{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:GAEAA7U3YUL3PO6NZNED5P3N25","short_pith_number":"pith:GAEAA7U3","canonical_record":{"source":{"id":"2603.24125","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-03-25T09:35:18Z","cross_cats_sorted":[],"title_canon_sha256":"b7263ca6adce8b93c8ff911975d92cb7539c432bc209b8473e80eff151c524d5","abstract_canon_sha256":"b358da891dc36defefb1874a0115d485852c0655cad89c05cd4e0975d8028b7f"},"schema_version":"1.0"},"canonical_sha256":"3008007e9bc517b7bbcdcb483ebf6dd764b99bf14feacbb25a5d358cdd6ae221","source":{"kind":"arxiv","id":"2603.24125","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2603.24125","created_at":"2026-05-18T03:10:03Z"},{"alias_kind":"arxiv_version","alias_value":"2603.24125v2","created_at":"2026-05-18T03:10:03Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.24125","created_at":"2026-05-18T03:10:03Z"},{"alias_kind":"pith_short_12","alias_value":"GAEAA7U3YUL3","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"GAEAA7U3YUL3PO6N","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"GAEAA7U3","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:GAEAA7U3YUL3PO6NZNED5P3N25","target":"record","payload":{"canonical_record":{"source":{"id":"2603.24125","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-03-25T09:35:18Z","cross_cats_sorted":[],"title_canon_sha256":"b7263ca6adce8b93c8ff911975d92cb7539c432bc209b8473e80eff151c524d5","abstract_canon_sha256":"b358da891dc36defefb1874a0115d485852c0655cad89c05cd4e0975d8028b7f"},"schema_version":"1.0"},"canonical_sha256":"3008007e9bc517b7bbcdcb483ebf6dd764b99bf14feacbb25a5d358cdd6ae221","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:10:03.407998Z","signature_b64":"fj0J1gTZ7aYm7ty4Lj52YrPIK20AQefHSW3qhflfHVdaSbvpbBN/F1I9kObIc/Z5E4FRx8BDNWwUie+5vi+ECw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3008007e9bc517b7bbcdcb483ebf6dd764b99bf14feacbb25a5d358cdd6ae221","last_reissued_at":"2026-05-18T03:10:03.407514Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:10:03.407514Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2603.24125","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":"2026-05-18T03:10:03Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"zb2dUyjUYlmTbJgps5UfNpIfw25Qo3kI8G2VFIr3YJ7Z2cD6KucF6V3q5R/+tvwRvG1y3bw0ZFzpIFXF4k+0Bg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-10T16:17:04.260271Z"},"content_sha256":"4dd3c0a3bd71686b6713619bdaf5c8b796cfe089151fa5e811b2ccf3d7abe4c4","schema_version":"1.0","event_id":"sha256:4dd3c0a3bd71686b6713619bdaf5c8b796cfe089151fa5e811b2ccf3d7abe4c4"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:GAEAA7U3YUL3PO6NZNED5P3N25","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Alignment Reduces Expressed but Not Encoded Gender Bias: A Unified Framework and Study","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Alignment reduces expressed gender bias in outputs but leaves measurable associations intact inside the model's representations.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Christophe Marsala, Marcin Detyniecki, Marie-Jeanne Lesot, Nour Bouchouchi, Thibault Laugel, Xavier Renard","submitted_at":"2026-03-25T09:35:18Z","abstract_excerpt":"During training, Large Language Models (LLMs) learn social regularities that can lead to gender bias in downstream applications. Most mitigation efforts focus on reducing bias in generated outputs, typically evaluated on structured benchmarks, which raises two concerns: output-level evaluation does not reveal whether alignment modifies the model's underlying representations, and structured benchmarks may not reflect realistic usage scenarios. We propose a unified framework to jointly analyze intrinsic and extrinsic gender bias in LLMs using identical neutral prompts, enabling direct comparison"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Contrary to prior work reporting weak or inconsistent correlations, we find a consistent association between latent gender information and expressed bias when measured under the unified protocol. [...] while the latter indeed reduces expressed bias, measurable gender-related associations are still present in internal representations, and can be reactivated under adversarial prompting.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the chosen neutral prompts and the specific methods for extracting latent gender information from internal representations accurately reflect the model's true encoded associations without being artifacts of prompt design or measurement choices.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Alignment reduces expressed gender bias in LLM outputs but does not remove the underlying encoded gender associations in internal representations.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Alignment reduces expressed gender bias in outputs but leaves measurable associations intact inside the model's representations.