{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:SRPBUEJTSQHSLITQGY5XEKSRZQ","short_pith_number":"pith:SRPBUEJT","schema_version":"1.0","canonical_sha256":"945e1a1133940f25a270363b722a51cc3a91719782d584eba169fae7577a466d","source":{"kind":"arxiv","id":"2305.00586","version":5},"attestation_state":"computed","paper":{"title":"How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Alexandre Variengien, Michael Hanna, Ollie Liu","submitted_at":"2023-04-30T21:44:21Z","abstract_excerpt":"Pre-trained language models can be surprisingly adept at tasks they were not explicitly trained on, but how they implement these capabilities is poorly understood. In this paper, we investigate the basic mathematical abilities often acquired by pre-trained language models. Concretely, we use mechanistic interpretability techniques to explain the (limited) mathematical abilities of GPT-2 small. As a case study, we examine its ability to take in sentences such as \"The war lasted from the year 1732 to the year 17\", and predict valid two-digit end years (years > 32). We first identify a circuit, a"},"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":"2305.00586","kind":"arxiv","version":5},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2023-04-30T21:44:21Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"ff44c4e816685cc1733badfd7165cf1711ab6bc7e08f530118527f3c6deec4d9","abstract_canon_sha256":"095c66b2a5fec85ed3463c2e7a64c7bac786ea41025652ffb473258715cfcb4c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:08:06.805347Z","signature_b64":"oGm1ZOCBlvxAjdyT+6BfOWj/89AXp6IU4cqvRKJxHdkaACPKOro56iipyQjJ+LEpQ7YWQugOOJ5Xz7pxFdQmBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"945e1a1133940f25a270363b722a51cc3a91719782d584eba169fae7577a466d","last_reissued_at":"2026-07-05T07:08:06.804802Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:08:06.804802Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Alexandre Variengien, Michael Hanna, Ollie Liu","submitted_at":"2023-04-30T21:44:21Z","abstract_excerpt":"Pre-trained language models can be surprisingly adept at tasks they were not explicitly trained on, but how they implement these capabilities is poorly understood. In this paper, we investigate the basic mathematical abilities often acquired by pre-trained language models. Concretely, we use mechanistic interpretability techniques to explain the (limited) mathematical abilities of GPT-2 small. As a case study, we examine its ability to take in sentences such as \"The war lasted from the year 1732 to the year 17\", and predict valid two-digit end years (years > 32). We first identify a circuit, a"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2305.00586","kind":"arxiv","version":5},"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/2305.00586/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":"2305.00586","created_at":"2026-07-05T07:08:06.804860+00:00"},{"alias_kind":"arxiv_version","alias_value":"2305.00586v5","created_at":"2026-07-05T07:08:06.804860+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2305.00586","created_at":"2026-07-05T07:08:06.804860+00:00"},{"alias_kind":"pith_short_12","alias_value":"SRPBUEJTSQHS","created_at":"2026-07-05T07:08:06.804860+00:00"},{"alias_kind":"pith_short_16","alias_value":"SRPBUEJTSQHSLITQ","created_at":"2026-07-05T07:08:06.804860+00:00"},{"alias_kind":"pith_short_8","alias_value":"SRPBUEJT","created_at":"2026-07-05T07:08:06.804860+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.24026","citing_title":"Can Language Model Agents be Helpful Circuit Explainers in Mechanistic Interpretability?","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05378","citing_title":"Pattern Selectivity is Not Task-Causal Structure: A Cross-Architecture Mechanistic Study of Composed-Task Circuits in 1B-Class Language Models","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2404.15255","citing_title":"How to use and interpret activation patching","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12809","citing_title":"Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces","ref_index":152,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/SRPBUEJTSQHSLITQGY5XEKSRZQ","json":"https://pith.science/pith/SRPBUEJTSQHSLITQGY5XEKSRZQ.json","graph_json":"https://pith.science/api/pith-number/SRPBUEJTSQHSLITQGY5XEKSRZQ/graph.json","events_json":"https://pith.science/api/pith-number/SRPBUEJTSQHSLITQGY5XEKSRZQ/events.json","paper":"https://pith.science/paper/SRPBUEJT"},"agent_actions":{"view_html":"https://pith.science/pith/SRPBUEJTSQHSLITQGY5XEKSRZQ","download_json":"https://pith.science/pith/SRPBUEJTSQHSLITQGY5XEKSRZQ.json","view_paper":"https://pith.science/paper/SRPBUEJT","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2305.00586&json=true","fetch_graph":"https://pith.science/api/pith-number/SRPBUEJTSQHSLITQGY5XEKSRZQ/graph.json","fetch_events":"https://pith.science/api/pith-number/SRPBUEJTSQHSLITQGY5XEKSRZQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/SRPBUEJTSQHSLITQGY5XEKSRZQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/SRPBUEJTSQHSLITQGY5XEKSRZQ/action/storage_attestation","attest_author":"https://pith.science/pith/SRPBUEJTSQHSLITQGY5XEKSRZQ/action/author_attestation","sign_citation":"https://pith.science/pith/SRPBUEJTSQHSLITQGY5XEKSRZQ/action/citation_signature","submit_replication":"https://pith.science/pith/SRPBUEJTSQHSLITQGY5XEKSRZQ/action/replication_record"}},"created_at":"2026-07-05T07:08:06.804860+00:00","updated_at":"2026-07-05T07:08:06.804860+00:00"}