{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:G5CNCAZNUCMSIHHPL2NYJGFIBI","short_pith_number":"pith:G5CNCAZN","canonical_record":{"source":{"id":"2604.18109","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-04-20T11:27:14Z","cross_cats_sorted":["cs.SD"],"title_canon_sha256":"138ba11b5117c03bfe7b7fa904510948da496a396a4ce53740c3d73335777772","abstract_canon_sha256":"0e89b29cdfa4a27602d82759503e798403a1d9d5a54a2c98914e6c3d1ef1eb09"},"schema_version":"1.0"},"canonical_sha256":"3744d1032da099241cef5e9b8498a80a24097823ee2d71a4ba463078a98d5b55","source":{"kind":"arxiv","id":"2604.18109","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2604.18109","created_at":"2026-06-19T16:11:23Z"},{"alias_kind":"arxiv_version","alias_value":"2604.18109v2","created_at":"2026-06-19T16:11:23Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.18109","created_at":"2026-06-19T16:11:23Z"},{"alias_kind":"pith_short_12","alias_value":"G5CNCAZNUCMS","created_at":"2026-06-19T16:11:23Z"},{"alias_kind":"pith_short_16","alias_value":"G5CNCAZNUCMSIHHP","created_at":"2026-06-19T16:11:23Z"},{"alias_kind":"pith_short_8","alias_value":"G5CNCAZN","created_at":"2026-06-19T16:11:23Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:G5CNCAZNUCMSIHHPL2NYJGFIBI","target":"record","payload":{"canonical_record":{"source":{"id":"2604.18109","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-04-20T11:27:14Z","cross_cats_sorted":["cs.SD"],"title_canon_sha256":"138ba11b5117c03bfe7b7fa904510948da496a396a4ce53740c3d73335777772","abstract_canon_sha256":"0e89b29cdfa4a27602d82759503e798403a1d9d5a54a2c98914e6c3d1ef1eb09"},"schema_version":"1.0"},"canonical_sha256":"3744d1032da099241cef5e9b8498a80a24097823ee2d71a4ba463078a98d5b55","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-19T16:11:23.236420Z","signature_b64":"+o4OGlgBNfygFc+LN2ZvwrGqZuu6gJjBrFic4aH2k9xodROWXhbjlR1kzyzVA+bd58pSayZ8FoE7xQoSwHVgBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3744d1032da099241cef5e9b8498a80a24097823ee2d71a4ba463078a98d5b55","last_reissued_at":"2026-06-19T16:11:23.236018Z","signature_status":"signed_v1","first_computed_at":"2026-06-19T16:11:23.236018Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2604.18109","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-19T16:11:23Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"RYfIE6us+RqQPJxZbwNHnHp3K7eHa+ChGrSuIdw5I5BDXQgDYRHseXmXnSAy5E+V6Lpse9CEKDdps832B6T7Aw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-20T18:39:32.723544Z"},"content_sha256":"20e7a89283fb8d8227149e25aa2485cf250e36f4ce1c85ea815d9d2110f07881","schema_version":"1.0","event_id":"sha256:20e7a89283fb8d8227149e25aa2485cf250e36f4ce1c85ea815d9d2110f07881"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:G5CNCAZNUCMSIHHPL2NYJGFIBI","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"FLiP: Towards understanding and interpreting multimodal multilingual sentence embeddings","license":"http://creativecommons.org/licenses/by/4.0/","headline":"FLiP recovers over 75% of lexical content from various sentence embeddings to diagnose biases.","cross_cats":["cs.SD"],"primary_cat":"cs.CL","authors_text":"Bolaji Yusuf, Old\\v{r}ich Plchot, Petr Schwarz, Santosh Kesiraju, \\v{S}imon Sedl\\'a\\v{c}ek","submitted_at":"2026-04-20T11:27:14Z","abstract_excerpt":"This paper presents factorized linear projection (FLiP) models for understanding pretrained sentence embedding spaces. We train FLiP models to recover the lexical content from multilingual (LaBSE), multimodal (SONAR) and API-based (Gemini) sentence embedding spaces in several high- and mid-resource languages. We show that FLiP can recall more than 75% of lexical content from the embeddings, significantly outperforming existing non-factorized baselines. Using this as a diagnostic tool, we uncover the modality and language biases across the selected sentence encoders and provide practitioners wi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We show that FLiP can recall more than 75% of lexical content from the embeddings, significantly outperforming existing non-factorized baselines.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That recovering lexical content via trained linear projections provides a faithful and unbiased diagnostic of the information actually present in the original sentence embeddings.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"FLiP recovers more than 75% lexical content from pretrained sentence embeddings across languages and modalities, outperforming non-factorized baselines and exposing intrinsic biases.