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"938acf540ba51d619c72fe2d9ea867745f110ab88966928f891218c32f28aabf"},"source":{"id":"2603.24125","kind":"arxiv","version":2},"verdict":{"id":"646e7624-6d4a-4f1b-adf2-02377bf94163","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T00:36:37.616118Z","strongest_claim":"Contrary to prior work reporting weak or inconsistent correlations, we find a consistent association between latent gender information and expressed bias when measured under the unified protocol. [...] while the latter indeed reduces expressed bias, measurable gender-related associations are still present in internal representations, and can be reactivated under adversarial prompting.","one_line_summary":"Alignment reduces expressed gender bias in LLM outputs but does not remove the underlying encoded gender associations in internal representations.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the chosen neutral prompts and the specific methods for extracting latent gender information from internal representations accurately reflect the model's true encoded associations without being artifacts of prompt design or measurement choices.","pith_extraction_headline":"Alignment reduces expressed gender bias in outputs but leaves measurable associations intact inside the model's representations."},"references":{"count":36,"sample":[{"doi":"","year":2024,"title":"In: NeurIPS (2024)","work_id":"17ac99e9-682f-4120-b540-f8f88a020ea4","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"In: NeurIPS (2016)","work_id":"b724b495-fd98-449b-9e8d-0bd784a36ef0","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Caliskan, A., Bryson, J.J., Narayanan, A.: Semantics derived automatically from language corpora contain human-like biases. Science356, 183–186 (2017)","work_id":"78b8f9aa-f16f-47f6-93a0-273f216c6eb6","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"In: ACL (2022)","work_id":"4a193780-15a1-4dbc-887c-3d061351c7f1","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2013,"title":"Psychology of sport and exercise 14, 136–144 (2013)","work_id":"d5769729-eae1-4296-aa39-85da84b349aa","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":36,"snapshot_sha256":"797286a345d43ff7182df78e7c5292858c6acbfc4eb1e87a92115fd15c98c5d5","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":"646e7624-6d4a-4f1b-adf2-02377bf94163"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T03:10:03Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"dmCX36ZquCzJJ0Gi01IIK5LVHXOPNnCwD95Z4bWMJrUyOE8VML5SyrEB2aKxymNF5jQpMrRhcvyWh5H6fr4cDA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-10T16:17:04.261600Z"},"content_sha256":"e319670f9572dc3d34e3980b4f8c5b48d32f8cc7fe4d80809f6728c0973b7b7a","schema_version":"1.0","event_id":"sha256:e319670f9572dc3d34e3980b4f8c5b48d32f8cc7fe4d80809f6728c0973b7b7a"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/GAEAA7U3YUL3PO6NZNED5P3N25/bundle.json","state_url":"https://pith.science/pith/GAEAA7U3YUL3PO6NZNED5P3N25/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/GAEAA7U3YUL3PO6NZNED5P3N25/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":"2026-06-10T16:17:04Z","links":{"resolver":"https://pith.science/pith/GAEAA7U3YUL3PO6NZNED5P3N25","bundle":"https://pith.science/pith/GAEAA7U3YUL3PO6NZNED5P3N25/bundle.json","state":"https://pith.science/pith/GAEAA7U3YUL3PO6NZNED5P3N25/state.json","well_known_bundle":"https://pith.science/.well-known/pith/GAEAA7U3YUL3PO6NZNED5P3N25/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:GAEAA7U3YUL3PO6NZNED5P3N25","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":"b358da891dc36defefb1874a0115d485852c0655cad89c05cd4e0975d8028b7f","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-03-25T09:35:18Z","title_canon_sha256":"b7263ca6adce8b93c8ff911975d92cb7539c432bc209b8473e80eff151c524d5"},"schema_version":"1.0","source":{"id":"2603.24125","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2603.24125","created_at":"2026-05-18T03:10:03Z"},{"alias_kind":"arxiv_version","alias_value":"2603.24125v2","created_at":"2026-05-18T03:10:03Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.24125","created_at":"2026-05-18T03:10:03Z"},{"alias_kind":"pith_short_12","alias_value":"GAEAA7U3YUL3","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"GAEAA7U3YUL3PO6N","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"GAEAA7U3","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:e319670f9572dc3d34e3980b4f8c5b48d32f8cc7fe4d80809f6728c0973b7b7a","target":"graph","created_at":"2026-05-18T03:10:03Z","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":"Contrary to prior work reporting weak or inconsistent correlations, we find a consistent association between latent gender information and expressed bias when measured under the unified protocol. [...] while the latter indeed reduces expressed bias, measurable gender-related associations are still present in internal representations, and can be reactivated under adversarial prompting."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the chosen neutral prompts and the specific methods for extracting latent gender information from internal representations accurately reflect the model's true encoded associations without being artifacts of prompt design or measurement choices."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Alignment reduces expressed gender bias in LLM outputs but does not remove the underlying encoded gender associations in internal representations."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Alignment reduces expressed gender bias in outputs but leaves measurable associations intact inside the model's representations."