objective in expectation","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"that retrospective verification against a sliding-window historical baseline can reliably identify genuine policy improvement without bias from window size, data distribution, or the specific choice of baseline statistics","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PIRL maximizes cumulative policy improvement across iterations instead of surrogate rewards and is proven aligned with final performance; PIPO implements it via retrospective verification for stable closed-loop optimization.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning for language models can be made self-correcting by directly maximizing cumulative verified policy improvement across iterations.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0771c5562a4e0c43d8a94a8533c81b3caba9b523c33214326d48d0abe76fd021"},"source":{"id":"2604.00860","kind":"arxiv","version":3},"verdict":{"id":"8f1b7e82-9cce-4ad9-9397-9b33dc804b4b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T22:44:54.577345Z","strongest_claim":"we prove this temporal objective is perfectly aligned with maximizing final task performance... PIPO performs ascent on the PIRL objective in expectation","one_line_summary":"PIRL maximizes cumulative policy improvement across iterations instead of surrogate rewards and is proven aligned with final performance; PIPO implements it via retrospective verification for stable closed-loop optimization.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"that retrospective verification against a sliding-window historical baseline can reliably identify genuine policy improvement without bias from window size, data distribution, or the specific choice of baseline statistics","pith_extraction_headline":"Reinforcement learning for language models can be made self-correcting by directly maximizing cumulative verified policy improvement across iterations."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.00860/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"}