{"paper":{"title":"TopoGeoScore: A Self-Supervised Source-Only Geometric Framework for OOD Checkpoint Selection","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Source embeddings encode global, local, and topological signals that identify which checkpoints will remain accurate under distribution shift.","cross_cats":["math.AT","math.DG"],"primary_cat":"cs.LG","authors_text":"Ali Zia, Farid Hazratian, Hien Duy Nguyen","submitted_at":"2026-05-09T10:46:47Z","abstract_excerpt":"Out-of-distribution (OOD) robustness is difficult to diagnose when target-domain labels are unavailable. We consider a more restrictive source-only variant of unsupervised accuracy estimation: selecting robust checkpoints using only source-domain representations, with no target samples or target labels. We propose \\textbf{TopoGeoScore}, a source-only geometric scorer for label-free OOD checkpoint selection. Given a trained checkpoint, we construct class-conditional mutual $k$-nearest-neighbour graphs from source embeddings and extract three interpretable signals: a torsion-inspired reduced Lap"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"source representations contain measurable global--local--topological evidence of robustness, supporting practical checkpoint selection before deployment under distribution shift.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the self-supervised objective, which enforces invariance under approximately geometry-preserving embedding views, selects for actual OOD robustness rather than some other incidental property of the source embeddings.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"TopoGeoScore combines a torsion-inspired Laplacian log-determinant, Ollivier-Ricci curvature, and higher-order topological summaries from source embeddings, with weights learned via self-supervised invariance to geometry-preserving views, to rank checkpoints by expected OOD robustness.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Source embeddings encode global, local, and topological signals that identify which checkpoints will remain accurate under distribution shift.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"41f81a5e70d02d699ca8dcf7a139eac016a49f7f60649f7ab3a51566525bdfd5"},"source":{"id":"2605.08870","kind":"arxiv","version":2},"verdict":{"id":"de77a6b1-491b-4abb-b7cb-f8d3be03c3a8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T01:03:55.107913Z","strongest_claim":"source representations contain measurable global--local--topological evidence of robustness, supporting practical checkpoint selection before deployment under distribution shift.","one_line_summary":"TopoGeoScore combines a torsion-inspired Laplacian log-determinant, Ollivier-Ricci curvature, and higher-order topological summaries from source embeddings, with weights learned via self-supervised invariance to geometry-preserving views, to rank checkpoints by expected OOD robustness.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the self-supervised objective, which enforces invariance under approximately geometry-preserving embedding views, selects for actual OOD robustness rather than some other incidental property of the source embeddings.","pith_extraction_headline":"Source embeddings encode global, local, and topological signals that identify which checkpoints will remain accurate under distribution shift."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.08870/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T08:42:01.989578Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T22:33:28.366938Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T14:01:21.631280Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T10:42:10.824671Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"d11f7e741dc1c899134003a92a57859fa128600b71cb11c48d2bc84280f7e52e"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":1,"snapshot_sha256":"a21244e7129561315180ff2f86ed675376c75e88290f797b2d087d83b407076c"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}