{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:5REKE2YFOLBZPHUCJAMHNWGBQ6","short_pith_number":"pith:5REKE2YF","canonical_record":{"source":{"id":"2603.20521","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-20T21:45:51Z","cross_cats_sorted":["cs.AI","math.OC","stat.ML"],"title_canon_sha256":"b5a4ed10115b97f5e0b14a6a1d2fa8282a51962e886c7331359a9191de8dacb8","abstract_canon_sha256":"79eaba7752d70acf5a9440197c725b056453e35f9feb612e9d7b5250db9682a9"},"schema_version":"1.0"},"canonical_sha256":"ec48a26b0572c3979e82481876d8c187ac4d48fb3eb8f4e8ecf5f90ea2cec608","source":{"kind":"arxiv","id":"2603.20521","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2603.20521","created_at":"2026-05-18T02:45:04Z"},{"alias_kind":"arxiv_version","alias_value":"2603.20521v2","created_at":"2026-05-18T02:45:04Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.20521","created_at":"2026-05-18T02:45:04Z"},{"alias_kind":"pith_short_12","alias_value":"5REKE2YFOLBZ","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"5REKE2YFOLBZPHUC","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"5REKE2YF","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:5REKE2YFOLBZPHUCJAMHNWGBQ6","target":"record","payload":{"canonical_record":{"source":{"id":"2603.20521","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-20T21:45:51Z","cross_cats_sorted":["cs.AI","math.OC","stat.ML"],"title_canon_sha256":"b5a4ed10115b97f5e0b14a6a1d2fa8282a51962e886c7331359a9191de8dacb8","abstract_canon_sha256":"79eaba7752d70acf5a9440197c725b056453e35f9feb612e9d7b5250db9682a9"},"schema_version":"1.0"},"canonical_sha256":"ec48a26b0572c3979e82481876d8c187ac4d48fb3eb8f4e8ecf5f90ea2cec608","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T02:45:04.689629Z","signature_b64":"y+BSrx+8N+ijdt4rRuPpQepx0wfdNY7tBLoTSxlrPe5FuNT/sK0mwB06GLgALrU/YXyYP+30Bm9LQgL/LMKECg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ec48a26b0572c3979e82481876d8c187ac4d48fb3eb8f4e8ecf5f90ea2cec608","last_reissued_at":"2026-05-18T02:45:04.689093Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T02:45:04.689093Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2603.20521","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":"2026-05-18T02:45:04Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"ODEKphwUX6v9FNynTp00CHyMlX4Hobop2L4WmwIXDkPNd9K40SeK5+gPqKAcZMbF6Ea5K2b2HlCxflP0nwdSBg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-10T11:35:51.855946Z"},"content_sha256":"458ce79e98c222b0bbf3e7f4d5996acf515845c02c35c9c36bf8645a77c1a316","schema_version":"1.0","event_id":"sha256:458ce79e98c222b0bbf3e7f4d5996acf515845c02c35c9c36bf8645a77c1a316"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:5REKE2YFOLBZPHUCJAMHNWGBQ6","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Delightful Distributed Policy Gradient","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Delightful policy gradient gates updates by advantage times surprisal to suppress harmful rare failures while keeping rare successes in distributed RL.","cross_cats":["cs.AI","math.OC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Ian Osband","submitted_at":"2026-03-20T21:45:51Z","abstract_excerpt":"Distributed reinforcement learning trains on data from stale, buggy, or mismatched actors, producing actions with high surprisal (negative log-probability) under the learner's policy. The core difficulty is not surprising data per se, but \\emph{negative learning from surprising data}. High-surprisal failures can dominate finite-batch updates through large perpendicular components, while high-surprisal successes reveal opportunities the current policy would otherwise miss. The \\textit{Delightful Policy Gradient} (DG) separates these cases by gating each update with delight, the product of advan"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On MNIST with simulated staleness, DG without off-policy correction outperforms importance-weighted PG with exact behavior probabilities; on a transformer sequence task with staleness, actor bugs, reward corruption, and rare discovery, DG often achieves nearly order-of-magnitude lower error, with sample-efficiency advantage growing with task complexity.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the advantage sign is essential for the surprisal-based filter and that the tabular suppression of perpendicular second moment generalizes to the neural network experiments without additional corrections or assumptions about policy overlap.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Delightful Policy Gradient gates updates with advantage times surprisal to suppress rare failures while preserving rare successes in distributed RL with stale or buggy data.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Delightful policy gradient gates updates by advantage times surprisal to suppress harmful rare failures while keeping rare successes in distributed RL.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"bf742e9ddda694e45436a0d770a433553a93a785a56fbb59a0a899166000d6f3"},"source":{"id":"2603.20521","kind":"arxiv","version":2},"verdict":{"id":"4c73da56-29d1-401f-9994-5d47cc2fe997","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T07:50:24.169905Z","strongest_claim":"On MNIST with simulated staleness, DG without off-policy correction outperforms importance-weighted PG with exact behavior probabilities; on a transformer sequence task with staleness, actor bugs, reward corruption, and rare discovery, DG often achieves nearly order-of-magnitude lower error, with sample-efficiency advantage growing with task complexity.","one_line_summary":"Delightful Policy Gradient gates updates with advantage times surprisal to suppress rare failures while preserving rare successes in distributed RL with stale or buggy data.