{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:SDCWQGR3PPYX37BEJWJGY4SMA7","short_pith_number":"pith:SDCWQGR3","schema_version":"1.0","canonical_sha256":"90c5681a3b7bf17dfc244d926c724c07ee235f85bfbadd4f55663bd12c709a0f","source":{"kind":"arxiv","id":"2502.01628","version":2},"attestation_state":"computed","paper":{"title":"Harmonic Loss Trains Interpretable AI Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"David D. Baek, Max Tegmark, Riya Tyagi, Ziming Liu","submitted_at":"2025-02-03T18:57:17Z","abstract_excerpt":"In this paper, we introduce harmonic loss as an alternative supervisory signal for training neural networks and large language models (LLMs). Harmonic loss differs from standard cross-entropy loss by (a) replacing the usual SoftMax normalization with a scale-invariant HarMax function and (b) computing logits via Euclidean distance rather than a dot product. Harmonic loss enables improved interpretability and faster convergence, owing to its scale invariance and finite convergence point by design, which can be interpreted as a class center. We first validate the performance of harmonic models a"},"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":"2502.01628","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-03T18:57:17Z","cross_cats_sorted":[],"title_canon_sha256":"1c1542dc07d653c93af39e9d027df6b7105861e5e9e77e7db425a15bde7374cb","abstract_canon_sha256":"b9db3442fa88994c15e66f1046e94f1593eff79d1087c4d93895dfe59be55c58"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:34:47.005689Z","signature_b64":"hSQDQcbDv9nbODwC8qCtYYH4hKR+g/S5p5zb/PTXUTQzlTvzJq/aaKYf45oiaB4LsWtK+yxleP9GAN5sutIqAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"90c5681a3b7bf17dfc244d926c724c07ee235f85bfbadd4f55663bd12c709a0f","last_reissued_at":"2026-07-05T11:34:47.005157Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:34:47.005157Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Harmonic Loss Trains Interpretable AI Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"David D. Baek, Max Tegmark, Riya Tyagi, Ziming Liu","submitted_at":"2025-02-03T18:57:17Z","abstract_excerpt":"In this paper, we introduce harmonic loss as an alternative supervisory signal for training neural networks and large language models (LLMs). Harmonic loss differs from standard cross-entropy loss by (a) replacing the usual SoftMax normalization with a scale-invariant HarMax function and (b) computing logits via Euclidean distance rather than a dot product. Harmonic loss enables improved interpretability and faster convergence, owing to its scale invariance and finite convergence point by design, which can be interpreted as a class center. We first validate the performance of harmonic models a"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.01628","kind":"arxiv","version":2},"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/2502.01628/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":"2502.01628","created_at":"2026-07-05T11:34:47.005221+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.01628v2","created_at":"2026-07-05T11:34:47.005221+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.01628","created_at":"2026-07-05T11:34:47.005221+00:00"},{"alias_kind":"pith_short_12","alias_value":"SDCWQGR3PPYX","created_at":"2026-07-05T11:34:47.005221+00:00"},{"alias_kind":"pith_short_16","alias_value":"SDCWQGR3PPYX37BE","created_at":"2026-07-05T11:34:47.005221+00:00"},{"alias_kind":"pith_short_8","alias_value":"SDCWQGR3","created_at":"2026-07-05T11:34:47.005221+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2504.16929","citing_title":"I-Con: A Unifying Framework for Representation Learning","ref_index":2019,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/SDCWQGR3PPYX37BEJWJGY4SMA7","json":"https://pith.science/pith/SDCWQGR3PPYX37BEJWJGY4SMA7.json","graph_json":"https://pith.science/api/pith-number/SDCWQGR3PPYX37BEJWJGY4SMA7/graph.json","events_json":"https://pith.science/api/pith-number/SDCWQGR3PPYX37BEJWJGY4SMA7/events.json","paper":"https://pith.science/paper/SDCWQGR3"},"agent_actions":{"view_html":"https://pith.science/pith/SDCWQGR3PPYX37BEJWJGY4SMA7","download_json":"https://pith.science/pith/SDCWQGR3PPYX37BEJWJGY4SMA7.json","view_paper":"https://pith.science/paper/SDCWQGR3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.01628&json=true","fetch_graph":"https://pith.science/api/pith-number/SDCWQGR3PPYX37BEJWJGY4SMA7/graph.json","fetch_events":"https://pith.science/api/pith-number/SDCWQGR3PPYX37BEJWJGY4SMA7/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/SDCWQGR3PPYX37BEJWJGY4SMA7/action/timestamp_anchor","attest_storage":"https://pith.science/pith/SDCWQGR3PPYX37BEJWJGY4SMA7/action/storage_attestation","attest_author":"https://pith.science/pith/SDCWQGR3PPYX37BEJWJGY4SMA7/action/author_attestation","sign_citation":"https://pith.science/pith/SDCWQGR3PPYX37BEJWJGY4SMA7/action/citation_signature","submit_replication":"https://pith.science/pith/SDCWQGR3PPYX37BEJWJGY4SMA7/action/replication_record"}},"created_at":"2026-07-05T11:34:47.005221+00:00","updated_at":"2026-07-05T11:34:47.005221+00:00"}