{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:CDWIKL7FJ34OLQDRKXUEVVYQP6","short_pith_number":"pith:CDWIKL7F","schema_version":"1.0","canonical_sha256":"10ec852fe54ef8e5c07155e84ad7107fbcfe74d48e0e3121af6a6fb885e2622a","source":{"kind":"arxiv","id":"2312.01648","version":3},"attestation_state":"computed","paper":{"title":"Characterizing Large Language Model Geometry Helps Solve Toxicity Detection and Generation","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.AI","authors_text":"Randall Balestriero, Romain Cosentino, Sarath Shekkizhar","submitted_at":"2023-12-04T06:01:32Z","abstract_excerpt":"Large Language Models (LLMs) drive current AI breakthroughs despite very little being known about their internal representations. In this work, we propose to shed the light on LLMs inner mechanisms through the lens of geometry. In particular, we develop in closed form $(i)$ the intrinsic dimension in which the Multi-Head Attention embeddings are constrained to exist and $(ii)$ the partition and per-region affine mappings of the feedforward (MLP) network of LLMs' layers. Our theoretical findings further enable the design of novel principled solutions applicable to state-of-the-art LLMs. First, "},"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":"2312.01648","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.AI","submitted_at":"2023-12-04T06:01:32Z","cross_cats_sorted":["cs.CL","cs.LG"],"title_canon_sha256":"783cf15eb2b94bdba76621bd93e00f503dc452d1af886acd1204952282c5a431","abstract_canon_sha256":"f5802bdeb9fa60cced8cc08dabf2b4fc67c901e0962c2f282621862885dda008"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:42:31.296416Z","signature_b64":"gCDsUorYbr42P/R+1EOsFSaa3JhgUi4gsF0CvsAocqog607NbJue+8c+4SvBBELn11Aq0UiaUMATo/ow7kDICw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"10ec852fe54ef8e5c07155e84ad7107fbcfe74d48e0e3121af6a6fb885e2622a","last_reissued_at":"2026-07-05T08:42:31.295983Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:42:31.295983Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Characterizing Large Language Model Geometry Helps Solve Toxicity Detection and Generation","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.AI","authors_text":"Randall Balestriero, Romain Cosentino, Sarath Shekkizhar","submitted_at":"2023-12-04T06:01:32Z","abstract_excerpt":"Large Language Models (LLMs) drive current AI breakthroughs despite very little being known about their internal representations. In this work, we propose to shed the light on LLMs inner mechanisms through the lens of geometry. In particular, we develop in closed form $(i)$ the intrinsic dimension in which the Multi-Head Attention embeddings are constrained to exist and $(ii)$ the partition and per-region affine mappings of the feedforward (MLP) network of LLMs' layers. Our theoretical findings further enable the design of novel principled solutions applicable to state-of-the-art LLMs. First, "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2312.01648","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/2312.01648/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":"2312.01648","created_at":"2026-07-05T08:42:31.296045+00:00"},{"alias_kind":"arxiv_version","alias_value":"2312.01648v3","created_at":"2026-07-05T08:42:31.296045+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2312.01648","created_at":"2026-07-05T08:42:31.296045+00:00"},{"alias_kind":"pith_short_12","alias_value":"CDWIKL7FJ34O","created_at":"2026-07-05T08:42:31.296045+00:00"},{"alias_kind":"pith_short_16","alias_value":"CDWIKL7FJ34OLQDR","created_at":"2026-07-05T08:42:31.296045+00:00"},{"alias_kind":"pith_short_8","alias_value":"CDWIKL7F","created_at":"2026-07-05T08:42:31.296045+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2502.09674","citing_title":"The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions","ref_index":6,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CDWIKL7FJ34OLQDRKXUEVVYQP6","json":"https://pith.science/pith/CDWIKL7FJ34OLQDRKXUEVVYQP6.json","graph_json":"https://pith.science/api/pith-number/CDWIKL7FJ34OLQDRKXUEVVYQP6/graph.json","events_json":"https://pith.science/api/pith-number/CDWIKL7FJ34OLQDRKXUEVVYQP6/events.json","paper":"https://pith.science/paper/CDWIKL7F"},"agent_actions":{"view_html":"https://pith.science/pith/CDWIKL7FJ34OLQDRKXUEVVYQP6","download_json":"https://pith.science/pith/CDWIKL7FJ34OLQDRKXUEVVYQP6.json","view_paper":"https://pith.science/paper/CDWIKL7F","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2312.01648&json=true","fetch_graph":"https://pith.science/api/pith-number/CDWIKL7FJ34OLQDRKXUEVVYQP6/graph.json","fetch_events":"https://pith.science/api/pith-number/CDWIKL7FJ34OLQDRKXUEVVYQP6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CDWIKL7FJ34OLQDRKXUEVVYQP6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CDWIKL7FJ34OLQDRKXUEVVYQP6/action/storage_attestation","attest_author":"https://pith.science/pith/CDWIKL7FJ34OLQDRKXUEVVYQP6/action/author_attestation","sign_citation":"https://pith.science/pith/CDWIKL7FJ34OLQDRKXUEVVYQP6/action/citation_signature","submit_replication":"https://pith.science/pith/CDWIKL7FJ34OLQDRKXUEVVYQP6/action/replication_record"}},"created_at":"2026-07-05T08:42:31.296045+00:00","updated_at":"2026-07-05T08:42:31.296045+00:00"}