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the advantage sign is essential for the surprisal-based filter and that the tabular suppression of perpendicular second moment generalizes to the neural network experiments without additional corrections or assumptions about policy overlap.","pith_extraction_headline":"Delightful policy gradient gates updates by advantage times surprisal to suppress harmful rare failures while keeping rare successes in distributed RL."},"references":{"count":28,"sample":[{"doi":"","year":2024,"title":"Preference optimization as probabilistic inference.arXiv e-prints, pages arXiv–2410, 2024","work_id":"71615a52-4d21-4dc1-a6c9-e88006c1bbae","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"Unifying count-based exploration and intrinsic motivation","work_id":"2fe189ce-b35d-4171-b8bf-a289fb2e3110","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Decision transformer: Reinforcement learning via sequence modeling","work_id":"a90a034a-ee4a-4779-927b-507900e52b5b","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"First return, then explore.Nature, 590(7847):580–586, 2021","work_id":"6bcd818c-8c01-4c75-9915-3201f47c9839","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"IMPALA: Scalable distributed deep-RL with importance weighted actor-learner architectures","work_id":"d50d1964-9261-4245-ae99-87a659d8cd5a","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":28,"snapshot_sha256":"e3d3ecb7dd411315f5219689f0adfac797575500b14f3cf2c4f03857af4e7b6c","internal_anchors":4},"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":"4c73da56-29d1-401f-9994-5d47cc2fe997"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T02:45:04Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"zo4ej0W7aA5zeB90Grp41qCgYpG+UjXa3xkA5XPr3UWjREWd4WrgdRwJHhb8SWRLRz9BzoCOULyK3jxi8/hCAg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-10T11:35:51.857194Z"},"content_sha256":"a1ce2be9f2e1c096657aa12c64bb8578ddb3955b1aba2cfe805252c5582e8abc","schema_version":"1.0","event_id":"sha256:a1ce2be9f2e1c096657aa12c64bb8578ddb3955b1aba2cfe805252c5582e8abc"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/5REKE2YFOLBZPHUCJAMHNWGBQ6/bundle.json","state_url":"https://pith.science/pith/5REKE2YFOLBZPHUCJAMHNWGBQ6/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/5REKE2YFOLBZPHUCJAMHNWGBQ6/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":"2026-06-10T11:35:51Z","links":{"resolver":"https://pith.science/pith/5REKE2YFOLBZPHUCJAMHNWGBQ6","bundle":"https://pith.science/pith/5REKE2YFOLBZPHUCJAMHNWGBQ6/bundle.json","state":"https://pith.science/pith/5REKE2YFOLBZPHUCJAMHNWGBQ6/state.json","well_known_bundle":"https://pith.science/.well-known/pith/5REKE2YFOLBZPHUCJAMHNWGBQ6/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:5REKE2YFOLBZPHUCJAMHNWGBQ6","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":"79eaba7752d70acf5a9440197c725b056453e35f9feb612e9d7b5250db9682a9","cross_cats_sorted":["cs.AI","math.OC","stat.ML"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-20T21:45:51Z","title_canon_sha256":"b5a4ed10115b97f5e0b14a6a1d2fa8282a51962e886c7331359a9191de8dacb8"},"schema_version":"1.0","source":{"id":"2603.20521","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2603.20521","created_at":"2026-05-18T02:45:04Z"},{"alias_kind":"arxiv_version","alias_value":"2603.20521v2","created_at":"2026-05-18T02:45:04Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.20521","created_at":"2026-05-18T02:45:04Z"},{"alias_kind":"pith_short_12","alias_value":"5REKE2YFOLBZ","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"5REKE2YFOLBZPHUC","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"5REKE2YF","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:a1ce2be9f2e1c096657aa12c64bb8578ddb3955b1aba2cfe805252c5582e8abc","target":"graph","created_at":"2026-05-18T02:45:04Z","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":"On MNIST with simulated staleness, DG without off-policy correction outperforms importance-weighted PG with exact behavior probabilities; on a transformer sequence task with staleness, actor bugs, reward corruption, and rare discovery, DG often achieves nearly order-of-magnitude lower error, with sample-efficiency advantage growing with task complexity."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the advantage sign is essential for the surprisal-based filter and that the tabular suppression of perpendicular second moment generalizes to the neural network experiments without additional corrections or assumptions about policy overlap."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Delightful Policy Gradient gates updates with advantage times surprisal to suppress rare failures while preserving rare successes in distributed RL with stale or buggy data."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Delightful policy gradient gates updates by advantage times surprisal to suppress harmful rare failures while keeping rare successes in distributed RL."