{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:2SQRZQWVLP3RE5QRKYYPLS5KIE","short_pith_number":"pith:2SQRZQWV","schema_version":"1.0","canonical_sha256":"d4a11cc2d55bf71276115630f5cbaa4115dd4ff6e69d2c2f6875ad656d7711a9","source":{"kind":"arxiv","id":"2605.14347","version":1},"attestation_state":"computed","paper":{"title":"Exemplar Partitioning for Mechanistic Interpretability","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Exemplar Partitioning constructs feature dictionaries for language model activations by clustering around observed exemplars, achieving near-SAE performance at much lower computational cost.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jessica Rumbelow","submitted_at":"2026-05-14T04:15:30Z","abstract_excerpt":"We introduce Exemplar Partitioning (EP), an unsupervised method for constructing interpretable feature dictionaries from large language model activations with $\\sim 10^{3}\\times$ fewer tokens than comparable sparse autoencoders (SAEs). An EP dictionary is a Voronoi partition of activation space, built by leader-clustering streamed activations within a distance threshold. Each region is anchored by an observed exemplar that serves as both its membership criterion and intervention direction; dictionary size is not prespecified, but determined by the activation geometry at that threshold. Because"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.14347","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-14T04:15:30Z","cross_cats_sorted":[],"title_canon_sha256":"325b159300e80cf7f2e654225d65059c4c87a20c84cfd4e299bf50ee7c414895","abstract_canon_sha256":"636aa3e6dcc915b5ec565a8db37bd6a75c7ef2486ee07309547e27d472f16e3a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:08.116256Z","signature_b64":"OishU94yVyU9rqjy84Ja6HH298sUpjSvs038Gd9Enw1xxx5DAxv7FytbOq8GWhEcOlrsWwI5Q2wwKjvsBblCDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d4a11cc2d55bf71276115630f5cbaa4115dd4ff6e69d2c2f6875ad656d7711a9","last_reissued_at":"2026-05-17T23:39:08.115573Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:08.115573Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Exemplar Partitioning for Mechanistic Interpretability","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Exemplar Partitioning constructs feature dictionaries for language model activations by clustering around observed exemplars, achieving near-SAE performance at much lower computational cost.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jessica Rumbelow","submitted_at":"2026-05-14T04:15:30Z","abstract_excerpt":"We introduce Exemplar Partitioning (EP), an unsupervised method for constructing interpretable feature dictionaries from large language model activations with $\\sim 10^{3}\\times$ fewer tokens than comparable sparse autoencoders (SAEs). An EP dictionary is a Voronoi partition of activation space, built by leader-clustering streamed activations within a distance threshold. Each region is anchored by an observed exemplar that serves as both its membership criterion and intervention direction; dictionary size is not prespecified, but determined by the activation geometry at that threshold. Because"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On AxBench latent concept detection at Gemma-2-2B-it L20, EP at p1 reaches mean AUROC 0.881, +0.126 over the canonical GemmaScope SAE leaderboard entry and within 0.030 of SAE-A's 0.911, at ~10^3× less build compute.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That nearest-exemplar Voronoi regions defined by a single distance threshold correspond to causally meaningful and human-interpretable features rather than arbitrary geometric clusters.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Exemplar Partitioning creates activation-space dictionaries via leader-clustered Voronoi partitions around real observed exemplars, delivering competitive concept-detection performance with far lower build cost than SAEs.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Exemplar Partitioning constructs feature dictionaries for language model activations by clustering around observed exemplars, achieving near-SAE performance at much lower computational cost.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"84c514c3d677e31c24932d9295469db9eea71cec33e28e05a1aee76929ee2f3b"},"source":{"id":"2605.14347","kind":"arxiv","version":1},"verdict":{"id":"6dda7827-0efe-46b3-ae2e-b9e7e178b789","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T01:30:59.455904Z","strongest_claim":"On AxBench latent concept detection at Gemma-2-2B-it L20, EP at p1 reaches mean AUROC 0.881, +0.126 over the canonical GemmaScope SAE leaderboard entry and within 0.030 of SAE-A's 0.911, at ~10^3× less build compute.","one_line_summary":"Exemplar Partitioning creates activation-space dictionaries via leader-clustered Voronoi partitions around real observed exemplars, delivering competitive concept-detection performance with far lower build cost than SAEs.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That nearest-exemplar Voronoi regions defined by a single distance threshold correspond to causally meaningful and human-interpretable features rather than arbitrary geometric clusters.","pith_extraction_headline":"Exemplar Partitioning constructs feature dictionaries for language model activations by clustering around observed exemplars, achieving near-SAE performance at much lower computational cost."},"references":{"count":39,"sample":[{"doi":"","year":null,"title":"Emergence of simple-cell receptive field properties by learning a sparse code for natural images , author=. Nature , volume=","work_id":"8cba53e0-2327-4f27-be37-7a0080774e16","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Toy Models of Superposition , author=. 2022 , note=","work_id":"b05a43bb-21aa-44ff-8424-df38240a48cf","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models , author=. 2023 , eprint=","work_id":"b0500e0d-343b-4ae2-8b3b-080070d36a77","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Towards Monosemanticity: Decomposing Language Models With Dictionary Learning , author=. 2023 , note=","work_id":"1231d466-a836-4494-a765-2e69c9918681","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet , author=. 2024 , note=","work_id":"bf57d145-a267-443c-823a-e0c0f0adab66","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":39,"snapshot_sha256":"428d10ed8bec878cb3741dcebb2734f4b5db5c5d3276c66845a1ac1e7056e392","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"ceb69e9c02c0c9c6389da959aed2f52ca38fde996724edf8b72b695460796fbe"},"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":"2605.14347","created_at":"2026-05-17T23:39:08.115676+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.14347v1","created_at":"2026-05-17T23:39:08.115676+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14347","created_at":"2026-05-17T23:39:08.115676+00:00"},{"alias_kind":"pith_short_12","alias_value":"2SQRZQWVLP3R","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"2SQRZQWVLP3RE5QR","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"2SQRZQWV","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE","json":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE.json","graph_json":"https://pith.science/api/pith-number/2SQRZQWVLP3RE5QRKYYPLS5KIE/graph.json","events_json":"https://pith.science/api/pith-number/2SQRZQWVLP3RE5QRKYYPLS5KIE/events.json","paper":"https://pith.science/paper/2SQRZQWV"},"agent_actions":{"view_html":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE","download_json":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE.json","view_paper":"https://pith.science/paper/2SQRZQWV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.14347&json=true","fetch_graph":"https://pith.science/api/pith-number/2SQRZQWVLP3RE5QRKYYPLS5KIE/graph.json","fetch_events":"https://pith.science/api/pith-number/2SQRZQWVLP3RE5QRKYYPLS5KIE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/action/storage_attestation","attest_author":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/action/author_attestation","sign_citation":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/action/citation_signature","submit_replication":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/action/replication_record"}},"created_at":"2026-05-17T23:39:08.115676+00:00","updated_at":"2026-05-17T23:39:08.115676+00:00"}