{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2021:PSQSOKGJ6QPPOLSC5Y6SWXPXBG","short_pith_number":"pith:PSQSOKGJ","schema_version":"1.0","canonical_sha256":"7ca12728c9f41ef72e42ee3d2b5df709ac3b27d50458bc215fddb58296eabf7d","source":{"kind":"arxiv","id":"2105.06990","version":2},"attestation_state":"computed","paper":{"title":"BERT Busters: Outlier Dimensions that Disrupt Transformers","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Anna Rogers, Anna Rumshisky, Olga Kovaleva, Saurabh Kulshreshtha","submitted_at":"2021-05-14T17:54:28Z","abstract_excerpt":"Multiple studies have shown that Transformers are remarkably robust to pruning. Contrary to this received wisdom, we demonstrate that pre-trained Transformer encoders are surprisingly fragile to the removal of a very small number of features in the layer outputs (<0.0001% of model weights). In case of BERT and other pre-trained encoder Transformers, the affected component is the scaling factors and biases in the LayerNorm. The outliers are high-magnitude normalization parameters that emerge early in pre-training and show up consistently in the same dimensional position throughout the model. We"},"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":"2105.06990","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2021-05-14T17:54:28Z","cross_cats_sorted":[],"title_canon_sha256":"fd765ba8f9e48937d608505bd2d442afaf240557d22431a8bf3b1b2652c93903","abstract_canon_sha256":"560d177629262e1e7524c54700b3db441c013b86e3c40d5a5b0a34e575518236"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T02:45:50.156477Z","signature_b64":"QLNlx7goQedIwCCBHsdRsOQ4m9uuuI+AaPyif7hgZ7pXIkbAALMeMqEQaiuAzTIuECJQz/joMcWEfwjF9iGEBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7ca12728c9f41ef72e42ee3d2b5df709ac3b27d50458bc215fddb58296eabf7d","last_reissued_at":"2026-07-05T02:45:50.155961Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T02:45:50.155961Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"BERT Busters: Outlier Dimensions that Disrupt Transformers","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Anna Rogers, Anna Rumshisky, Olga Kovaleva, Saurabh Kulshreshtha","submitted_at":"2021-05-14T17:54:28Z","abstract_excerpt":"Multiple studies have shown that Transformers are remarkably robust to pruning. Contrary to this received wisdom, we demonstrate that pre-trained Transformer encoders are surprisingly fragile to the removal of a very small number of features in the layer outputs (<0.0001% of model weights). In case of BERT and other pre-trained encoder Transformers, the affected component is the scaling factors and biases in the LayerNorm. The outliers are high-magnitude normalization parameters that emerge early in pre-training and show up consistently in the same dimensional position throughout the model. We"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2105.06990","kind":"arxiv","version":2},"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/2105.06990/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":"2105.06990","created_at":"2026-07-05T02:45:50.156029+00:00"},{"alias_kind":"arxiv_version","alias_value":"2105.06990v2","created_at":"2026-07-05T02:45:50.156029+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2105.06990","created_at":"2026-07-05T02:45:50.156029+00:00"},{"alias_kind":"pith_short_12","alias_value":"PSQSOKGJ6QPP","created_at":"2026-07-05T02:45:50.156029+00:00"},{"alias_kind":"pith_short_16","alias_value":"PSQSOKGJ6QPPOLSC","created_at":"2026-07-05T02:45:50.156029+00:00"},{"alias_kind":"pith_short_8","alias_value":"PSQSOKGJ","created_at":"2026-07-05T02:45:50.156029+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.20743","citing_title":"Massive Activations Are Architecturally Robust: A Controlled Scratch/Commitment Residual Stream Test","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2208.07339","citing_title":"LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale","ref_index":143,"is_internal_anchor":false},{"citing_arxiv_id":"2303.08112","citing_title":"Eliciting Latent Predictions from Transformers with the Tuned Lens","ref_index":49,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PSQSOKGJ6QPPOLSC5Y6SWXPXBG","json":"https://pith.science/pith/PSQSOKGJ6QPPOLSC5Y6SWXPXBG.json","graph_json":"https://pith.science/api/pith-number/PSQSOKGJ6QPPOLSC5Y6SWXPXBG/graph.json","events_json":"https://pith.science/api/pith-number/PSQSOKGJ6QPPOLSC5Y6SWXPXBG/events.json","paper":"https://pith.science/paper/PSQSOKGJ"},"agent_actions":{"view_html":"https://pith.science/pith/PSQSOKGJ6QPPOLSC5Y6SWXPXBG","download_json":"https://pith.science/pith/PSQSOKGJ6QPPOLSC5Y6SWXPXBG.json","view_paper":"https://pith.science/paper/PSQSOKGJ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2105.06990&json=true","fetch_graph":"https://pith.science/api/pith-number/PSQSOKGJ6QPPOLSC5Y6SWXPXBG/graph.json","fetch_events":"https://pith.science/api/pith-number/PSQSOKGJ6QPPOLSC5Y6SWXPXBG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PSQSOKGJ6QPPOLSC5Y6SWXPXBG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PSQSOKGJ6QPPOLSC5Y6SWXPXBG/action/storage_attestation","attest_author":"https://pith.science/pith/PSQSOKGJ6QPPOLSC5Y6SWXPXBG/action/author_attestation","sign_citation":"https://pith.science/pith/PSQSOKGJ6QPPOLSC5Y6SWXPXBG/action/citation_signature","submit_replication":"https://pith.science/pith/PSQSOKGJ6QPPOLSC5Y6SWXPXBG/action/replication_record"}},"created_at":"2026-07-05T02:45:50.156029+00:00","updated_at":"2026-07-05T02:45:50.156029+00:00"}