{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:M6V5KM57MO5SS5AHTPEZ6MATBI","short_pith_number":"pith:M6V5KM57","schema_version":"1.0","canonical_sha256":"67abd533bf63bb2974079bc99f30130a31cdc1bf9b3a1ca8282382084106427f","source":{"kind":"arxiv","id":"2408.10920","version":1},"attestation_state":"computed","paper":{"title":"Recurrent Neural Networks Learn to Store and Generate Sequences using Non-Linear Representations","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.NE"],"primary_cat":"cs.LG","authors_text":"Atticus Geiger, Christopher D. Manning, Christopher Potts, R\\'obert Csord\\'as","submitted_at":"2024-08-20T15:04:37Z","abstract_excerpt":"The Linear Representation Hypothesis (LRH) states that neural networks learn to encode concepts as directions in activation space, and a strong version of the LRH states that models learn only such encodings. In this paper, we present a counterexample to this strong LRH: when trained to repeat an input token sequence, gated recurrent neural networks (RNNs) learn to represent the token at each position with a particular order of magnitude, rather than a direction. These representations have layered features that are impossible to locate in distinct linear subspaces. To show this, we train inter"},"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":"2408.10920","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-08-20T15:04:37Z","cross_cats_sorted":["cs.AI","cs.NE"],"title_canon_sha256":"2d94a859f3072775edae53963aabd871b435fc0cff734e391d66216189570c7b","abstract_canon_sha256":"9c38918a2e54405e4f3712a44032dcf284627d94ba3359fa16ecfa5d769c388f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:57:19.689344Z","signature_b64":"2binOLZv3MbbRU4g506i73hJdsfX6VGCsmTCo+qw8deolEKQ5tGOMOhQXGkbJc14eV6HpyfXVQYBL5nqVnuDBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"67abd533bf63bb2974079bc99f30130a31cdc1bf9b3a1ca8282382084106427f","last_reissued_at":"2026-07-05T08:57:19.688897Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:57:19.688897Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Recurrent Neural Networks Learn to Store and Generate Sequences using Non-Linear Representations","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.NE"],"primary_cat":"cs.LG","authors_text":"Atticus Geiger, Christopher D. Manning, Christopher Potts, R\\'obert Csord\\'as","submitted_at":"2024-08-20T15:04:37Z","abstract_excerpt":"The Linear Representation Hypothesis (LRH) states that neural networks learn to encode concepts as directions in activation space, and a strong version of the LRH states that models learn only such encodings. In this paper, we present a counterexample to this strong LRH: when trained to repeat an input token sequence, gated recurrent neural networks (RNNs) learn to represent the token at each position with a particular order of magnitude, rather than a direction. These representations have layered features that are impossible to locate in distinct linear subspaces. To show this, we train inter"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2408.10920","kind":"arxiv","version":1},"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/2408.10920/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":"2408.10920","created_at":"2026-07-05T08:57:19.688950+00:00"},{"alias_kind":"arxiv_version","alias_value":"2408.10920v1","created_at":"2026-07-05T08:57:19.688950+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2408.10920","created_at":"2026-07-05T08:57:19.688950+00:00"},{"alias_kind":"pith_short_12","alias_value":"M6V5KM57MO5S","created_at":"2026-07-05T08:57:19.688950+00:00"},{"alias_kind":"pith_short_16","alias_value":"M6V5KM57MO5SS5AH","created_at":"2026-07-05T08:57:19.688950+00:00"},{"alias_kind":"pith_short_8","alias_value":"M6V5KM57","created_at":"2026-07-05T08:57:19.688950+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2509.02967","citing_title":"AR-KAN: Autoregressive-Weight-Enhanced Kolmogorov-Arnold Network for Time Series Forecasting","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05115","citing_title":"Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior","ref_index":207,"is_internal_anchor":false},{"citing_arxiv_id":"2604.21691","citing_title":"There Will Be a Scientific Theory of Deep Learning","ref_index":221,"is_internal_anchor":false},{"citing_arxiv_id":"2604.15557","citing_title":"Predicting Where Steering Vectors Succeed","ref_index":3,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/M6V5KM57MO5SS5AHTPEZ6MATBI","json":"https://pith.science/pith/M6V5KM57MO5SS5AHTPEZ6MATBI.json","graph_json":"https://pith.science/api/pith-number/M6V5KM57MO5SS5AHTPEZ6MATBI/graph.json","events_json":"https://pith.science/api/pith-number/M6V5KM57MO5SS5AHTPEZ6MATBI/events.json","paper":"https://pith.science/paper/M6V5KM57"},"agent_actions":{"view_html":"https://pith.science/pith/M6V5KM57MO5SS5AHTPEZ6MATBI","download_json":"https://pith.science/pith/M6V5KM57MO5SS5AHTPEZ6MATBI.json","view_paper":"https://pith.science/paper/M6V5KM57","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2408.10920&json=true","fetch_graph":"https://pith.science/api/pith-number/M6V5KM57MO5SS5AHTPEZ6MATBI/graph.json","fetch_events":"https://pith.science/api/pith-number/M6V5KM57MO5SS5AHTPEZ6MATBI/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/M6V5KM57MO5SS5AHTPEZ6MATBI/action/timestamp_anchor","attest_storage":"https://pith.science/pith/M6V5KM57MO5SS5AHTPEZ6MATBI/action/storage_attestation","attest_author":"https://pith.science/pith/M6V5KM57MO5SS5AHTPEZ6MATBI/action/author_attestation","sign_citation":"https://pith.science/pith/M6V5KM57MO5SS5AHTPEZ6MATBI/action/citation_signature","submit_replication":"https://pith.science/pith/M6V5KM57MO5SS5AHTPEZ6MATBI/action/replication_record"}},"created_at":"2026-07-05T08:57:19.688950+00:00","updated_at":"2026-07-05T08:57:19.688950+00:00"}