{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:O6M274A44LUGGWHN2K3BP56J2B","short_pith_number":"pith:O6M274A4","canonical_record":{"source":{"id":"2605.14075","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T19:51:25Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"cce309cae85c94074707337ba9d8d89c0247c9aa05e900492fbcf668743cb6e2","abstract_canon_sha256":"1a2a2a80efea60ded40ab4e08481d98a95ec7b3f2a73a0fbd19be6c9d577819f"},"schema_version":"1.0"},"canonical_sha256":"7799aff01ce2e86358edd2b617f7c9d078ee5f7225d4c99a86ac22c2ead54a40","source":{"kind":"arxiv","id":"2605.14075","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.14075","created_at":"2026-05-17T23:39:12Z"},{"alias_kind":"arxiv_version","alias_value":"2605.14075v1","created_at":"2026-05-17T23:39:12Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14075","created_at":"2026-05-17T23:39:12Z"},{"alias_kind":"pith_short_12","alias_value":"O6M274A44LUG","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"O6M274A44LUGGWHN","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"O6M274A4","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:O6M274A44LUGGWHN2K3BP56J2B","target":"record","payload":{"canonical_record":{"source":{"id":"2605.14075","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T19:51:25Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"cce309cae85c94074707337ba9d8d89c0247c9aa05e900492fbcf668743cb6e2","abstract_canon_sha256":"1a2a2a80efea60ded40ab4e08481d98a95ec7b3f2a73a0fbd19be6c9d577819f"},"schema_version":"1.0"},"canonical_sha256":"7799aff01ce2e86358edd2b617f7c9d078ee5f7225d4c99a86ac22c2ead54a40","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:12.374497Z","signature_b64":"m8nhrjhinvdfe6ds5HPWCsTQ+M0D+cYPjnoCaJwhgYnIWds25d4kh3bEKr7Ev8ntNKSHAobfUCW7+6YPJuWdDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7799aff01ce2e86358edd2b617f7c9d078ee5f7225d4c99a86ac22c2ead54a40","last_reissued_at":"2026-05-17T23:39:12.373908Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:12.373908Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2605.14075","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":"2026-05-17T23:39:12Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"JBBkvZJBlh+uMJeg6CQcCzoOTn4hdJHpeuD/fRYhpPsyR5UlT/fce6qWxfwc4MHE+bZxasMibfAUezMsvaHrDQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-25T18:31:34.034827Z"},"content_sha256":"c6cae5f80d9b351065172ffa40030afadbca30463bc6063f60490834a4fb3e0d","schema_version":"1.0","event_id":"sha256:c6cae5f80d9b351065172ffa40030afadbca30463bc6063f60490834a4fb3e0d"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:O6M274A44LUGGWHN2K3BP56J2B","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Rethinking Layer Relevance in Large Language Models Beyond Cosine Similarity","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Cosine similarity can be arbitrarily low for a layer that is still essential to an LLM's performance.","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Andres Carvallo De Ferari, Christ Devia, Cristian Hinostroza, Denis Parra, Eugenio Herrera-Berg, Jorge F Silva, Rodrigo Toro Icarte","submitted_at":"2026-05-13T19:51:25Z","abstract_excerpt":"Large language models (LLMs) have revolutionized natural language processing. Understanding their internal mechanisms is crucial for developing more interpretable and optimized architectures. Mechanistic interpretability has led to the development of various methods for assessing layer relevance, with cosine similarity being a widely used tool in the field. On this work, we demonstrate that cosine similarity is a poor proxy for the actual performance degradation caused by layer removal. Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while sti"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while still being crucial to the model's performance. Empirical evidence confirms that the correlation between cosine similarity and actual performance degradation is often weak or moderate.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That removing a single layer and measuring accuracy drop on held-out tasks gives a faithful picture of that layer's contribution inside the intact model, without major compensatory effects from remaining layers.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Cosine similarity poorly predicts performance degradation from layer removal in LLMs, making direct accuracy-drop ablation a more reliable relevance metric.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Cosine similarity can be arbitrarily low for a layer that is still essential to an LLM's performance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"879e6f5afa5d12097b9f3dcc9a061b620e1ed965fb25c993a320a98621c5d324"},"source":{"id":"2605.14075","kind":"arxiv","version":1},"verdict":{"id":"581f27de-ba7a-49dd-902d-8ce3abd0c06d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T05:01:38.973870Z","strongest_claim":"Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while still being crucial to the model's performance. Empirical evidence confirms that the correlation between cosine similarity and actual performance degradation is often weak or moderate.","one_line_summary":"Cosine similarity poorly predicts performance degradation from layer removal in LLMs, making direct accuracy-drop ablation a more reliable relevance metric.