{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2025:M67MY6ZAN6K7XVKPTZJ5GESVBL","short_pith_number":"pith:M67MY6ZA","canonical_record":{"source":{"id":"2512.13399","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2025-12-15T14:50:08Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"18e7e9880fc372af62925814be9070d0e90dc57ac5299358a60bf65189958dcb","abstract_canon_sha256":"83b62e3afefe8ebba1f359eb106851879c9fd34e368906c9292498060439aedc"},"schema_version":"1.0"},"canonical_sha256":"67becc7b206f95fbd54f9e53d312550ada4e9b3dc46b0f9efb9163b934c0d914","source":{"kind":"arxiv","id":"2512.13399","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2512.13399","created_at":"2026-05-18T03:09:32Z"},{"alias_kind":"arxiv_version","alias_value":"2512.13399v2","created_at":"2026-05-18T03:09:32Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2512.13399","created_at":"2026-05-18T03:09:32Z"},{"alias_kind":"pith_short_12","alias_value":"M67MY6ZAN6K7","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"M67MY6ZAN6K7XVKP","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"M67MY6ZA","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2025:M67MY6ZAN6K7XVKPTZJ5GESVBL","target":"record","payload":{"canonical_record":{"source":{"id":"2512.13399","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2025-12-15T14:50:08Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"18e7e9880fc372af62925814be9070d0e90dc57ac5299358a60bf65189958dcb","abstract_canon_sha256":"83b62e3afefe8ebba1f359eb106851879c9fd34e368906c9292498060439aedc"},"schema_version":"1.0"},"canonical_sha256":"67becc7b206f95fbd54f9e53d312550ada4e9b3dc46b0f9efb9163b934c0d914","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:09:32.606985Z","signature_b64":"EjZbgvhXnHoKc1F6dak3pkBIrW4Y+FmKrfz+UpnWsVG29p9S0+2Tq4bHsxUM8iqKdGevTf2nmlIgU4aYV/+sCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"67becc7b206f95fbd54f9e53d312550ada4e9b3dc46b0f9efb9163b934c0d914","last_reissued_at":"2026-05-18T03:09:32.606466Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:09:32.606466Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2512.13399","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T03:09:32Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"MrxlKehuW48lBmgFjxuug2zFQk7cNc4+dgW5OT6GV/whcHSp7gwYNHYo1c1ImUdU9HnxmK7N0uuPnDkV+rkmAQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-05T14:46:56.101366Z"},"content_sha256":"ca6513ef815793a4e8ca0b30e12ee0b8c727ebaded26e1a879cca6181e3bce84","schema_version":"1.0","event_id":"sha256:ca6513ef815793a4e8ca0b30e12ee0b8c727ebaded26e1a879cca6181e3bce84"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2025:M67MY6ZAN6K7XVKPTZJ5GESVBL","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Differentiable Evolutionary Reinforcement Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"DERL makes evolutionary reward search differentiable by updating a meta-optimizer with policy gradients from inner-loop validation performance.","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Difan Zou, Sitao Cheng, Tianle Li, Xuhan Huang, Xunjian Yin","submitted_at":"2025-12-15T14:50:08Z","abstract_excerpt":"Crafting effective reward signals remains a central challenge in Reinforcement Learning (RL), especially for complex reasoning tasks. Existing automated reward optimization methods typically rely on derivative-free search heuristics that treat the reward function as a black box, failing to exploit the causal dynamics between reward structure modifications and policy performance. We introduce Differentiable Evolutionary Reinforcement Learning (DERL), a bi-level framework for the autonomous discovery of optimal reward structures. DERL employs a Meta-Optimizer that evolves a reward function throu"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"DERL achieves state-of-the-art performance on agent benchmarks, substantially outperforming non-differentiable baselines-especially in out-of-distribution generalization.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That policy gradients derived from inner-loop validation performance provide stable, unbiased, and dense feedback for updating the meta-optimizer without introducing circularity or instability in the bi-level optimization.