{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:27IACSK5HFEEQL3VKBQC7H3NCO","short_pith_number":"pith:27IACSK5","schema_version":"1.0","canonical_sha256":"d7d001495d3948482f7550602f9f6d139077ab08c7b054ff42439be81cf717d0","source":{"kind":"arxiv","id":"2301.05062","version":5},"attestation_state":"computed","paper":{"title":"Tracr: Compiled Transformers as a Laboratory for Interpretability","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"David Lindner, J\\'anos Kram\\'ar, Matthew Rahtz, Sebastian Farquhar, Thomas McGrath, Vladimir Mikulik","submitted_at":"2023-01-12T14:59:19Z","abstract_excerpt":"We show how to \"compile\" human-readable programs into standard decoder-only transformer models. Our compiler, Tracr, generates models with known structure. This structure can be used to design experiments. For example, we use it to study \"superposition\" in transformers that execute multi-step algorithms. Additionally, the known structure of Tracr-compiled models can serve as ground-truth for evaluating interpretability methods. Commonly, because the \"programs\" learned by transformers are unknown it is unclear whether an interpretation succeeded. We demonstrate our approach by implementing and "},"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":"2301.05062","kind":"arxiv","version":5},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2023-01-12T14:59:19Z","cross_cats_sorted":["cs.AI","stat.ML"],"title_canon_sha256":"528d937f5367a1e932d3e420873094c788d76299c3d522c115d6155ff41da5cd","abstract_canon_sha256":"a8c169bea738473c3744178a6660c39491c9ec95e3dadcb710e7fe8eefc7b849"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:08:40.443660Z","signature_b64":"c74Nl+HQGjrsXmDg6Ia1COdCOoJegB7bebn9h00z6+UrvvCadFk9fZuh77TPUn2XxJSkFadteIT1H4k08DikCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d7d001495d3948482f7550602f9f6d139077ab08c7b054ff42439be81cf717d0","last_reissued_at":"2026-07-05T07:08:40.443197Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:08:40.443197Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Tracr: Compiled Transformers as a Laboratory for Interpretability","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"David Lindner, J\\'anos Kram\\'ar, Matthew Rahtz, Sebastian Farquhar, Thomas McGrath, Vladimir Mikulik","submitted_at":"2023-01-12T14:59:19Z","abstract_excerpt":"We show how to \"compile\" human-readable programs into standard decoder-only transformer models. Our compiler, Tracr, generates models with known structure. This structure can be used to design experiments. For example, we use it to study \"superposition\" in transformers that execute multi-step algorithms. Additionally, the known structure of Tracr-compiled models can serve as ground-truth for evaluating interpretability methods. Commonly, because the \"programs\" learned by transformers are unknown it is unclear whether an interpretation succeeded. We demonstrate our approach by implementing and "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2301.05062","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/2301.05062/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":"2301.05062","created_at":"2026-07-05T07:08:40.443255+00:00"},{"alias_kind":"arxiv_version","alias_value":"2301.05062v5","created_at":"2026-07-05T07:08:40.443255+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2301.05062","created_at":"2026-07-05T07:08:40.443255+00:00"},{"alias_kind":"pith_short_12","alias_value":"27IACSK5HFEE","created_at":"2026-07-05T07:08:40.443255+00:00"},{"alias_kind":"pith_short_16","alias_value":"27IACSK5HFEEQL3V","created_at":"2026-07-05T07:08:40.443255+00:00"},{"alias_kind":"pith_short_8","alias_value":"27IACSK5","created_at":"2026-07-05T07:08:40.443255+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2309.16797","citing_title":"Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution","ref_index":209,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12809","citing_title":"Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces","ref_index":144,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/27IACSK5HFEEQL3VKBQC7H3NCO","json":"https://pith.science/pith/27IACSK5HFEEQL3VKBQC7H3NCO.json","graph_json":"https://pith.science/api/pith-number/27IACSK5HFEEQL3VKBQC7H3NCO/graph.json","events_json":"https://pith.science/api/pith-number/27IACSK5HFEEQL3VKBQC7H3NCO/events.json","paper":"https://pith.science/paper/27IACSK5"},"agent_actions":{"view_html":"https://pith.science/pith/27IACSK5HFEEQL3VKBQC7H3NCO","download_json":"https://pith.science/pith/27IACSK5HFEEQL3VKBQC7H3NCO.json","view_paper":"https://pith.science/paper/27IACSK5","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2301.05062&json=true","fetch_graph":"https://pith.science/api/pith-number/27IACSK5HFEEQL3VKBQC7H3NCO/graph.json","fetch_events":"https://pith.science/api/pith-number/27IACSK5HFEEQL3VKBQC7H3NCO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/27IACSK5HFEEQL3VKBQC7H3NCO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/27IACSK5HFEEQL3VKBQC7H3NCO/action/storage_attestation","attest_author":"https://pith.science/pith/27IACSK5HFEEQL3VKBQC7H3NCO/action/author_attestation","sign_citation":"https://pith.science/pith/27IACSK5HFEEQL3VKBQC7H3NCO/action/citation_signature","submit_replication":"https://pith.science/pith/27IACSK5HFEEQL3VKBQC7H3NCO/action/replication_record"}},"created_at":"2026-07-05T07:08:40.443255+00:00","updated_at":"2026-07-05T07:08:40.443255+00:00"}