{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:FBM372NADXDRALWOQEOMAXRNEF","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":"3c82a4cd4a13228faa8e3f8f0452df549492de1680669473df93427cd848dcde","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-14T01:57:47Z","title_canon_sha256":"33c4e39637b4370251316a412bb40e8ff8a91afb48e987994e2aaed1d6ea0104"},"schema_version":"1.0","source":{"id":"2605.14258","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.14258","created_at":"2026-05-17T23:39:10Z"},{"alias_kind":"arxiv_version","alias_value":"2605.14258v1","created_at":"2026-05-17T23:39:10Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14258","created_at":"2026-05-17T23:39:10Z"},{"alias_kind":"pith_short_12","alias_value":"FBM372NADXDR","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"FBM372NADXDRALWO","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"FBM372NA","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:0dffb7d81d9f20a90203c2942d5325ed33190dd778b6a6cf0eb8ec9617938f76","target":"graph","created_at":"2026-05-17T23:39:10Z","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":"training installs a monotonic spectral gradient through depth -- from non-normal, rotation-dominated early layers to near-symmetric late layers -- together with a cumulative low-rank bottleneck that funnels perturbations into a small fraction of the residual stream's effective dimensions. ... the topological positioning of graph communities predicts whether the Jacobian amplifies or suppresses them, with the sign of the coupling determined by the local operator type, a relationship absent at initialization."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the local linearization given by the Jacobian at each layer remains a faithful description of perturbation propagation even though the actual layer update is nonlinear, and that the chosen graph-community detection procedure yields communities whose functional role is independent of the Jacobian analysis itself."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Training installs a depth-dependent spectral gradient and low-rank bottleneck in LLM residual streams whose amplification or suppression of graph communities is predicted by local operator type."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Training installs a monotonic spectral gradient in LLMs from non-normal early layers to near-symmetric late layers, creating a low-rank bottleneck for perturbations."}],"snapshot_sha256":"d0b63f5b93407b2658d6600e22a504a1c59749e78e5c2faee2a9541dd2ce7c71"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"abstract_excerpt":"Large language models are remarkably capable, yet how computation propagates through their layers remains poorly understood. A growing line of work treats depth as discrete time and the residual stream as a dynamical system, where each layer's nonlinear update has a local linear description. However, previous analyses have relied on scalar summaries or approximate linearizations, leaving the full spectral geometry of trained LLMs unknown. We perform full Jacobian eigendecomposition across three production--scale LLMs and show that training installs a monotonic spectral gradient through depth -","authors_text":"Grigori Guitchounts, Jesseba Fernando","cross_cats":["cs.AI"],"headline":"Training installs a monotonic spectral gradient in LLMs from non-normal early layers to near-symmetric late layers, creating a low-rank bottleneck for perturbations.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-14T01:57:47Z","title":"Dynamics of the Transformer Residual Stream: Coupling Spectral Geometry to Network Topology"},"references":{"count":36,"internal_anchors":9,"resolved_work":36,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Dubey, Abhimanyu and Jauhri, Abhinav and Pandey, Abhinav and Kadian, Abhishek and Al-Dahle, Ahmad and Letman, Aiesha and Mathur, Akhil and Schelten, Alan and Yang, Amy and Fan, Angela and others , jou","work_id":"4b2a0186-c896-47b0-90bc-8ca5ed406e43","year":2024},{"cited_arxiv_id":"2512.13961","doi":"","is_internal_anchor":true,"ref_index":2,"title":"arXiv preprint arXiv:2512.13961 , year =","work_id":"74de5f5e-0a69-4f73-862d-e5705fa9f4bb","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"International Conference on Learning Representations (ICLR) , year =","work_id":"9977b5a1-8392-4548-86a2-6df546b46a4e","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"and Waltman, Ludo and van Eck, Nees Jan , journal =","work_id":"7d365822-e548-4921-bfc3-38fb30081518","year":2019},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Advances in Neural Information Processing Systems (NeurIPS) , volume =","work_id":"938fdbe5-c940-4267-b842-a6c4dff1e574","year":null}],"snapshot_sha256":"608d536a6cb7c89072a54b30cc0713e12782038451060926ef2319bc66c6a1df"},"source":{"id":"2605.14258","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-15T02:34:23.335145Z","id":"c4564d46-f565-4deb-8fc6-49873088a306","model_set":{"reader":"grok-4.3"},"one_line_summary":"Training installs a depth-dependent spectral gradient and low-rank bottleneck in LLM residual streams whose amplification or suppression of graph communities is predicted by local operator type.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Training installs a monotonic spectral gradient in LLMs from non-normal early layers to near-symmetric late layers, creating a low-rank bottleneck for perturbations.","strongest_claim":"training installs a monotonic spectral gradient through depth -- from non-normal, rotation-dominated early layers to near-symmetric late layers -- together with a cumulative low-rank bottleneck that funnels perturbations into a small fraction of the residual stream's effective dimensions. ... the topological positioning of graph communities predicts whether the Jacobian amplifies or suppresses them, with the sign of the coupling determined by the local operator type, a relationship absent at initialization.","weakest_assumption":"That the local linearization given by the Jacobian at each layer remains a faithful description of perturbation propagation even though the actual layer update is nonlinear, and that the chosen graph-community detection procedure yields communities whose functional role is independent of the Jacobian analysis itself."}},"verdict_id":"c4564d46-f565-4deb-8fc6-49873088a306"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:5cd0792e6d78f1d04d9c7a8fa1711bbae357120c17f62b8560487732edc5f3d1","target":"record","created_at":"2026-05-17T23:39:10Z","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":"3c82a4cd4a13228faa8e3f8f0452df549492de1680669473df93427cd848dcde","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-14T01:57:47Z","title_canon_sha256":"33c4e39637b4370251316a412bb40e8ff8a91afb48e987994e2aaed1d6ea0104"},"schema_version":"1.0","source":{"id":"2605.14258","kind":"arxiv","version":1}},"canonical_sha256":"2859bfe9a01dc7102ece811cc05e2d21771e876b088515806634f0e2f559c134","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"2859bfe9a01dc7102ece811cc05e2d21771e876b088515806634f0e2f559c134","first_computed_at":"2026-05-17T23:39:10.508262Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:39:10.508262Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"wl1fsUa0s84BoFp8kVhQHFSE/9OvINm1Nlq6e0qeyhpmm0w+i2LQGBwxx3FeuojpM36WGz/ZLJ39zZcHjsPUCg==","signature_status":"signed_v1","signed_at":"2026-05-17T23:39:10.508707Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.14258","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:5cd0792e6d78f1d04d9c7a8fa1711bbae357120c17f62b8560487732edc5f3d1","sha256:0dffb7d81d9f20a90203c2942d5325ed33190dd778b6a6cf0eb8ec9617938f76"],"state_sha256":"4a1367539666541fa1fe84b480137c39ae2b4926cc490194611738ad6c52487c"}