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DERL is a differentiable bi-level method that evolves optimal reward structures for RL policies by composing atomic primitives and using meta-gradients from validation performance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"DERL makes evolutionary reward search differentiable by updating a meta-optimizer with policy gradients from inner-loop validation performance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"1857ac5ef5358f128e12d8cd0e2ee7334965015a4d74a0defd0439d094673c5b"},"source":{"id":"2512.13399","kind":"arxiv","version":2},"verdict":{"id":"657f36d0-ebe4-401b-acc4-7362da9e4e1a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T22:31:21.753597Z","strongest_claim":"DERL achieves state-of-the-art performance on agent benchmarks, substantially outperforming non-differentiable baselines-especially in out-of-distribution generalization.","one_line_summary":"DERL is a differentiable bi-level method that evolves optimal reward structures for RL policies by composing atomic primitives and using meta-gradients from validation performance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That policy gradients derived from inner-loop validation performance provide stable, unbiased, and dense feedback for updating the meta-optimizer without introducing circularity or instability in the bi-level optimization.","pith_extraction_headline":"DERL makes evolutionary reward search differentiable by updating a meta-optimizer with policy gradients from inner-loop validation performance."},"references":{"count":29,"sample":[{"doi":"","year":null,"title":"Concrete Problems in AI Safety","work_id":"c8d14fbe-6eab-464a-95b3-778aabd82fa3","ref_index":1,"cited_arxiv_id":"1606.06565","is_internal_anchor":true},{"doi":"","year":null,"title":"2016 , month = nov, journal =","work_id":"9919eb6a-fde4-4d39-8047-21197365166d","ref_index":2,"cited_arxiv_id":"1606.04474","is_internal_anchor":true},{"doi":"","year":2017,"title":"under review","work_id":"f199d869-08bd-48c0-9a77-0e81bc5884b6","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, Yifeng Lu, and Quoc V","work_id":"553680c9-3aa5-4d00-8f57-a63d5e24c52d","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Xingwu Chen, Tianle Li, and Difan Zou","work_id":"0aa79354-51a4-4a06-8f4d-ea1af0dd88e4","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":29,"snapshot_sha256":"f25d1dd2cca3b1cc4383665edd18d12dbcca3c66ec103ad55c42977c6c514e6a","internal_anchors":18},"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"},"verdict_id":"657f36d0-ebe4-401b-acc4-7362da9e4e1a"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T03:09:32Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"PHvnrcq4+VPeHRNVAu8INSd83YY4A/Zx052pyZt+HJGNKmnc48GIGPzHObUmQ+leB1HN7pmUpG48GBusqdbnBw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-05T14:46:56.101899Z"},"content_sha256":"989b83b4c02822b757395e5d4c17131a4fc0297e2ea98f777e87fb0b72a2fb7d","schema_version":"1.0","event_id":"sha256:989b83b4c02822b757395e5d4c17131a4fc0297e2ea98f777e87fb0b72a2fb7d"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/M67MY6ZAN6K7XVKPTZJ5GESVBL/bundle.json","state_url":"https://pith.science/pith/M67MY6ZAN6K7XVKPTZJ5GESVBL/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/M67MY6ZAN6K7XVKPTZJ5GESVBL/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-06-05T14:46:56Z","links":{"resolver":"https://pith.science/pith/M67MY6ZAN6K7XVKPTZJ5GESVBL","bundle":"https://pith.science/pith/M67MY6ZAN6K7XVKPTZJ5GESVBL/bundle.json","state":"https://pith.science/pith/M67MY6ZAN6K7XVKPTZJ5GESVBL/state.json","well_known_bundle":"https://pith.science/.well-known/pith/M67MY6ZAN6K7XVKPTZJ5GESVBL/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2025:M67MY6ZAN6K7XVKPTZJ5GESVBL","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"83b62e3afefe8ebba1f359eb106851879c9fd34e368906c9292498060439aedc","cross_cats_sorted":["cs.CL"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2025-12-15T14:50:08Z","title_canon_sha256":"18e7e9880fc372af62925814be9070d0e90dc57ac5299358a60bf65189958dcb"},"schema_version":"1.0","source":{"id":"2512.13399","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2512.13399","created_at":"2026-05-18T03:09:32Z"},{"alias_kind":"arxiv_version","alias_value":"2512.13399v2","created_at":"2026-05-18T03:09:32Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2512.13399","created_at":"2026-05-18T03:09:32Z"},{"alias_kind":"pith_short_12","alias_value":"M67MY6ZAN6K7","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"M67MY6ZAN6K7XVKP","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"M67MY6ZA","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:989b83b4c02822b757395e5d4c17131a4fc0297e2ea98f777e87fb0b72a2fb7d","target":"graph","created_at":"2026-05-18T03:09:32Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"DERL achieves state-of-the-art performance on agent benchmarks, substantially outperforming non-differentiable baselines-especially in out-of-distribution generalization."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That policy gradients derived from inner-loop validation performance provide stable, unbiased, and dense feedback for updating the meta-optimizer without introducing circularity or instability in the bi-level optimization."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"DERL is a differentiable bi-level method that evolves optimal reward structures for RL policies by composing atomic primitives and using meta-gradients from validation performance."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"DERL makes evolutionary reward search differentiable by updating a meta-optimizer with policy gradients from inner-loop validation performance."