{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:KDMIU4AROC7ML2AZBRZZMZYCAC","short_pith_number":"pith:KDMIU4AR","schema_version":"1.0","canonical_sha256":"50d88a701170bec5e8190c7396670200856a5ae6a2df5dc15ab2b1e4f75baa82","source":{"kind":"arxiv","id":"2605.28918","version":1},"attestation_state":"computed","paper":{"title":"When LLM Reward Design Fails: Diagnostic-Driven Refinement for Sparse Structured RL","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.IR"],"primary_cat":"cs.LG","authors_text":"Bowen Liu, Dingyan Shang, Xuan Liu, Youting Wang, Yuan Tang","submitted_at":"2026-05-27T17:57:43Z","abstract_excerpt":"For sparse, structured reinforcement-learning tasks with semantic reward-function interfaces, LLM-generated reward shaping is better framed as debugging than one-shot generation. We study PPO-trained agents using MiniGrid as core evaluation and MuJoCo as boundary stress test. Our audit finds two dominant one-shot failure modes -- reward flooding and semantic/API misunderstanding -- plus a rarer weak-shaping case. We propose diagnostic-driven iterative refinement, where training diagnostics and a failure-mode taxonomy guide targeted reward-function revision. Refinement improves DoorKey-8x8 from"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.28918","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-27T17:57:43Z","cross_cats_sorted":["cs.IR"],"title_canon_sha256":"f540683a55d2fd48830c428694a4fc56de7f9bf3beb4fab7eb5006681e2f3c25","abstract_canon_sha256":"8f9d091625e0debafd3b8bdc0003c8c4e725b67b48b299974331c2fd6cf2a1a0"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-29T00:04:16.529002Z","signature_b64":"5gTQaIffv+ak0QF4xmKdaSzLRBIBMKtXci5rHS1+E7aZc6jJbRFssAN7dbIWuCkSSR8mHa0j3fEWE8vAplpYBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"50d88a701170bec5e8190c7396670200856a5ae6a2df5dc15ab2b1e4f75baa82","last_reissued_at":"2026-05-29T00:04:16.528510Z","signature_status":"signed_v1","first_computed_at":"2026-05-29T00:04:16.528510Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"When LLM Reward Design Fails: Diagnostic-Driven Refinement for Sparse Structured RL","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.IR"],"primary_cat":"cs.LG","authors_text":"Bowen Liu, Dingyan Shang, Xuan Liu, Youting Wang, Yuan Tang","submitted_at":"2026-05-27T17:57:43Z","abstract_excerpt":"For sparse, structured reinforcement-learning tasks with semantic reward-function interfaces, LLM-generated reward shaping is better framed as debugging than one-shot generation. We study PPO-trained agents using MiniGrid as core evaluation and MuJoCo as boundary stress test. Our audit finds two dominant one-shot failure modes -- reward flooding and semantic/API misunderstanding -- plus a rarer weak-shaping case. We propose diagnostic-driven iterative refinement, where training diagnostics and a failure-mode taxonomy guide targeted reward-function revision. Refinement improves DoorKey-8x8 from"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.28918","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.28918/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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.28918","created_at":"2026-05-29T00:04:16.528582+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.28918v1","created_at":"2026-05-29T00:04:16.528582+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.28918","created_at":"2026-05-29T00:04:16.528582+00:00"},{"alias_kind":"pith_short_12","alias_value":"KDMIU4AROC7M","created_at":"2026-05-29T00:04:16.528582+00:00"},{"alias_kind":"pith_short_16","alias_value":"KDMIU4AROC7ML2AZ","created_at":"2026-05-29T00:04:16.528582+00:00"},{"alias_kind":"pith_short_8","alias_value":"KDMIU4AR","created_at":"2026-05-29T00:04:16.528582+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/KDMIU4AROC7ML2AZBRZZMZYCAC","json":"https://pith.science/pith/KDMIU4AROC7ML2AZBRZZMZYCAC.json","graph_json":"https://pith.science/api/pith-number/KDMIU4AROC7ML2AZBRZZMZYCAC/graph.json","events_json":"https://pith.science/api/pith-number/KDMIU4AROC7ML2AZBRZZMZYCAC/events.json","paper":"https://pith.science/paper/KDMIU4AR"},"agent_actions":{"view_html":"https://pith.science/pith/KDMIU4AROC7ML2AZBRZZMZYCAC","download_json":"https://pith.science/pith/KDMIU4AROC7ML2AZBRZZMZYCAC.json","view_paper":"https://pith.science/paper/KDMIU4AR","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.28918&json=true","fetch_graph":"https://pith.science/api/pith-number/KDMIU4AROC7ML2AZBRZZMZYCAC/graph.json","fetch_events":"https://pith.science/api/pith-number/KDMIU4AROC7ML2AZBRZZMZYCAC/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/KDMIU4AROC7ML2AZBRZZMZYCAC/action/timestamp_anchor","attest_storage":"https://pith.science/pith/KDMIU4AROC7ML2AZBRZZMZYCAC/action/storage_attestation","attest_author":"https://pith.science/pith/KDMIU4AROC7ML2AZBRZZMZYCAC/action/author_attestation","sign_citation":"https://pith.science/pith/KDMIU4AROC7ML2AZBRZZMZYCAC/action/citation_signature","submit_replication":"https://pith.science/pith/KDMIU4AROC7ML2AZBRZZMZYCAC/action/replication_record"}},"created_at":"2026-05-29T00:04:16.528582+00:00","updated_at":"2026-05-29T00:04:16.528582+00:00"}