}],"snapshot_sha256":"bf742e9ddda694e45436a0d770a433553a93a785a56fbb59a0a899166000d6f3"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"abstract_excerpt":"Distributed reinforcement learning trains on data from stale, buggy, or mismatched actors, producing actions with high surprisal (negative log-probability) under the learner's policy. The core difficulty is not surprising data per se, but \\emph{negative learning from surprising data}. High-surprisal failures can dominate finite-batch updates through large perpendicular components, while high-surprisal successes reveal opportunities the current policy would otherwise miss. The \\textit{Delightful Policy Gradient} (DG) separates these cases by gating each update with delight, the product of advan","authors_text":"Ian Osband","cross_cats":["cs.AI","math.OC","stat.ML"],"headline":"Delightful policy gradient gates updates by advantage times surprisal to suppress harmful rare failures while keeping rare successes in distributed RL.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-20T21:45:51Z","title":"Delightful Distributed Policy Gradient"},"references":{"count":28,"internal_anchors":4,"resolved_work":28,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Preference optimization as probabilistic inference.arXiv e-prints, pages arXiv–2410, 2024","work_id":"71615a52-4d21-4dc1-a6c9-e88006c1bbae","year":2024},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Unifying count-based exploration and intrinsic motivation","work_id":"2fe189ce-b35d-4171-b8bf-a289fb2e3110","year":2016},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"Decision transformer: Reinforcement learning via sequence modeling","work_id":"a90a034a-ee4a-4779-927b-507900e52b5b","year":2021},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"First return, then explore.Nature, 590(7847):580–586, 2021","work_id":"6bcd818c-8c01-4c75-9915-3201f47c9839","year":2021},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"IMPALA: Scalable distributed deep-RL with importance weighted actor-learner architectures","work_id":"d50d1964-9261-4245-ae99-87a659d8cd5a","year":2018}],"snapshot_sha256":"e3d3ecb7dd411315f5219689f0adfac797575500b14f3cf2c4f03857af4e7b6c"},"source":{"id":"2603.20521","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-15T07:50:24.169905Z","id":"4c73da56-29d1-401f-9994-5d47cc2fe997","model_set":{"reader":"grok-4.3"},"one_line_summary":"Delightful Policy Gradient gates updates with advantage times surprisal to suppress rare failures while preserving rare successes in distributed RL with stale or buggy data.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Delightful policy gradient gates updates by advantage times surprisal to suppress harmful rare failures while keeping rare successes in distributed RL.","strongest_claim":"On MNIST with simulated staleness, DG without off-policy correction outperforms importance-weighted PG with exact behavior probabilities; on a transformer sequence task with staleness, actor bugs, reward corruption, and rare discovery, DG often achieves nearly order-of-magnitude lower error, with sample-efficiency advantage growing with task complexity.","weakest_assumption":"That the advantage sign is essential for the surprisal-based filter and that the tabular suppression of perpendicular second moment generalizes to the neural network experiments without additional corrections or assumptions about policy overlap."}},"verdict_id":"4c73da56-29d1-401f-9994-5d47cc2fe997"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:458ce79e98c222b0bbf3e7f4d5996acf515845c02c35c9c36bf8645a77c1a316","target":"record","created_at":"2026-05-18T02:45:04Z","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":"79eaba7752d70acf5a9440197c725b056453e35f9feb612e9d7b5250db9682a9","cross_cats_sorted":["cs.AI","math.OC","stat.ML"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-20T21:45:51Z","title_canon_sha256":"b5a4ed10115b97f5e0b14a6a1d2fa8282a51962e886c7331359a9191de8dacb8"},"schema_version":"1.0","source":{"id":"2603.20521","kind":"arxiv","version":2}},"canonical_sha256":"ec48a26b0572c3979e82481876d8c187ac4d48fb3eb8f4e8ecf5f90ea2cec608","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"ec48a26b0572c3979e82481876d8c187ac4d48fb3eb8f4e8ecf5f90ea2cec608","first_computed_at":"2026-05-18T02:45:04.689093Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T02:45:04.689093Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"y+BSrx+8N+ijdt4rRuPpQepx0wfdNY7tBLoTSxlrPe5FuNT/sK0mwB06GLgALrU/YXyYP+30Bm9LQgL/LMKECg==","signature_status":"signed_v1","signed_at":"2026-05-18T02:45:04.689629Z","signed_message":"canonical_sha256_bytes"},"source_id":"2603.20521","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:458ce79e98c222b0bbf3e7f4d5996acf515845c02c35c9c36bf8645a77c1a316","sha256:a1ce2be9f2e1c096657aa12c64bb8578ddb3955b1aba2cfe805252c5582e8abc"],"state_sha256":"04426cf19b4ed0f5ba8ad5f18a6f1d09c69b33cc6c156a0b3bfea2589d353ad4"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"goJMGT58Z3k469G8SgUNZEpPQVmOEha6MSHyBX04EXLyudfBTg+9vrUdAZF1uGleKkkDmlbBpK/bfn90XPLPCA==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-10T11:35:51.861153Z","bundle_sha256":"65dc32e84a2b781541d7d7abbacab2c6dee8623bfe20abb771f2de4833e3561b"}}