{"paper":{"title":"ODRPO: Ordinal Decompositions of Discrete Rewards for Robust Policy Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Decomposing discrete rewards into ordinal binary indicators isolates evaluation noise and stabilizes policy updates in RLAIF without extra compute.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Fei Wang, Inderjit Dhillon, Nirmal Patel","submitted_at":"2026-05-12T19:17:14Z","abstract_excerpt":"The alignment of Large Language Models (LLMs) utilizes Reinforcement Learning from AI Feedback (RLAIF) for non-verifiable domains such as long-form question answering and open-ended instruction following. These domains often rely on LLM based auto-raters to provide granular, multi-tier discrete rewards (e.g., 1-10 rubrics) that are inherently stochastic due to prompt sensitivity and sampling randomness. We empirically verify the stochasticity of auto-raters that can propagate and corrupt standard advantage estimators like GRPO and MaxRL, as a noisy reward samples can skew normalization statist"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"ODRPO achieves robust performance on Qwen2.5-7B and Qwen3-4B models, outperforming baselines with relative improvements of upto 14.8% on FACTS-grounding-v2 and 7.5% on Alpaca-Evals. Critically, these gains are achieved with negligible training-time overhead, as ODRPO requires no additional compute per step compared to standard estimators.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That decomposing discrete rewards into ordinal binary indicators structurally isolates evaluation noise and prevents outlier evaluations from corrupting the global update, as stated in the abstract description of the framework.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"ODRPO decomposes discrete rewards into ordinal binary indicators to compute independent advantages and reduce noise corruption in RLAIF policy optimization.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Decomposing discrete rewards into ordinal binary indicators isolates evaluation noise and stabilizes policy updates in RLAIF without extra compute.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"98a7cb3a2c2621f4d7630879b5e9eb097a0ae8b35271b05abecc38f1c6e4bc6c"},"source":{"id":"2605.12667","kind":"arxiv","version":1},"verdict":{"id":"3589802d-eb98-4d60-a062-08a520eeac62","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T21:13:13.575935Z","strongest_claim":"ODRPO achieves robust performance on Qwen2.5-7B and Qwen3-4B models, outperforming baselines with relative improvements of upto 14.8% on FACTS-grounding-v2 and 7.5% on Alpaca-Evals. Critically, these gains are achieved with negligible training-time overhead, as ODRPO requires no additional compute per step compared to standard estimators.","one_line_summary":"ODRPO decomposes discrete rewards into ordinal binary indicators to compute independent advantages and reduce noise corruption in RLAIF policy optimization.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That decomposing discrete rewards into ordinal binary indicators structurally isolates evaluation noise and prevents outlier evaluations from corrupting the global update, as stated in the abstract description of the framework.","pith_extraction_headline":"Decomposing discrete rewards into ordinal binary indicators isolates evaluation noise and stabilizes policy updates in RLAIF without extra compute."},"references":{"count":41,"sample":[{"doi":"10.1038/s41586-025-09422-z","year":null,"title":"Nature645(8081), 633–638 (2025) https://doi.org/10.1038/s41586-025-09422-z","work_id":"9835b482-5032-4135-93dd-82a066677569","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles , author=. 2025 , eprint=","work_id":"7ea2960e-3de4-4048-be17-ddc3bb8327a6","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Hansen and Duo Peng and Yuhui Zhang and Alejandro Lozano and Min Woo Sun and Emma Lundberg and Serena Yeung-Levy , year=","work_id":"24cdde52-11e0-4409-8656-4a46c92a8552","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"ViCrit: A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs , author=. 2025 , eprint=","work_id":"08971129-3986-43de-9574-8e903aff78da","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Examining Reasoning LLMs-as-Judges in Non-Verifiable LLM Post-Training , author=. 2026 , eprint=","work_id":"01f55215-0262-4791-87f3-0a2c1cd0bf25","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":41,"snapshot_sha256":"4a64a2eec651d75f8ad382c46ecf342602c7241135526e5dbb9583aacf526677","internal_anchors":2},"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"}