{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2022:UDD7EK2NNZHLRQNKQTCNNUXDYM","short_pith_number":"pith:UDD7EK2N","schema_version":"1.0","canonical_sha256":"a0c7f22b4d6e4eb8c1aa84c4d6d2e3c32884202ea5bec8d382ac3cd197bcfefc","source":{"kind":"arxiv","id":"2207.04153","version":3},"attestation_state":"computed","paper":{"title":"Probing Classifiers are Unreliable for Concept Removal and Detection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Abhinav Kumar, Amit Sharma, Chenhao Tan","submitted_at":"2022-07-08T23:15:26Z","abstract_excerpt":"Neural network models trained on text data have been found to encode undesirable linguistic or sensitive concepts in their representation. Removing such concepts is non-trivial because of a complex relationship between the concept, text input, and the learnt representation. Recent work has proposed post-hoc and adversarial methods to remove such unwanted concepts from a model's representation. Through an extensive theoretical and empirical analysis, we show that these methods can be counter-productive: they are unable to remove the concepts entirely, and in the worst case may end up destroying"},"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":"2207.04153","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2022-07-08T23:15:26Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"5d76f34bd3db5c272c9b292890c12f58fbf5654f6eb2bc8b8fa3eca1f66e29d3","abstract_canon_sha256":"9c49b1f430dfa4ed2aec184577d3684207eaace50353c5a48f0da2a8e9d6d5f7"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T06:21:48.125841Z","signature_b64":"rOViFGIl8aRY7YMb4SuNfNp2m5pitlPYrObTg9sPf1m5PQwdkFEiFkVZvdJKmy0T2rI5qTGzvxOa0sc6ov/7DQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a0c7f22b4d6e4eb8c1aa84c4d6d2e3c32884202ea5bec8d382ac3cd197bcfefc","last_reissued_at":"2026-07-05T06:21:48.125431Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T06:21:48.125431Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Probing Classifiers are Unreliable for Concept Removal and Detection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Abhinav Kumar, Amit Sharma, Chenhao Tan","submitted_at":"2022-07-08T23:15:26Z","abstract_excerpt":"Neural network models trained on text data have been found to encode undesirable linguistic or sensitive concepts in their representation. Removing such concepts is non-trivial because of a complex relationship between the concept, text input, and the learnt representation. Recent work has proposed post-hoc and adversarial methods to remove such unwanted concepts from a model's representation. Through an extensive theoretical and empirical analysis, we show that these methods can be counter-productive: they are unable to remove the concepts entirely, and in the worst case may end up destroying"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2207.04153","kind":"arxiv","version":3},"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/2207.04153/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":"2207.04153","created_at":"2026-07-05T06:21:48.125485+00:00"},{"alias_kind":"arxiv_version","alias_value":"2207.04153v3","created_at":"2026-07-05T06:21:48.125485+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2207.04153","created_at":"2026-07-05T06:21:48.125485+00:00"},{"alias_kind":"pith_short_12","alias_value":"UDD7EK2NNZHL","created_at":"2026-07-05T06:21:48.125485+00:00"},{"alias_kind":"pith_short_16","alias_value":"UDD7EK2NNZHLRQNK","created_at":"2026-07-05T06:21:48.125485+00:00"},{"alias_kind":"pith_short_8","alias_value":"UDD7EK2N","created_at":"2026-07-05T06:21:48.125485+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.12088","citing_title":"Debiasing Without Protected Attributes: Latent Concept Erasure from Textual Profiles","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2303.08112","citing_title":"Eliciting Latent Predictions from Transformers with the Tuned Lens","ref_index":50,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UDD7EK2NNZHLRQNKQTCNNUXDYM","json":"https://pith.science/pith/UDD7EK2NNZHLRQNKQTCNNUXDYM.json","graph_json":"https://pith.science/api/pith-number/UDD7EK2NNZHLRQNKQTCNNUXDYM/graph.json","events_json":"https://pith.science/api/pith-number/UDD7EK2NNZHLRQNKQTCNNUXDYM/events.json","paper":"https://pith.science/paper/UDD7EK2N"},"agent_actions":{"view_html":"https://pith.science/pith/UDD7EK2NNZHLRQNKQTCNNUXDYM","download_json":"https://pith.science/pith/UDD7EK2NNZHLRQNKQTCNNUXDYM.json","view_paper":"https://pith.science/paper/UDD7EK2N","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2207.04153&json=true","fetch_graph":"https://pith.science/api/pith-number/UDD7EK2NNZHLRQNKQTCNNUXDYM/graph.json","fetch_events":"https://pith.science/api/pith-number/UDD7EK2NNZHLRQNKQTCNNUXDYM/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UDD7EK2NNZHLRQNKQTCNNUXDYM/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UDD7EK2NNZHLRQNKQTCNNUXDYM/action/storage_attestation","attest_author":"https://pith.science/pith/UDD7EK2NNZHLRQNKQTCNNUXDYM/action/author_attestation","sign_citation":"https://pith.science/pith/UDD7EK2NNZHLRQNKQTCNNUXDYM/action/citation_signature","submit_replication":"https://pith.science/pith/UDD7EK2NNZHLRQNKQTCNNUXDYM/action/replication_record"}},"created_at":"2026-07-05T06:21:48.125485+00:00","updated_at":"2026-07-05T06:21:48.125485+00:00"}