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That removing a single layer and measuring accuracy drop on held-out tasks gives a faithful picture of that layer's contribution inside the intact model, without major compensatory effects from remaining layers.","pith_extraction_headline":"Cosine similarity can be arbitrarily low for a layer that is still essential to an LLM's performance."},"references":{"count":77,"sample":[{"doi":"","year":null,"title":"Advances in neural information processing systems , booktitle=NIPS17, year=","work_id":"56b7b024-8ba8-46f3-8640-d02f359647af","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"All bark and no bite: Rogue dimensions in transformer language models obscure representational quality , author=","work_id":"79b261d5-8dbe-4d1a-97de-3b0c488318de","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"A Survey on Transformers in Reinforcement Learning , author=. 2023 , journal=","work_id":"647752ce-367f-4a84-9714-affff33f36ad","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1992,"title":"Breakthroughs in statistics: Methodology and distribution , pages=","work_id":"0c98474c-5657-4f19-b191-a492f6cdd19b","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1937,"title":"Journal of the american statistical association , volume=","work_id":"88946fe3-5dc0-4ea4-bd02-93cfb0e267c2","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":77,"snapshot_sha256":"0698fed82a6a97b4f307d0b3d4e9a5a03c4205f331271ba0b621c8023866411a","internal_anchors":7},"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"},"verdict_id":"581f27de-ba7a-49dd-902d-8ce3abd0c06d"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:39:12Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"s39+K2Ck2w4kPg8tQqFQ0QScVcfdCL/Ygm6k6L1lu/jAAp8EzXnxptCVQeVN3XzPiXerp7iafFBvPAmoLbB4Bg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-25T18:31:34.035906Z"},"content_sha256":"a1160a27bc49b08c8d6ade89e24fda0bd04c583974be6d7dc2d2193ce718a803","schema_version":"1.0","event_id":"sha256:a1160a27bc49b08c8d6ade89e24fda0bd04c583974be6d7dc2d2193ce718a803"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/O6M274A44LUGGWHN2K3BP56J2B/bundle.json","state_url":"https://pith.science/pith/O6M274A44LUGGWHN2K3BP56J2B/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/O6M274A44LUGGWHN2K3BP56J2B/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":"2026-05-25T18:31:34Z","links":{"resolver":"https://pith.science/pith/O6M274A44LUGGWHN2K3BP56J2B","bundle":"https://pith.science/pith/O6M274A44LUGGWHN2K3BP56J2B/bundle.json","state":"https://pith.science/pith/O6M274A44LUGGWHN2K3BP56J2B/state.json","well_known_bundle":"https://pith.science/.well-known/pith/O6M274A44LUGGWHN2K3BP56J2B/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:O6M274A44LUGGWHN2K3BP56J2B","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":"1a2a2a80efea60ded40ab4e08481d98a95ec7b3f2a73a0fbd19be6c9d577819f","cross_cats_sorted":["cs.CL"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T19:51:25Z","title_canon_sha256":"cce309cae85c94074707337ba9d8d89c0247c9aa05e900492fbcf668743cb6e2"},"schema_version":"1.0","source":{"id":"2605.14075","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.14075","created_at":"2026-05-17T23:39:12Z"},{"alias_kind":"arxiv_version","alias_value":"2605.14075v1","created_at":"2026-05-17T23:39:12Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14075","created_at":"2026-05-17T23:39:12Z"},{"alias_kind":"pith_short_12","alias_value":"O6M274A44LUG","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"O6M274A44LUGGWHN","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"O6M274A4","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:a1160a27bc49b08c8d6ade89e24fda0bd04c583974be6d7dc2d2193ce718a803","target":"graph","created_at":"2026-05-17T23:39:12Z","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":"Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while still being crucial to the model's performance. Empirical evidence confirms that the correlation between cosine similarity and actual performance degradation is often weak or moderate."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That removing a single layer and measuring accuracy drop on held-out tasks gives a faithful picture of that layer's contribution inside the intact model, without major compensatory effects from remaining layers."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Cosine similarity poorly predicts performance degradation from layer removal in LLMs, making direct accuracy-drop ablation a more reliable relevance metric."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Cosine similarity can be arbitrarily low for a layer that is still essential to an LLM's performance."