{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:2SQRZQWVLP3RE5QRKYYPLS5KIE","short_pith_number":"pith:2SQRZQWV","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"},"canonical_sha256":"d4a11cc2d55bf71276115630f5cbaa4115dd4ff6e69d2c2f6875ad656d7711a9","source":{"kind":"arxiv","id":"2605.14347","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.14347","created_at":"2026-05-17T23:39:08Z"},{"alias_kind":"arxiv_version","alias_value":"2605.14347v1","created_at":"2026-05-17T23:39:08Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14347","created_at":"2026-05-17T23:39:08Z"},{"alias_kind":"pith_short_12","alias_value":"2SQRZQWVLP3R","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"2SQRZQWVLP3RE5QR","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"2SQRZQWV","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:2SQRZQWVLP3RE5QRKYYPLS5KIE","target":"record","payload":{"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"},"canonical_sha256":"d4a11cc2d55bf71276115630f5cbaa4115dd4ff6e69d2c2f6875ad656d7711a9","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"},"source_kind":"arxiv","source_id":"2605.14347","source_version":1,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:39:08Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"l3QXNaofNQSiGefLMzaavLx2H5PLQLyIwDpehs+m9ckCeocghSTzbioi+YVbt/3bTCULkWl0ojvYxW+hEOWqAw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-25T15:01:05.194806Z"},"content_sha256":"e04de711c2a68e8ca813a0865c8279b8d330461233a223682bd18c1c7d300541","schema_version":"1.0","event_id":"sha256:e04de711c2a68e8ca813a0865c8279b8d330461233a223682bd18c1c7d300541"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:2SQRZQWVLP3RE5QRKYYPLS5KIE","target":"graph","payload":{"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"},"verdict_id":"6dda7827-0efe-46b3-ae2e-b9e7e178b789"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:39:08Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"tRSLthcHPb9dYTK0WCrkM9YeC+LQrqJmSS4PKjgw5lkiuMooIX+VQCA4E1nIWYlM56JzqERVPk1Ozsd2qdOYBQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-25T15:01:05.195898Z"},"content_sha256":"5c57e18b43828d8ebe55264cdca8faa02ff31b33be0f9bdd30b54ce9c4a9ce1c","schema_version":"1.0","event_id":"sha256:5c57e18b43828d8ebe55264cdca8faa02ff31b33be0f9bdd30b54ce9c4a9ce1c"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/bundle.json","state_url":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-05-25T15:01:05Z","links":{"resolver":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE","bundle":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/bundle.json","state":"https://pith.science/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/state.json","well_known_bundle":"https://pith.science/.well-known/pith/2SQRZQWVLP3RE5QRKYYPLS5KIE/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:2SQRZQWVLP3RE5QRKYYPLS5KIE","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"636aa3e6dcc915b5ec565a8db37bd6a75c7ef2486ee07309547e27d472f16e3a","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-14T04:15:30Z","title_canon_sha256":"325b159300e80cf7f2e654225d65059c4c87a20c84cfd4e299bf50ee7c414895"},"schema_version":"1.0","source":{"id":"2605.14347","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.14347","created_at":"2026-05-17T23:39:08Z"},{"alias_kind":"arxiv_version","alias_value":"2605.14347v1","created_at":"2026-05-17T23:39:08Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14347","created_at":"2026-05-17T23:39:08Z"},{"alias_kind":"pith_short_12","alias_value":"2SQRZQWVLP3R","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"2SQRZQWVLP3RE5QR","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"2SQRZQWV","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:5c57e18b43828d8ebe55264cdca8faa02ff31b33be0f9bdd30b54ce9c4a9ce1c","target":"graph","created_at":"2026-05-17T23:39:08Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Exemplar Partitioning constructs feature dictionaries for language model activations by clustering around observed exemplars, achieving near-SAE performance at much lower computational cost."}],"snapshot_sha256":"84c514c3d677e31c24932d9295469db9eea71cec33e28e05a1aee76929ee2f3b"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"ceb69e9c02c0c9c6389da959aed2f52ca38fde996724edf8b72b695460796fbe"},"paper":{"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","authors_text":"Jessica Rumbelow","cross_cats":[],"headline":"Exemplar Partitioning constructs feature dictionaries for language model activations by clustering around observed exemplars, achieving near-SAE performance at much lower computational cost.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-14T04:15:30Z","title":"Exemplar Partitioning for Mechanistic Interpretability"},"references":{"count":39,"internal_anchors":0,"resolved_work":39,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"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","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Toy Models of Superposition , author=. 2022 , note=","work_id":"b05a43bb-21aa-44ff-8424-df38240a48cf","year":2022},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models , author=. 2023 , eprint=","work_id":"b0500e0d-343b-4ae2-8b3b-080070d36a77","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Towards Monosemanticity: Decomposing Language Models With Dictionary Learning , author=. 2023 , note=","work_id":"1231d466-a836-4494-a765-2e69c9918681","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet , author=. 2024 , note=","work_id":"bf57d145-a267-443c-823a-e0c0f0adab66","year":2024}],"snapshot_sha256":"428d10ed8bec878cb3741dcebb2734f4b5db5c5d3276c66845a1ac1e7056e392"},"source":{"id":"2605.14347","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-15T01:30:59.455904Z","id":"6dda7827-0efe-46b3-ae2e-b9e7e178b789","model_set":{"reader":"grok-4.3"},"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","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.","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.","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."}},"verdict_id":"6dda7827-0efe-46b3-ae2e-b9e7e178b789"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:e04de711c2a68e8ca813a0865c8279b8d330461233a223682bd18c1c7d300541","target":"record","created_at":"2026-05-17T23:39:08Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"636aa3e6dcc915b5ec565a8db37bd6a75c7ef2486ee07309547e27d472f16e3a","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-14T04:15:30Z","title_canon_sha256":"325b159300e80cf7f2e654225d65059c4c87a20c84cfd4e299bf50ee7c414895"},"schema_version":"1.0","source":{"id":"2605.14347","kind":"arxiv","version":1}},"canonical_sha256":"d4a11cc2d55bf71276115630f5cbaa4115dd4ff6e69d2c2f6875ad656d7711a9","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"d4a11cc2d55bf71276115630f5cbaa4115dd4ff6e69d2c2f6875ad656d7711a9","first_computed_at":"2026-05-17T23:39:08.115573Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:39:08.115573Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"OishU94yVyU9rqjy84Ja6HH298sUpjSvs038Gd9Enw1xxx5DAxv7FytbOq8GWhEcOlrsWwI5Q2wwKjvsBblCDA==","signature_status":"signed_v1","signed_at":"2026-05-17T23:39:08.116256Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.14347","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:e04de711c2a68e8ca813a0865c8279b8d330461233a223682bd18c1c7d300541","sha256:5c57e18b43828d8ebe55264cdca8faa02ff31b33be0f9bdd30b54ce9c4a9ce1c"],"state_sha256":"680ed39585820c1a36d80844823a0d2c3b00c1cb9d039b3ef1505a7d620822e4"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"EPD1Z9ZfP3Wyiz9ngbmYIq24trmocQRZqlyeuYrLqSEZWA+Ca3/hR8ictCtd3wkBLUPm54rzo8sEQ1igqTwbDg==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-25T15:01:05.200654Z","bundle_sha256":"fdf636da3153a6e74bd504e7e2f062281d2132481e12a6285f0c15324883d217"}}