{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:VU6HH5IEEBY2KLOM642PRGCLGX","short_pith_number":"pith:VU6HH5IE","schema_version":"1.0","canonical_sha256":"ad3c73f5042071a52dccf734f8984b35ea81b5d130f573786dbd43e961ab44ac","source":{"kind":"arxiv","id":"2405.14860","version":3},"attestation_state":"computed","paper":{"title":"Not All Language Model Features Are One-Dimensionally Linear","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Eric J. Michaud, Isaac Liao, Joshua Engels, Max Tegmark, Wes Gurnee","submitted_at":"2024-05-23T17:59:04Z","abstract_excerpt":"Recent work has proposed that language models perform computation by manipulating one-dimensional representations of concepts (\"features\") in activation space. In contrast, we explore whether some language model representations may be inherently multi-dimensional. We begin by developing a rigorous definition of irreducible multi-dimensional features based on whether they can be decomposed into either independent or non-co-occurring lower-dimensional features. Motivated by these definitions, we design a scalable method that uses sparse autoencoders to automatically find multi-dimensional featur"},"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":"2405.14860","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-05-23T17:59:04Z","cross_cats_sorted":[],"title_canon_sha256":"336e7c0495c7c712112877672e5368cc25791ebe9a93791fba57e01de667bcfd","abstract_canon_sha256":"a8ee70684c431d927b0cfdcaaa8141fecb69743547ebf686e84ea641e680f985"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:20:37.070796Z","signature_b64":"0aIt4Qyt7uVL9lNCxM+Z/Z9mp5JjddIEriKUXGkG/mTCLwaY59J2A6XX2lhihzEauYqkegWDqMiDaAf8BaJ9Bg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ad3c73f5042071a52dccf734f8984b35ea81b5d130f573786dbd43e961ab44ac","last_reissued_at":"2026-07-05T10:20:37.070275Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:20:37.070275Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Not All Language Model Features Are One-Dimensionally Linear","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Eric J. Michaud, Isaac Liao, Joshua Engels, Max Tegmark, Wes Gurnee","submitted_at":"2024-05-23T17:59:04Z","abstract_excerpt":"Recent work has proposed that language models perform computation by manipulating one-dimensional representations of concepts (\"features\") in activation space. In contrast, we explore whether some language model representations may be inherently multi-dimensional. We begin by developing a rigorous definition of irreducible multi-dimensional features based on whether they can be decomposed into either independent or non-co-occurring lower-dimensional features. Motivated by these definitions, we design a scalable method that uses sparse autoencoders to automatically find multi-dimensional featur"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2405.14860","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/2405.14860/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":"2405.14860","created_at":"2026-07-05T10:20:37.070342+00:00"},{"alias_kind":"arxiv_version","alias_value":"2405.14860v3","created_at":"2026-07-05T10:20:37.070342+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2405.14860","created_at":"2026-07-05T10:20:37.070342+00:00"},{"alias_kind":"pith_short_12","alias_value":"VU6HH5IEEBY2","created_at":"2026-07-05T10:20:37.070342+00:00"},{"alias_kind":"pith_short_16","alias_value":"VU6HH5IEEBY2KLOM","created_at":"2026-07-05T10:20:37.070342+00:00"},{"alias_kind":"pith_short_8","alias_value":"VU6HH5IE","created_at":"2026-07-05T10:20:37.070342+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":25,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.25234","citing_title":"Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifolds","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09658","citing_title":"Muon Learns More Robust and Transferable Features than Adam","ref_index":120,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10198","citing_title":"Density Ridge Selective Prediction for LLM and VLM Hallucination Detection under Calibration Label Scarcity","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09607","citing_title":"Closure-Validated Circuit Discovery in Attention Heads: Co-activation Proposes, Ablation Disposes","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05194","citing_title":"Temporal Preference Concepts and their Functions in a Large Language Model","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2606.15054","citing_title":"Size Doesn't Matter: Cosine-Scored Sparse Autoencoders","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2606.29522","citing_title":"Do Models Read What They Write? Causal Registers in Scratchpad Reasoning","ref_index":1,"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":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.27970","citing_title":"Geometry of Human Perceptual Domains Emerges Transiently in LLM Representations","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06333","citing_title":"Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2606.26396","citing_title":"At the Edge of Understanding: Sparse Autoencoders Trace The Limits of Transformer Generalization","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2603.01227","citing_title":"The Lattice Representation Hypothesis of Large Language Models","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2601.19208","citing_title":"How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic Interpretability","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.14075","citing_title":"Rethinking Layer Relevance in Large Language Models Beyond Cosine Similarity","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":39,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12770","citing_title":"WriteSAE: Sparse Autoencoders for Recurrent State","ref_index":39,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09967","citing_title":"Tensor Product Representation Probes Reveal Shared Structure Across Linear Directions","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07922","citing_title":"Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders","ref_index":8,"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":205,"is_internal_anchor":false},{"citing_arxiv_id":"2604.21691","citing_title":"There Will Be a Scientific Theory of Deep Learning","ref_index":217,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07922","citing_title":"Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07990","citing_title":"Tool Calling is Linearly Readable and Steerable in Language Models","ref_index":48,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00847","citing_title":"H-Probes: Extracting Hierarchical Structures From Latent Representations of Language Models","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2604.15557","citing_title":"Predicting Where Steering Vectors Succeed","ref_index":5,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VU6HH5IEEBY2KLOM642PRGCLGX","json":"https://pith.science/pith/VU6HH5IEEBY2KLOM642PRGCLGX.json","graph_json":"https://pith.science/api/pith-number/VU6HH5IEEBY2KLOM642PRGCLGX/graph.json","events_json":"https://pith.science/api/pith-number/VU6HH5IEEBY2KLOM642PRGCLGX/events.json","paper":"https://pith.science/paper/VU6HH5IE"},"agent_actions":{"view_html":"https://pith.science/pith/VU6HH5IEEBY2KLOM642PRGCLGX","download_json":"https://pith.science/pith/VU6HH5IEEBY2KLOM642PRGCLGX.json","view_paper":"https://pith.science/paper/VU6HH5IE","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2405.14860&json=true","fetch_graph":"https://pith.science/api/pith-number/VU6HH5IEEBY2KLOM642PRGCLGX/graph.json","fetch_events":"https://pith.science/api/pith-number/VU6HH5IEEBY2KLOM642PRGCLGX/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VU6HH5IEEBY2KLOM642PRGCLGX/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VU6HH5IEEBY2KLOM642PRGCLGX/action/storage_attestation","attest_author":"https://pith.science/pith/VU6HH5IEEBY2KLOM642PRGCLGX/action/author_attestation","sign_citation":"https://pith.science/pith/VU6HH5IEEBY2KLOM642PRGCLGX/action/citation_signature","submit_replication":"https://pith.science/pith/VU6HH5IEEBY2KLOM642PRGCLGX/action/replication_record"}},"created_at":"2026-07-05T10:20:37.070342+00:00","updated_at":"2026-07-05T10:20:37.070342+00:00"}