{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:OGZLHNURTYEF2GKLCX47BDMA6Z","short_pith_number":"pith:OGZLHNUR","schema_version":"1.0","canonical_sha256":"71b2b3b6919e085d194b15f9f08d80f6479e6192cd5cfd3757ee179d56986a4f","source":{"kind":"arxiv","id":"2604.09414","version":3},"attestation_state":"computed","paper":{"title":"Beyond Augmented-Action Surrogates for Multi-Expert Learning-to-Defer","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A decoupled surrogate for multi-expert deferral separates class posteriors from expert utilities to eliminate gradient pathologies.","cross_cats":["cs.LG"],"primary_cat":"stat.ML","authors_text":"Axel Carlier, Lai Xing Ng, Wei Tsang Ooi, Yannis Montreuil","submitted_at":"2026-04-10T15:27:23Z","abstract_excerpt":"A learning-to-defer (L2D) system decides, for each input, whether to predict on its own or to hand it to one of several available experts. The very well established recipe trains classifier and router jointly by treating the $K$ classes and $J$ experts as competing actions in one shared $(K{+}J)$-action geometry. Subsequent work has proposed a series of incremental fixes within this geometry; we show that each still suffers, to varying severity, from an optimization-level pathology (target distortion, gradient amplification, winner-take-all starvation, set-mass collapse, or class--expert coupl"},"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":"2604.09414","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"stat.ML","submitted_at":"2026-04-10T15:27:23Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"83f2b28c8f52052fc2db313cd65d563c2e06a269909b5cae66d090668fd80fd0","abstract_canon_sha256":"2e3effd8e43beae8df323db019730c924c99af1ecc353a6751b8328185d08c8a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-21T01:05:18.755089Z","signature_b64":"l4pISyEHMzT0nkB+wLMDWKagXJo8iicUdR95iy3VyIjDqfXlM8xwKtnkX90TgzH1Q9TxvXDVQ/FyXuLbLtc+Dw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"71b2b3b6919e085d194b15f9f08d80f6479e6192cd5cfd3757ee179d56986a4f","last_reissued_at":"2026-05-21T01:05:18.754410Z","signature_status":"signed_v1","first_computed_at":"2026-05-21T01:05:18.754410Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Beyond Augmented-Action Surrogates for Multi-Expert Learning-to-Defer","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A decoupled surrogate for multi-expert deferral separates class posteriors from expert utilities to eliminate gradient pathologies.","cross_cats":["cs.LG"],"primary_cat":"stat.ML","authors_text":"Axel Carlier, Lai Xing Ng, Wei Tsang Ooi, Yannis Montreuil","submitted_at":"2026-04-10T15:27:23Z","abstract_excerpt":"A learning-to-defer (L2D) system decides, for each input, whether to predict on its own or to hand it to one of several available experts. The very well established recipe trains classifier and router jointly by treating the $K$ classes and $J$ experts as competing actions in one shared $(K{+}J)$-action geometry. Subsequent work has proposed a series of incremental fixes within this geometry; we show that each still suffers, to varying severity, from an optimization-level pathology (target distortion, gradient amplification, winner-take-all starvation, set-mass collapse, or class--expert coupl"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"The decoupled surrogate is the only method that avoids amplification under redundancy, preserves rare specialists, and consistently improves over a standalone classifier across all settings.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That separating class posterior estimation (softmax) from expert utility estimation (independent sigmoids) removes the amplification, starvation, and coupling pathologies without introducing new failure modes under realistic expert correlation structures.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A decoupled surrogate separates class posterior estimation from per-expert utility estimation, yielding a J-independent H-consistency bound and avoiding the amplification, starvation, and coupling issues of prior augmented-action surrogates.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A decoupled surrogate for multi-expert deferral separates class posteriors from expert utilities to eliminate gradient pathologies.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"76e26b7761ff5928528b11873060af77b2fea16410162feb3ef856d04191ea23"},"source":{"id":"2604.09414","kind":"arxiv","version":3},"verdict":{"id":"eb4bd779-9319-4528-90b7-09f3ac8d31d7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T16:36:12.150652Z","strongest_claim":"The decoupled surrogate is the only method that avoids amplification under redundancy, preserves rare specialists, and consistently improves over a standalone classifier across all settings.","one_line_summary":"A decoupled surrogate separates class posterior estimation from per-expert utility estimation, yielding a J-independent H-consistency bound and avoiding the amplification, starvation, and coupling issues of prior augmented-action surrogates.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That separating class posterior estimation (softmax) from expert utility estimation (independent sigmoids) removes the amplification, starvation, and coupling pathologies without introducing new failure modes under realistic expert correlation structures.","pith_extraction_headline":"A decoupled surrogate for multi-expert deferral separates class posteriors from expert utilities to eliminate gradient pathologies."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.09414/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":"2604.09414","created_at":"2026-05-21T01:05:18.754485+00:00"},{"alias_kind":"arxiv_version","alias_value":"2604.09414v3","created_at":"2026-05-21T01:05:18.754485+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.09414","created_at":"2026-05-21T01:05:18.754485+00:00"},{"alias_kind":"pith_short_12","alias_value":"OGZLHNURTYEF","created_at":"2026-05-21T01:05:18.754485+00:00"},{"alias_kind":"pith_short_16","alias_value":"OGZLHNURTYEF2GKL","created_at":"2026-05-21T01:05:18.754485+00:00"},{"alias_kind":"pith_short_8","alias_value":"OGZLHNUR","created_at":"2026-05-21T01:05:18.754485+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2605.12340","citing_title":"Online Learning-to-Defer with Varying Experts","ref_index":170,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OGZLHNURTYEF2GKLCX47BDMA6Z","json":"https://pith.science/pith/OGZLHNURTYEF2GKLCX47BDMA6Z.json","graph_json":"https://pith.science/api/pith-number/OGZLHNURTYEF2GKLCX47BDMA6Z/graph.json","events_json":"https://pith.science/api/pith-number/OGZLHNURTYEF2GKLCX47BDMA6Z/events.json","paper":"https://pith.science/paper/OGZLHNUR"},"agent_actions":{"view_html":"https://pith.science/pith/OGZLHNURTYEF2GKLCX47BDMA6Z","download_json":"https://pith.science/pith/OGZLHNURTYEF2GKLCX47BDMA6Z.json","view_paper":"https://pith.science/paper/OGZLHNUR","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2604.09414&json=true","fetch_graph":"https://pith.science/api/pith-number/OGZLHNURTYEF2GKLCX47BDMA6Z/graph.json","fetch_events":"https://pith.science/api/pith-number/OGZLHNURTYEF2GKLCX47BDMA6Z/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OGZLHNURTYEF2GKLCX47BDMA6Z/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OGZLHNURTYEF2GKLCX47BDMA6Z/action/storage_attestation","attest_author":"https://pith.science/pith/OGZLHNURTYEF2GKLCX47BDMA6Z/action/author_attestation","sign_citation":"https://pith.science/pith/OGZLHNURTYEF2GKLCX47BDMA6Z/action/citation_signature","submit_replication":"https://pith.science/pith/OGZLHNURTYEF2GKLCX47BDMA6Z/action/replication_record"}},"created_at":"2026-05-21T01:05:18.754485+00:00","updated_at":"2026-05-21T01:05:18.754485+00:00"}