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"FLiP recovers over 75% of lexical content from various sentence embeddings to diagnose biases.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"71b65dc9dd7181890a5256832a271bfa3fd04ec0825345e0c418a2a5b98bb2d8"},"source":{"id":"2604.18109","kind":"arxiv","version":2},"verdict":{"id":"3009d9bf-6f4b-4f34-9ce3-d3ede66fcd0a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T05:18:50.417185Z","strongest_claim":"We show that FLiP can recall more than 75% of lexical content from the embeddings, significantly outperforming existing non-factorized baselines.","one_line_summary":"FLiP recovers more than 75% lexical content from pretrained sentence embeddings across languages and modalities, outperforming non-factorized baselines and exposing intrinsic biases.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That recovering lexical content via trained linear projections provides a faithful and unbiased diagnostic of the information actually present in the original sentence embeddings.","pith_extraction_headline":"FLiP recovers over 75% of lexical content from various sentence embeddings to diagnose biases."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.18109/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-20T04:25:57.455004Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"b35a2884f748f19d2de14995575533febb59c7a0b95d7f591728477dd9a4a08e"},"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"},"verdict_id":"3009d9bf-6f4b-4f34-9ce3-d3ede66fcd0a"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-19T16:11:23Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"AfQaSz+tEdh01UgQaV32WiKWMg7dKtD74JTWagKG2HnBUjRC23jgiNaMDQHPJ2cKITcWWrk+1HaHNg2E6mgsDQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-20T18:39:32.724393Z"},"content_sha256":"1aefc505e0f0ab5b710aa4a005c3a581ee3ff43510d09025e045f07bb821c33a","schema_version":"1.0","event_id":"sha256:1aefc505e0f0ab5b710aa4a005c3a581ee3ff43510d09025e045f07bb821c33a"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/G5CNCAZNUCMSIHHPL2NYJGFIBI/bundle.json","state_url":"https://pith.science/pith/G5CNCAZNUCMSIHHPL2NYJGFIBI/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/G5CNCAZNUCMSIHHPL2NYJGFIBI/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-08-20T18:39:32Z","links":{"resolver":"https://pith.science/pith/G5CNCAZNUCMSIHHPL2NYJGFIBI","bundle":"https://pith.science/pith/G5CNCAZNUCMSIHHPL2NYJGFIBI/bundle.json","state":"https://pith.science/pith/G5CNCAZNUCMSIHHPL2NYJGFIBI/state.json","well_known_bundle":"https://pith.science/.well-known/pith/G5CNCAZNUCMSIHHPL2NYJGFIBI/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:G5CNCAZNUCMSIHHPL2NYJGFIBI","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"0e89b29cdfa4a27602d82759503e798403a1d9d5a54a2c98914e6c3d1ef1eb09","cross_cats_sorted":["cs.SD"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-04-20T11:27:14Z","title_canon_sha256":"138ba11b5117c03bfe7b7fa904510948da496a396a4ce53740c3d73335777772"},"schema_version":"1.0","source":{"id":"2604.18109","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2604.18109","created_at":"2026-06-19T16:11:23Z"},{"alias_kind":"arxiv_version","alias_value":"2604.18109v2","created_at":"2026-06-19T16:11:23Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.18109","created_at":"2026-06-19T16:11:23Z"},{"alias_kind":"pith_short_12","alias_value":"G5CNCAZNUCMS","created_at":"2026-06-19T16:11:23Z"},{"alias_kind":"pith_short_16","alias_value":"G5CNCAZNUCMSIHHP","created_at":"2026-06-19T16:11:23Z"},{"alias_kind":"pith_short_8","alias_value":"G5CNCAZN","created_at":"2026-06-19T16:11:23Z"}],"graph_snapshots":[{"event_id":"sha256:1aefc505e0f0ab5b710aa4a005c3a581ee3ff43510d09025e045f07bb821c33a","target":"graph","created_at":"2026-06-19T16:11:23Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"We show that FLiP can recall more than 75% of lexical content from the embeddings, significantly outperforming existing non-factorized baselines."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That recovering lexical content via trained linear projections provides a faithful and unbiased diagnostic of the information actually present in the original sentence embeddings."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"FLiP recovers more than 75% lexical content from pretrained sentence embeddings across languages and modalities, outperforming non-factorized baselines and exposing intrinsic biases."