{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:INK6UH3LZ2YHYAWKIGQZHE4PNW","short_pith_number":"pith:INK6UH3L","schema_version":"1.0","canonical_sha256":"4355ea1f6bceb07c02ca41a193938f6d9393a84dda7825e6496723b0e522cc79","source":{"kind":"arxiv","id":"2505.17769","version":2},"attestation_state":"computed","paper":{"title":"Inference-Time Decomposition of Activations (ITDA): A Scalable Approach to Interpreting Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Neel Nanda, Noura Al Moubayed, Patrick Leask","submitted_at":"2025-05-23T11:41:41Z","abstract_excerpt":"Sparse autoencoders (SAEs) are a popular method for decomposing Large Langage Models (LLM) activations into interpretable latents. However, due to their substantial training cost, most academic research uses open-source SAEs which are only available for a restricted set of models of up to 27B parameters. SAE latents are also learned from a dataset of activations, which means they do not transfer between models. Motivated by relative representation similarity measures, we introduce Inference-Time Decomposition of Activations (ITDA) models, an alternative method for decomposing language model ac"},"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":"2505.17769","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-05-23T11:41:41Z","cross_cats_sorted":[],"title_canon_sha256":"769b677372b54a618e12c604334cf1b0dbf447027808e04f1590ab6f54966e9b","abstract_canon_sha256":"e59e528737749b37119c7a8091de2bd64f1fe69cbd6146345d72ee7c27a2d9be"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:20:12.356479Z","signature_b64":"XKQ37yUx4Z6hUy35qQF4FNYURx0W5A9ymAZo8sTh4MnuYLxNOOVLUYn8NAmJaMWRCd81iViP3chwrC9YB/bkAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4355ea1f6bceb07c02ca41a193938f6d9393a84dda7825e6496723b0e522cc79","last_reissued_at":"2026-07-05T11:20:12.356045Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:20:12.356045Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Inference-Time Decomposition of Activations (ITDA): A Scalable Approach to Interpreting Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Neel Nanda, Noura Al Moubayed, Patrick Leask","submitted_at":"2025-05-23T11:41:41Z","abstract_excerpt":"Sparse autoencoders (SAEs) are a popular method for decomposing Large Langage Models (LLM) activations into interpretable latents. However, due to their substantial training cost, most academic research uses open-source SAEs which are only available for a restricted set of models of up to 27B parameters. SAE latents are also learned from a dataset of activations, which means they do not transfer between models. Motivated by relative representation similarity measures, we introduce Inference-Time Decomposition of Activations (ITDA) models, an alternative method for decomposing language model ac"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.17769","kind":"arxiv","version":2},"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/2505.17769/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":"2505.17769","created_at":"2026-07-05T11:20:12.356095+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.17769v2","created_at":"2026-07-05T11:20:12.356095+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.17769","created_at":"2026-07-05T11:20:12.356095+00:00"},{"alias_kind":"pith_short_12","alias_value":"INK6UH3LZ2YH","created_at":"2026-07-05T11:20:12.356095+00:00"},{"alias_kind":"pith_short_16","alias_value":"INK6UH3LZ2YHYAWK","created_at":"2026-07-05T11:20:12.356095+00:00"},{"alias_kind":"pith_short_8","alias_value":"INK6UH3L","created_at":"2026-07-05T11:20:12.356095+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.11722","citing_title":"ICA Lens: Interpreting Language Models Without Training Another Dictionary","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06333","citing_title":"Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability","ref_index":31,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/INK6UH3LZ2YHYAWKIGQZHE4PNW","json":"https://pith.science/pith/INK6UH3LZ2YHYAWKIGQZHE4PNW.json","graph_json":"https://pith.science/api/pith-number/INK6UH3LZ2YHYAWKIGQZHE4PNW/graph.json","events_json":"https://pith.science/api/pith-number/INK6UH3LZ2YHYAWKIGQZHE4PNW/events.json","paper":"https://pith.science/paper/INK6UH3L"},"agent_actions":{"view_html":"https://pith.science/pith/INK6UH3LZ2YHYAWKIGQZHE4PNW","download_json":"https://pith.science/pith/INK6UH3LZ2YHYAWKIGQZHE4PNW.json","view_paper":"https://pith.science/paper/INK6UH3L","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.17769&json=true","fetch_graph":"https://pith.science/api/pith-number/INK6UH3LZ2YHYAWKIGQZHE4PNW/graph.json","fetch_events":"https://pith.science/api/pith-number/INK6UH3LZ2YHYAWKIGQZHE4PNW/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/INK6UH3LZ2YHYAWKIGQZHE4PNW/action/timestamp_anchor","attest_storage":"https://pith.science/pith/INK6UH3LZ2YHYAWKIGQZHE4PNW/action/storage_attestation","attest_author":"https://pith.science/pith/INK6UH3LZ2YHYAWKIGQZHE4PNW/action/author_attestation","sign_citation":"https://pith.science/pith/INK6UH3LZ2YHYAWKIGQZHE4PNW/action/citation_signature","submit_replication":"https://pith.science/pith/INK6UH3LZ2YHYAWKIGQZHE4PNW/action/replication_record"}},"created_at":"2026-07-05T11:20:12.356095+00:00","updated_at":"2026-07-05T11:20:12.356095+00:00"}