}],"snapshot_sha256":"879e6f5afa5d12097b9f3dcc9a061b620e1ed965fb25c993a320a98621c5d324"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"abstract_excerpt":"Large language models (LLMs) have revolutionized natural language processing. Understanding their internal mechanisms is crucial for developing more interpretable and optimized architectures. Mechanistic interpretability has led to the development of various methods for assessing layer relevance, with cosine similarity being a widely used tool in the field. On this work, we demonstrate that cosine similarity is a poor proxy for the actual performance degradation caused by layer removal. Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while sti","authors_text":"Andres Carvallo De Ferari, Christ Devia, Cristian Hinostroza, Denis Parra, Eugenio Herrera-Berg, Jorge F Silva, Rodrigo Toro Icarte","cross_cats":["cs.CL"],"headline":"Cosine similarity can be arbitrarily low for a layer that is still essential to an LLM's performance.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T19:51:25Z","title":"Rethinking Layer Relevance in Large Language Models Beyond Cosine Similarity"},"references":{"count":77,"internal_anchors":7,"resolved_work":77,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Advances in neural information processing systems , booktitle=NIPS17, year=","work_id":"56b7b024-8ba8-46f3-8640-d02f359647af","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"All bark and no bite: Rogue dimensions in transformer language models obscure representational quality , author=","work_id":"79b261d5-8dbe-4d1a-97de-3b0c488318de","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"A Survey on Transformers in Reinforcement Learning , author=. 2023 , journal=","work_id":"647752ce-367f-4a84-9714-affff33f36ad","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Breakthroughs in statistics: Methodology and distribution , pages=","work_id":"0c98474c-5657-4f19-b191-a492f6cdd19b","year":1992},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Journal of the american statistical association , volume=","work_id":"88946fe3-5dc0-4ea4-bd02-93cfb0e267c2","year":1937}],"snapshot_sha256":"0698fed82a6a97b4f307d0b3d4e9a5a03c4205f331271ba0b621c8023866411a"},"source":{"id":"2605.14075","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-15T05:01:38.973870Z","id":"581f27de-ba7a-49dd-902d-8ce3abd0c06d","model_set":{"reader":"grok-4.3"},"one_line_summary":"Cosine similarity poorly predicts performance degradation from layer removal in LLMs, making direct accuracy-drop ablation a more reliable relevance metric.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Cosine similarity can be arbitrarily low for a layer that is still essential to an LLM's performance.","strongest_claim":"Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while still being crucial to the model's performance. Empirical evidence confirms that the correlation between cosine similarity and actual performance degradation is often weak or moderate.","weakest_assumption":"That removing a single layer and measuring accuracy drop on held-out tasks gives a faithful picture of that layer's contribution inside the intact model, without major compensatory effects from remaining layers."}},"verdict_id":"581f27de-ba7a-49dd-902d-8ce3abd0c06d"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:c6cae5f80d9b351065172ffa40030afadbca30463bc6063f60490834a4fb3e0d","target":"record","created_at":"2026-05-17T23:39:12Z","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":"1a2a2a80efea60ded40ab4e08481d98a95ec7b3f2a73a0fbd19be6c9d577819f","cross_cats_sorted":["cs.CL"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T19:51:25Z","title_canon_sha256":"cce309cae85c94074707337ba9d8d89c0247c9aa05e900492fbcf668743cb6e2"},"schema_version":"1.0","source":{"id":"2605.14075","kind":"arxiv","version":1}},"canonical_sha256":"7799aff01ce2e86358edd2b617f7c9d078ee5f7225d4c99a86ac22c2ead54a40","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"7799aff01ce2e86358edd2b617f7c9d078ee5f7225d4c99a86ac22c2ead54a40","first_computed_at":"2026-05-17T23:39:12.373908Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:39:12.373908Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"m8nhrjhinvdfe6ds5HPWCsTQ+M0D+cYPjnoCaJwhgYnIWds25d4kh3bEKr7Ev8ntNKSHAobfUCW7+6YPJuWdDQ==","signature_status":"signed_v1","signed_at":"2026-05-17T23:39:12.374497Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.14075","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:c6cae5f80d9b351065172ffa40030afadbca30463bc6063f60490834a4fb3e0d","sha256:a1160a27bc49b08c8d6ade89e24fda0bd04c583974be6d7dc2d2193ce718a803"],"state_sha256":"232c4b37b84c2f4c5aaba39124853ac790cace52e14d2e745610ffa0af05d217"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"XNva2wwQ/N7LtUtr9Yqus/vWdx1Zgas8IRQndA85GTM/h6aMx2yY03o55m5f1J193/jto8oLRSSu93EEsbKgCQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-25T18:31:34.040458Z","bundle_sha256":"182554e3a4e2e5a364090e58c844a9d672a00136b7b767b21413172eab66dac7"}}