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"FLiP recovers over 75% of lexical content from various sentence embeddings to diagnose biases."}],"snapshot_sha256":"71b65dc9dd7181890a5256832a271bfa3fd04ec0825345e0c418a2a5b98bb2d8"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[{"findings_count":0,"name":"doi_compliance","ran_at":"2026-05-20T04:25:57.455004Z","status":"completed","version":"1.0.0"}],"endpoint":"/pith/2604.18109/integrity.json","findings":[],"snapshot_sha256":"b35a2884f748f19d2de14995575533febb59c7a0b95d7f591728477dd9a4a08e","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"This paper presents factorized linear projection (FLiP) models for understanding pretrained sentence embedding spaces. We train FLiP models to recover the lexical content from multilingual (LaBSE), multimodal (SONAR) and API-based (Gemini) sentence embedding spaces in several high- and mid-resource languages. We show that FLiP can recall more than 75% of lexical content from the embeddings, significantly outperforming existing non-factorized baselines. Using this as a diagnostic tool, we uncover the modality and language biases across the selected sentence encoders and provide practitioners wi","authors_text":"Bolaji Yusuf, Old\\v{r}ich Plchot, Petr Schwarz, Santosh Kesiraju, \\v{S}imon Sedl\\'a\\v{c}ek","cross_cats":["cs.SD"],"headline":"FLiP recovers over 75% of lexical content from various sentence embeddings to diagnose biases.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-04-20T11:27:14Z","title":"FLiP: Towards understanding and interpreting multimodal multilingual sentence embeddings"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2604.18109","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-10T05:18:50.417185Z","id":"3009d9bf-6f4b-4f34-9ce3-d3ede66fcd0a","model_set":{"reader":"grok-4.3"},"one_line_summary":"FLiP recovers more than 75% lexical content from pretrained sentence embeddings across languages and modalities, outperforming non-factorized baselines and exposing intrinsic biases.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"FLiP recovers over 75% of lexical content from various sentence embeddings to diagnose biases.","strongest_claim":"We show that FLiP can recall more than 75% of lexical content from the embeddings, significantly outperforming existing non-factorized baselines.","weakest_assumption":"That recovering lexical content via trained linear projections provides a faithful and unbiased diagnostic of the information actually present in the original sentence embeddings."}},"verdict_id":"3009d9bf-6f4b-4f34-9ce3-d3ede66fcd0a"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:20e7a89283fb8d8227149e25aa2485cf250e36f4ce1c85ea815d9d2110f07881","target":"record","created_at":"2026-06-19T16:11:23Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"0e89b29cdfa4a27602d82759503e798403a1d9d5a54a2c98914e6c3d1ef1eb09","cross_cats_sorted":["cs.SD"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-04-20T11:27:14Z","title_canon_sha256":"138ba11b5117c03bfe7b7fa904510948da496a396a4ce53740c3d73335777772"},"schema_version":"1.0","source":{"id":"2604.18109","kind":"arxiv","version":2}},"canonical_sha256":"3744d1032da099241cef5e9b8498a80a24097823ee2d71a4ba463078a98d5b55","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"3744d1032da099241cef5e9b8498a80a24097823ee2d71a4ba463078a98d5b55","first_computed_at":"2026-06-19T16:11:23.236018Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-06-19T16:11:23.236018Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"+o4OGlgBNfygFc+LN2ZvwrGqZuu6gJjBrFic4aH2k9xodROWXhbjlR1kzyzVA+bd58pSayZ8FoE7xQoSwHVgBQ==","signature_status":"signed_v1","signed_at":"2026-06-19T16:11:23.236420Z","signed_message":"canonical_sha256_bytes"},"source_id":"2604.18109","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:20e7a89283fb8d8227149e25aa2485cf250e36f4ce1c85ea815d9d2110f07881","sha256:1aefc505e0f0ab5b710aa4a005c3a581ee3ff43510d09025e045f07bb821c33a"],"state_sha256":"deaa97a39e6a289b055e81eac1a40cc376a7da36de49a0e89e6c7e51578ce300"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"ByvcuvYbcHd4tRdBTeFEirq/qBMLEW5bLPZynwv/RHzL53md3yNI9xuaBSXzgBnb8Z+Cu/jVPhd7JlUpyKPoAA==","signed_message":"bundle_sha256_bytes","signed_at":"2026-08-20T18:39:32.730941Z","bundle_sha256":"1441fb82df0244e266c40776e7854dee89f1a89319696b766957d7af77a8d020"}}