{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2020:EN77RGPB7JSUOILWPYXWW3R6ZJ","short_pith_number":"pith:EN77RGPB","schema_version":"1.0","canonical_sha256":"237ff899e1fa654721767e2f6b6e3eca79c802b72ae4cfb7beac16bb7f39a0d7","source":{"kind":"arxiv","id":"2012.05862","version":1},"attestation_state":"computed","paper":{"title":"Understanding Learned Reward Functions","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Adam Gleave, Eric J. Michaud, Stuart Russell","submitted_at":"2020-12-10T18:19:48Z","abstract_excerpt":"In many real-world tasks, it is not possible to procedurally specify an RL agent's reward function. In such cases, a reward function must instead be learned from interacting with and observing humans. However, current techniques for reward learning may fail to produce reward functions which accurately reflect user preferences. Absent significant advances in reward learning, it is thus important to be able to audit learned reward functions to verify whether they truly capture user preferences. In this paper, we investigate techniques for interpreting learned reward functions. In particular, we "},"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":"2012.05862","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2020-12-10T18:19:48Z","cross_cats_sorted":[],"title_canon_sha256":"07b378047c3e246410c4c94cb14b2c4b3ff139c3b088cec31f826c79f8aee26c","abstract_canon_sha256":"e8a4d15e2ebad6bc2fb3aace995c1cbc9d1dafd01d1f2f3e62939fd576e87756"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T01:58:37.876144Z","signature_b64":"Mnt5Yd2bKAtkvoPBJIAypvLbAssXmZnlQTWW8vvsfkvx+4s1zVacCoBQh6M5U/UdN6sXDy+bCTzMTk2pIeB7Cg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"237ff899e1fa654721767e2f6b6e3eca79c802b72ae4cfb7beac16bb7f39a0d7","last_reissued_at":"2026-07-05T01:58:37.875781Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T01:58:37.875781Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Understanding Learned Reward Functions","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Adam Gleave, Eric J. Michaud, Stuart Russell","submitted_at":"2020-12-10T18:19:48Z","abstract_excerpt":"In many real-world tasks, it is not possible to procedurally specify an RL agent's reward function. In such cases, a reward function must instead be learned from interacting with and observing humans. However, current techniques for reward learning may fail to produce reward functions which accurately reflect user preferences. Absent significant advances in reward learning, it is thus important to be able to audit learned reward functions to verify whether they truly capture user preferences. In this paper, we investigate techniques for interpreting learned reward functions. In particular, we "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2012.05862","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/2012.05862/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":"2012.05862","created_at":"2026-07-05T01:58:37.875837+00:00"},{"alias_kind":"arxiv_version","alias_value":"2012.05862v1","created_at":"2026-07-05T01:58:37.875837+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2012.05862","created_at":"2026-07-05T01:58:37.875837+00:00"},{"alias_kind":"pith_short_12","alias_value":"EN77RGPB7JSU","created_at":"2026-07-05T01:58:37.875837+00:00"},{"alias_kind":"pith_short_16","alias_value":"EN77RGPB7JSUOILW","created_at":"2026-07-05T01:58:37.875837+00:00"},{"alias_kind":"pith_short_8","alias_value":"EN77RGPB","created_at":"2026-07-05T01:58:37.875837+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2310.15288","citing_title":"Active teacher selection for reward learning","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2509.03403","citing_title":"Beyond Correctness: Harmonizing Process and Outcome Rewards through RL Training","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2304.06767","citing_title":"RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment","ref_index":127,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/EN77RGPB7JSUOILWPYXWW3R6ZJ","json":"https://pith.science/pith/EN77RGPB7JSUOILWPYXWW3R6ZJ.json","graph_json":"https://pith.science/api/pith-number/EN77RGPB7JSUOILWPYXWW3R6ZJ/graph.json","events_json":"https://pith.science/api/pith-number/EN77RGPB7JSUOILWPYXWW3R6ZJ/events.json","paper":"https://pith.science/paper/EN77RGPB"},"agent_actions":{"view_html":"https://pith.science/pith/EN77RGPB7JSUOILWPYXWW3R6ZJ","download_json":"https://pith.science/pith/EN77RGPB7JSUOILWPYXWW3R6ZJ.json","view_paper":"https://pith.science/paper/EN77RGPB","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2012.05862&json=true","fetch_graph":"https://pith.science/api/pith-number/EN77RGPB7JSUOILWPYXWW3R6ZJ/graph.json","fetch_events":"https://pith.science/api/pith-number/EN77RGPB7JSUOILWPYXWW3R6ZJ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/EN77RGPB7JSUOILWPYXWW3R6ZJ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/EN77RGPB7JSUOILWPYXWW3R6ZJ/action/storage_attestation","attest_author":"https://pith.science/pith/EN77RGPB7JSUOILWPYXWW3R6ZJ/action/author_attestation","sign_citation":"https://pith.science/pith/EN77RGPB7JSUOILWPYXWW3R6ZJ/action/citation_signature","submit_replication":"https://pith.science/pith/EN77RGPB7JSUOILWPYXWW3R6ZJ/action/replication_record"}},"created_at":"2026-07-05T01:58:37.875837+00:00","updated_at":"2026-07-05T01:58:37.875837+00:00"}