{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:3P3TYENET7I3KKDEL2ISSJWTXF","short_pith_number":"pith:3P3TYENE","schema_version":"1.0","canonical_sha256":"dbf73c11a49fd1b528645e912926d3b96a9381a390a83f195beef8eb93630a88","source":{"kind":"arxiv","id":"2405.15943","version":3},"attestation_state":"computed","paper":{"title":"Transformers represent belief state geometry in their residual stream","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Adam S. Shai, Alexander Gietelink Oldenziel, Lucas Teixeira, Paul M. Riechers, Sarah E. Marzen","submitted_at":"2024-05-24T21:14:10Z","abstract_excerpt":"What computational structure are we building into large language models when we train them on next-token prediction? Here, we present evidence that this structure is given by the meta-dynamics of belief updating over hidden states of the data-generating process. Leveraging the theory of optimal prediction, we anticipate and then find that belief states are linearly represented in the residual stream of transformers, even in cases where the predicted belief state geometry has highly nontrivial fractal structure. We investigate cases where the belief state geometry is represented in the final re"},"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":"2405.15943","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-05-24T21:14:10Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"d7b18c6b8a0c841f100c518f62e118f0fc6d11672477cedd0be557aaf31a2246","abstract_canon_sha256":"8b03ddcd544100caa83e47ca597116ead5918fce3dd6d9c8a029a7a24ed7b3a1"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:09:04.557715Z","signature_b64":"p2kyM9Bjfuxw256yP24YWxwQKqSMlnvz3l15K07GdHkjM6YYlJQzXDqcjCI5wNESjOyQlhEZpPTc8SRSSABqBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"dbf73c11a49fd1b528645e912926d3b96a9381a390a83f195beef8eb93630a88","last_reissued_at":"2026-07-05T10:09:04.557135Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:09:04.557135Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Transformers represent belief state geometry in their residual stream","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Adam S. Shai, Alexander Gietelink Oldenziel, Lucas Teixeira, Paul M. Riechers, Sarah E. Marzen","submitted_at":"2024-05-24T21:14:10Z","abstract_excerpt":"What computational structure are we building into large language models when we train them on next-token prediction? Here, we present evidence that this structure is given by the meta-dynamics of belief updating over hidden states of the data-generating process. Leveraging the theory of optimal prediction, we anticipate and then find that belief states are linearly represented in the residual stream of transformers, even in cases where the predicted belief state geometry has highly nontrivial fractal structure. We investigate cases where the belief state geometry is represented in the final re"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2405.15943","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/2405.15943/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":"2405.15943","created_at":"2026-07-05T10:09:04.557199+00:00"},{"alias_kind":"arxiv_version","alias_value":"2405.15943v3","created_at":"2026-07-05T10:09:04.557199+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2405.15943","created_at":"2026-07-05T10:09:04.557199+00:00"},{"alias_kind":"pith_short_12","alias_value":"3P3TYENET7I3","created_at":"2026-07-05T10:09:04.557199+00:00"},{"alias_kind":"pith_short_16","alias_value":"3P3TYENET7I3KKDE","created_at":"2026-07-05T10:09:04.557199+00:00"},{"alias_kind":"pith_short_8","alias_value":"3P3TYENE","created_at":"2026-07-05T10:09:04.557199+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":8,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.30813","citing_title":"Gradient Smoothing: Coupling Layer-wise Updates for Improved Optimization","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07568","citing_title":"A Systematic Study of Behavioral Cloning for Scientific Data Annotation","ref_index":82,"is_internal_anchor":false},{"citing_arxiv_id":"2512.23752","citing_title":"Geometric Scaling of Bayesian Inference in LLMs","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02028","citing_title":"Language models fail at extended rule following","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":58,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":58,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02685","citing_title":"Finding Belief Geometries with Sparse Autoencoders","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02028","citing_title":"Language models fail at extended rule following","ref_index":45,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3P3TYENET7I3KKDEL2ISSJWTXF","json":"https://pith.science/pith/3P3TYENET7I3KKDEL2ISSJWTXF.json","graph_json":"https://pith.science/api/pith-number/3P3TYENET7I3KKDEL2ISSJWTXF/graph.json","events_json":"https://pith.science/api/pith-number/3P3TYENET7I3KKDEL2ISSJWTXF/events.json","paper":"https://pith.science/paper/3P3TYENE"},"agent_actions":{"view_html":"https://pith.science/pith/3P3TYENET7I3KKDEL2ISSJWTXF","download_json":"https://pith.science/pith/3P3TYENET7I3KKDEL2ISSJWTXF.json","view_paper":"https://pith.science/paper/3P3TYENE","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2405.15943&json=true","fetch_graph":"https://pith.science/api/pith-number/3P3TYENET7I3KKDEL2ISSJWTXF/graph.json","fetch_events":"https://pith.science/api/pith-number/3P3TYENET7I3KKDEL2ISSJWTXF/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3P3TYENET7I3KKDEL2ISSJWTXF/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3P3TYENET7I3KKDEL2ISSJWTXF/action/storage_attestation","attest_author":"https://pith.science/pith/3P3TYENET7I3KKDEL2ISSJWTXF/action/author_attestation","sign_citation":"https://pith.science/pith/3P3TYENET7I3KKDEL2ISSJWTXF/action/citation_signature","submit_replication":"https://pith.science/pith/3P3TYENET7I3KKDEL2ISSJWTXF/action/replication_record"}},"created_at":"2026-07-05T10:09:04.557199+00:00","updated_at":"2026-07-05T10:09:04.557199+00:00"}