}],"snapshot_sha256":"1857ac5ef5358f128e12d8cd0e2ee7334965015a4d74a0defd0439d094673c5b"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"abstract_excerpt":"Crafting effective reward signals remains a central challenge in Reinforcement Learning (RL), especially for complex reasoning tasks. Existing automated reward optimization methods typically rely on derivative-free search heuristics that treat the reward function as a black box, failing to exploit the causal dynamics between reward structure modifications and policy performance. We introduce Differentiable Evolutionary Reinforcement Learning (DERL), a bi-level framework for the autonomous discovery of optimal reward structures. DERL employs a Meta-Optimizer that evolves a reward function throu","authors_text":"Difan Zou, Sitao Cheng, Tianle Li, Xuhan Huang, Xunjian Yin","cross_cats":["cs.CL"],"headline":"DERL makes evolutionary reward search differentiable by updating a meta-optimizer with policy gradients from inner-loop validation performance.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2025-12-15T14:50:08Z","title":"Differentiable Evolutionary Reinforcement Learning"},"references":{"count":29,"internal_anchors":18,"resolved_work":29,"sample":[{"cited_arxiv_id":"1606.06565","doi":"","is_internal_anchor":true,"ref_index":1,"title":"Concrete Problems in AI Safety","work_id":"c8d14fbe-6eab-464a-95b3-778aabd82fa3","year":null},{"cited_arxiv_id":"1606.04474","doi":"","is_internal_anchor":true,"ref_index":2,"title":"2016 , month = nov, journal =","work_id":"9919eb6a-fde4-4d39-8047-21197365166d","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"under review","work_id":"f199d869-08bd-48c0-9a77-0e81bc5884b6","year":2017},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, Yifeng Lu, and Quoc V","work_id":"553680c9-3aa5-4d00-8f57-a63d5e24c52d","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Xingwu Chen, Tianle Li, and Difan Zou","work_id":"0aa79354-51a4-4a06-8f4d-ea1af0dd88e4","year":2023}],"snapshot_sha256":"f25d1dd2cca3b1cc4383665edd18d12dbcca3c66ec103ad55c42977c6c514e6a"},"source":{"id":"2512.13399","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-16T22:31:21.753597Z","id":"657f36d0-ebe4-401b-acc4-7362da9e4e1a","model_set":{"reader":"grok-4.3"},"one_line_summary":"DERL is a differentiable bi-level method that evolves optimal reward structures for RL policies by composing atomic primitives and using meta-gradients from validation performance.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"DERL makes evolutionary reward search differentiable by updating a meta-optimizer with policy gradients from inner-loop validation performance.","strongest_claim":"DERL achieves state-of-the-art performance on agent benchmarks, substantially outperforming non-differentiable baselines-especially in out-of-distribution generalization.","weakest_assumption":"That policy gradients derived from inner-loop validation performance provide stable, unbiased, and dense feedback for updating the meta-optimizer without introducing circularity or instability in the bi-level optimization."}},"verdict_id":"657f36d0-ebe4-401b-acc4-7362da9e4e1a"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:ca6513ef815793a4e8ca0b30e12ee0b8c727ebaded26e1a879cca6181e3bce84","target":"record","created_at":"2026-05-18T03:09:32Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"83b62e3afefe8ebba1f359eb106851879c9fd34e368906c9292498060439aedc","cross_cats_sorted":["cs.CL"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2025-12-15T14:50:08Z","title_canon_sha256":"18e7e9880fc372af62925814be9070d0e90dc57ac5299358a60bf65189958dcb"},"schema_version":"1.0","source":{"id":"2512.13399","kind":"arxiv","version":2}},"canonical_sha256":"67becc7b206f95fbd54f9e53d312550ada4e9b3dc46b0f9efb9163b934c0d914","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"67becc7b206f95fbd54f9e53d312550ada4e9b3dc46b0f9efb9163b934c0d914","first_computed_at":"2026-05-18T03:09:32.606466Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T03:09:32.606466Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"EjZbgvhXnHoKc1F6dak3pkBIrW4Y+FmKrfz+UpnWsVG29p9S0+2Tq4bHsxUM8iqKdGevTf2nmlIgU4aYV/+sCw==","signature_status":"signed_v1","signed_at":"2026-05-18T03:09:32.606985Z","signed_message":"canonical_sha256_bytes"},"source_id":"2512.13399","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:ca6513ef815793a4e8ca0b30e12ee0b8c727ebaded26e1a879cca6181e3bce84","sha256:989b83b4c02822b757395e5d4c17131a4fc0297e2ea98f777e87fb0b72a2fb7d"],"state_sha256":"9de0311e04cb4507245cfa3a9bbcfcbc8a6b27da09795798a98d9c04c00730cc"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"4NpdqsDacQCFb3Y5u/ooJv+qOsTpdKt+AA94OThWLoTcgsZvkv+omM+tA2qihom7oe7WBEjKayqPddvXbAHcBw==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-05T14:46:56.104361Z","bundle_sha256":"eb583d091e016c6d995bb140766eba9e579fda479a6d27089597cd1346ec38fa"}}