}],"snapshot_sha256":"938acf540ba51d619c72fe2d9ea867745f110ab88966928f891218c32f28aabf"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"abstract_excerpt":"During training, Large Language Models (LLMs) learn social regularities that can lead to gender bias in downstream applications. Most mitigation efforts focus on reducing bias in generated outputs, typically evaluated on structured benchmarks, which raises two concerns: output-level evaluation does not reveal whether alignment modifies the model's underlying representations, and structured benchmarks may not reflect realistic usage scenarios. We propose a unified framework to jointly analyze intrinsic and extrinsic gender bias in LLMs using identical neutral prompts, enabling direct comparison","authors_text":"Christophe Marsala, Marcin Detyniecki, Marie-Jeanne Lesot, Nour Bouchouchi, Thibault Laugel, Xavier Renard","cross_cats":[],"headline":"Alignment reduces expressed gender bias in outputs but leaves measurable associations intact inside the model's representations.","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-03-25T09:35:18Z","title":"Alignment Reduces Expressed but Not Encoded Gender Bias: A Unified Framework and Study"},"references":{"count":36,"internal_anchors":0,"resolved_work":36,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"In: NeurIPS (2024)","work_id":"17ac99e9-682f-4120-b540-f8f88a020ea4","year":2024},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"In: NeurIPS (2016)","work_id":"b724b495-fd98-449b-9e8d-0bd784a36ef0","year":2016},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"Caliskan, A., Bryson, J.J., Narayanan, A.: Semantics derived automatically from language corpora contain human-like biases. Science356, 183–186 (2017)","work_id":"78b8f9aa-f16f-47f6-93a0-273f216c6eb6","year":2017},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"In: ACL (2022)","work_id":"4a193780-15a1-4dbc-887c-3d061351c7f1","year":2022},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Psychology of sport and exercise 14, 136–144 (2013)","work_id":"d5769729-eae1-4296-aa39-85da84b349aa","year":2013}],"snapshot_sha256":"797286a345d43ff7182df78e7c5292858c6acbfc4eb1e87a92115fd15c98c5d5"},"source":{"id":"2603.24125","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-15T00:36:37.616118Z","id":"646e7624-6d4a-4f1b-adf2-02377bf94163","model_set":{"reader":"grok-4.3"},"one_line_summary":"Alignment reduces expressed gender bias in LLM outputs but does not remove the underlying encoded gender associations in internal representations.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Alignment reduces expressed gender bias in outputs but leaves measurable associations intact inside the model's representations.","strongest_claim":"Contrary to prior work reporting weak or inconsistent correlations, we find a consistent association between latent gender information and expressed bias when measured under the unified protocol. [...] while the latter indeed reduces expressed bias, measurable gender-related associations are still present in internal representations, and can be reactivated under adversarial prompting.","weakest_assumption":"That the chosen neutral prompts and the specific methods for extracting latent gender information from internal representations accurately reflect the model's true encoded associations without being artifacts of prompt design or measurement choices."}},"verdict_id":"646e7624-6d4a-4f1b-adf2-02377bf94163"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:4dd3c0a3bd71686b6713619bdaf5c8b796cfe089151fa5e811b2ccf3d7abe4c4","target":"record","created_at":"2026-05-18T03:10:03Z","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":"b358da891dc36defefb1874a0115d485852c0655cad89c05cd4e0975d8028b7f","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-03-25T09:35:18Z","title_canon_sha256":"b7263ca6adce8b93c8ff911975d92cb7539c432bc209b8473e80eff151c524d5"},"schema_version":"1.0","source":{"id":"2603.24125","kind":"arxiv","version":2}},"canonical_sha256":"3008007e9bc517b7bbcdcb483ebf6dd764b99bf14feacbb25a5d358cdd6ae221","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"3008007e9bc517b7bbcdcb483ebf6dd764b99bf14feacbb25a5d358cdd6ae221","first_computed_at":"2026-05-18T03:10:03.407514Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T03:10:03.407514Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"fj0J1gTZ7aYm7ty4Lj52YrPIK20AQefHSW3qhflfHVdaSbvpbBN/F1I9kObIc/Z5E4FRx8BDNWwUie+5vi+ECw==","signature_status":"signed_v1","signed_at":"2026-05-18T03:10:03.407998Z","signed_message":"canonical_sha256_bytes"},"source_id":"2603.24125","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:4dd3c0a3bd71686b6713619bdaf5c8b796cfe089151fa5e811b2ccf3d7abe4c4","sha256:e319670f9572dc3d34e3980b4f8c5b48d32f8cc7fe4d80809f6728c0973b7b7a"],"state_sha256":"e40b98e0f5b0f7aac44ef84f7821369396b339cf8d75e5e50faf606e39cced06"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"EcbPV98gdhSks0jgld3Bmq/R0MZgDo+nQZ5OcXcqE3toX3u4ccqC6zGyrlj6uIA6WEmhVIhC8YCM0nuHRTJZCQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-10T16:17:04.267742Z","bundle_sha256":"8b5351bab27c8285c6bbc269d64919a5632a924e8b3ebf21b9996ff2178b81a3"}}