{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:F74ANAEN6IIIFBA7HADZW7ADI3","short_pith_number":"pith:F74ANAEN","schema_version":"1.0","canonical_sha256":"2ff806808df21082841f38079b7c0346ce63c0d5cb3d3d63649d5a6af647f5a7","source":{"kind":"arxiv","id":"2410.20526","version":1},"attestation_state":"computed","paper":{"title":"Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Frances Liu, Junxuan Wang, Lingjie Chen, Qipeng Guo, Wentao Shu, Xipeng Qiu, Xuanjing Huang, Xuyang Ge, Yu-Gang Jiang, Yunhua Zhou, Zhengfu He, Zuxuan Wu","submitted_at":"2024-10-27T17:33:49Z","abstract_excerpt":"Sparse Autoencoders (SAEs) have emerged as a powerful unsupervised method for extracting sparse representations from language models, yet scalable training remains a significant challenge. We introduce a suite of 256 SAEs, trained on each layer and sublayer of the Llama-3.1-8B-Base model, with 32K and 128K features. Modifications to a state-of-the-art SAE variant, Top-K SAEs, are evaluated across multiple dimensions. In particular, we assess the generalizability of SAEs trained on base models to longer contexts and fine-tuned models. Additionally, we analyze the geometry of learned SAE latents"},"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":"2410.20526","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-10-27T17:33:49Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"4aa76a59736974149fbeaf5983bb66cf2671221a180c16a274a4691752b61f15","abstract_canon_sha256":"252ba682d83db693763d5f75206b16e763b51915a05e8393fb5a94810e1a2d87"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:27:08.234756Z","signature_b64":"o6s+YqbOSiT5HJd4yb9NiUw6oMuPCF5ef5syNsR36P0UB0dCf1H+kAXdba6JzsYSOYkQBbK6evWezRBsHx8dDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2ff806808df21082841f38079b7c0346ce63c0d5cb3d3d63649d5a6af647f5a7","last_reissued_at":"2026-07-05T09:27:08.234310Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:27:08.234310Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Frances Liu, Junxuan Wang, Lingjie Chen, Qipeng Guo, Wentao Shu, Xipeng Qiu, Xuanjing Huang, Xuyang Ge, Yu-Gang Jiang, Yunhua Zhou, Zhengfu He, Zuxuan Wu","submitted_at":"2024-10-27T17:33:49Z","abstract_excerpt":"Sparse Autoencoders (SAEs) have emerged as a powerful unsupervised method for extracting sparse representations from language models, yet scalable training remains a significant challenge. We introduce a suite of 256 SAEs, trained on each layer and sublayer of the Llama-3.1-8B-Base model, with 32K and 128K features. Modifications to a state-of-the-art SAE variant, Top-K SAEs, are evaluated across multiple dimensions. In particular, we assess the generalizability of SAEs trained on base models to longer contexts and fine-tuned models. Additionally, we analyze the geometry of learned SAE latents"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.20526","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/2410.20526/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":"2410.20526","created_at":"2026-07-05T09:27:08.234374+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.20526v1","created_at":"2026-07-05T09:27:08.234374+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.20526","created_at":"2026-07-05T09:27:08.234374+00:00"},{"alias_kind":"pith_short_12","alias_value":"F74ANAEN6III","created_at":"2026-07-05T09:27:08.234374+00:00"},{"alias_kind":"pith_short_16","alias_value":"F74ANAEN6IIIFBA7","created_at":"2026-07-05T09:27:08.234374+00:00"},{"alias_kind":"pith_short_8","alias_value":"F74ANAEN","created_at":"2026-07-05T09:27:08.234374+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":19,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.26620","citing_title":"Discovering Millions of Interpretable Features with Sparse Autoencoders","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08267","citing_title":"Post-AGI Economies: Superposition and the Second Fundamental Theorem of Welfare Economics","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2607.00397","citing_title":"NeuroCogMap Reveals Cognitive Organization of Large Language Models","ref_index":83,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00356","citing_title":"How Far Do Auto-Interpretation Labels Generalize: A Controlled Study Across Languages, Scripts, and Rewordings","ref_index":47,"is_internal_anchor":false},{"citing_arxiv_id":"2605.03052","citing_title":"How Language Models Process Negation","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2605.27819","citing_title":"ReSAE: Residualized Sparse Autoencoders for Multi-Layer Transformer Interventions","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28149","citing_title":"Sign-Aware Gated Sparse Autoencoders: Modeling Anticorrelated Features with Bi-Jump-ReLU Activations","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28649","citing_title":"Interpretability-Guided Layer Selection over Subspace Projection: SAEs as Stethoscopes, Not Scalpels, for Raw Task Vector Model Editing","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2605.23036","citing_title":"Multilingual Steering by Design: Multilingual Sparse Autoencoders and Principled Layer Selection","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2509.18127","citing_title":"Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2601.14004","citing_title":"Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models","ref_index":107,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12874","citing_title":"Descriptive Collision in Sparse Autoencoder Auto-Interpretability: When One Explanation Describes Many Features","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2604.23877","citing_title":"Knowledge Vector of Logical Reasoning in Large Language Models","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2604.11297","citing_title":"The Past Is Not Past: Memory-Enhanced Dynamic Reward Shaping","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14090","citing_title":"From Weights to Activations: Is Steering the Next Frontier of Adaptation?","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19974","citing_title":"Are LLM Uncertainty and Correctness Encoded by the Same Features? A Functional Dissociation via Sparse Autoencoders","ref_index":39,"is_internal_anchor":false},{"citing_arxiv_id":"2605.03052","citing_title":"How Language Models Process Negation","ref_index":63,"is_internal_anchor":false},{"citing_arxiv_id":"2605.03160","citing_title":"Steering grids for sparse-autoencoder features: when a top-context label names an activation regime rather than a causal axis","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02234","citing_title":"Bucketing the Good Apples: A Method for Diagnosing and Improving Causal Abstraction","ref_index":11,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/F74ANAEN6IIIFBA7HADZW7ADI3","json":"https://pith.science/pith/F74ANAEN6IIIFBA7HADZW7ADI3.json","graph_json":"https://pith.science/api/pith-number/F74ANAEN6IIIFBA7HADZW7ADI3/graph.json","events_json":"https://pith.science/api/pith-number/F74ANAEN6IIIFBA7HADZW7ADI3/events.json","paper":"https://pith.science/paper/F74ANAEN"},"agent_actions":{"view_html":"https://pith.science/pith/F74ANAEN6IIIFBA7HADZW7ADI3","download_json":"https://pith.science/pith/F74ANAEN6IIIFBA7HADZW7ADI3.json","view_paper":"https://pith.science/paper/F74ANAEN","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.20526&json=true","fetch_graph":"https://pith.science/api/pith-number/F74ANAEN6IIIFBA7HADZW7ADI3/graph.json","fetch_events":"https://pith.science/api/pith-number/F74ANAEN6IIIFBA7HADZW7ADI3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/F74ANAEN6IIIFBA7HADZW7ADI3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/F74ANAEN6IIIFBA7HADZW7ADI3/action/storage_attestation","attest_author":"https://pith.science/pith/F74ANAEN6IIIFBA7HADZW7ADI3/action/author_attestation","sign_citation":"https://pith.science/pith/F74ANAEN6IIIFBA7HADZW7ADI3/action/citation_signature","submit_replication":"https://pith.science/pith/F74ANAEN6IIIFBA7HADZW7ADI3/action/replication_record"}},"created_at":"2026-07-05T09:27:08.234374+00:00","updated_at":"2026-07-05T09:27:08.234374+00:00"}