{"paper":{"title":"Reward-Conditioned Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Conditioning RL agents on reward parameters during single-objective training enables zero-shot adaptation to new rewards via replay data alone.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Marek Cygan, Michal Nauman, Pieter Abbeel","submitted_at":"2026-03-05T11:29:17Z","abstract_excerpt":"Single-task RL agents are typically trained under a fixed reward function, which limits their robustness to reward misspecification and their ability to adapt to changing preferences. We introduce Reward-Conditioned Reinforcement Learning (RCRL), an off-policy method that conditions agents on reward parameterizations while collecting experience under a single nominal objective. By recomputing counterfactual rewards from shared replay data, RCRL exposes the agent to multiple reward objectives without additional environment interaction, connecting single-task RL with ideas from multi-objective a"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"RCRL improves sample efficiency under the nominal reward parameterization, enables efficient adaptation to new parameterizations, and supports zero-shot behavioral adjustment at deployment.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That recomputing counterfactual rewards from replay data collected under the nominal policy produces unbiased training signals for other reward parameterizations.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"RCRL conditions RL policies on reward parameters and uses shared replay data to train for multiple objectives under a single nominal reward, improving efficiency and adaptability.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Conditioning RL agents on reward parameters during single-objective training enables zero-shot adaptation to new rewards via replay data alone.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ec667cc40c61ec88d2bb4450fb4e3a40c40f419c5ecff6bbb8650ca19c74c784"},"source":{"id":"2603.05066","kind":"arxiv","version":3},"verdict":{"id":"abd32904-dc8c-45a7-ab96-799f4e4d18f8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T16:05:50.794154Z","strongest_claim":"RCRL improves sample efficiency under the nominal reward parameterization, enables efficient adaptation to new parameterizations, and supports zero-shot behavioral adjustment at deployment.","one_line_summary":"RCRL conditions RL policies on reward parameters and uses shared replay data to train for multiple objectives under a single nominal reward, improving efficiency and adaptability.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That recomputing counterfactual rewards from replay data collected under the nominal policy produces unbiased training signals for other reward parameterizations.","pith_extraction_headline":"Conditioning RL agents on reward parameters during single-objective training enables zero-shot adaptation to new rewards via replay data alone."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2603.05066/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":2,"snapshot_sha256":"ef1afa9b6df446ae10a438a688321c081a7446c6e8340ec9c70cb8bf942e7a4e"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}