{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:25HTPJLGS2LDN42WLPYMLPYZCW","short_pith_number":"pith:25HTPJLG","schema_version":"1.0","canonical_sha256":"d74f37a566969636f3565bf0c5bf191587d8e60b3ba680bc60ac9c9197b32e27","source":{"kind":"arxiv","id":"2502.00921","version":2},"attestation_state":"computed","paper":{"title":"Blink of an eye: a simple theory for feature localization in generative models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Aayush Karan, Marvin Li, Sitan Chen","submitted_at":"2025-02-02T21:19:53Z","abstract_excerpt":"Large language models can exhibit unexpected behavior in the blink of an eye. In a recent computer use demo, a language model switched from coding to Googling pictures of Yellowstone, and these sudden shifts in behavior have also been observed in reasoning patterns and jailbreaks. This phenomenon is not unique to autoregressive models: in diffusion models, key features of the final output are decided in narrow ``critical windows'' of the generation process. In this work we develop a simple, unifying theory to explain this phenomenon using the formalism of stochastic localization samplers. We s"},"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":"2502.00921","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-02T21:19:53Z","cross_cats_sorted":[],"title_canon_sha256":"cd40a489cbf5beeed53f157dd2dc98bd630b11f295af162643fb2ab195eacd83","abstract_canon_sha256":"4ef35aade16b43d26517d5b1811ca2a2ab02355e2b8b16a50e486bb3741ed242"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:16:31.729566Z","signature_b64":"38JxgC9J82g57CBbbrUti/7jycW7lpqLH+bzRP14Gh3yG1GmjDycaHK16+ooqq4f6Meua+f3HjzUhmL45IPeBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d74f37a566969636f3565bf0c5bf191587d8e60b3ba680bc60ac9c9197b32e27","last_reissued_at":"2026-07-05T11:16:31.729040Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:16:31.729040Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Blink of an eye: a simple theory for feature localization in generative models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Aayush Karan, Marvin Li, Sitan Chen","submitted_at":"2025-02-02T21:19:53Z","abstract_excerpt":"Large language models can exhibit unexpected behavior in the blink of an eye. In a recent computer use demo, a language model switched from coding to Googling pictures of Yellowstone, and these sudden shifts in behavior have also been observed in reasoning patterns and jailbreaks. This phenomenon is not unique to autoregressive models: in diffusion models, key features of the final output are decided in narrow ``critical windows'' of the generation process. In this work we develop a simple, unifying theory to explain this phenomenon using the formalism of stochastic localization samplers. We s"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.00921","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/2502.00921/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":"2502.00921","created_at":"2026-07-05T11:16:31.729107+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.00921v2","created_at":"2026-07-05T11:16:31.729107+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.00921","created_at":"2026-07-05T11:16:31.729107+00:00"},{"alias_kind":"pith_short_12","alias_value":"25HTPJLGS2LD","created_at":"2026-07-05T11:16:31.729107+00:00"},{"alias_kind":"pith_short_16","alias_value":"25HTPJLGS2LDN42W","created_at":"2026-07-05T11:16:31.729107+00:00"},{"alias_kind":"pith_short_8","alias_value":"25HTPJLG","created_at":"2026-07-05T11:16:31.729107+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.30327","citing_title":"Reasoning with Sampling: Cutting at Decision Points","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2503.03206","citing_title":"An Analytical Theory of Spectral Bias in the Learning Dynamics of Diffusion Models","ref_index":39,"is_internal_anchor":false},{"citing_arxiv_id":"2508.06614","citing_title":"Local Diffusion Models and Phases of Data Distributions","ref_index":25,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/25HTPJLGS2LDN42WLPYMLPYZCW","json":"https://pith.science/pith/25HTPJLGS2LDN42WLPYMLPYZCW.json","graph_json":"https://pith.science/api/pith-number/25HTPJLGS2LDN42WLPYMLPYZCW/graph.json","events_json":"https://pith.science/api/pith-number/25HTPJLGS2LDN42WLPYMLPYZCW/events.json","paper":"https://pith.science/paper/25HTPJLG"},"agent_actions":{"view_html":"https://pith.science/pith/25HTPJLGS2LDN42WLPYMLPYZCW","download_json":"https://pith.science/pith/25HTPJLGS2LDN42WLPYMLPYZCW.json","view_paper":"https://pith.science/paper/25HTPJLG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.00921&json=true","fetch_graph":"https://pith.science/api/pith-number/25HTPJLGS2LDN42WLPYMLPYZCW/graph.json","fetch_events":"https://pith.science/api/pith-number/25HTPJLGS2LDN42WLPYMLPYZCW/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/25HTPJLGS2LDN42WLPYMLPYZCW/action/timestamp_anchor","attest_storage":"https://pith.science/pith/25HTPJLGS2LDN42WLPYMLPYZCW/action/storage_attestation","attest_author":"https://pith.science/pith/25HTPJLGS2LDN42WLPYMLPYZCW/action/author_attestation","sign_citation":"https://pith.science/pith/25HTPJLGS2LDN42WLPYMLPYZCW/action/citation_signature","submit_replication":"https://pith.science/pith/25HTPJLGS2LDN42WLPYMLPYZCW/action/replication_record"}},"created_at":"2026-07-05T11:16:31.729107+00:00","updated_at":"2026-07-05T11:16:31.729107+00:00"}