{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:64RTIIKSYJ6CYRNQXZLNL7TJ73","short_pith_number":"pith:64RTIIKS","schema_version":"1.0","canonical_sha256":"f723342152c27c2c45b0be56d5fe69fee84ef3d6443cab97931902a915612a45","source":{"kind":"arxiv","id":"2403.17887","version":2},"attestation_state":"computed","paper":{"title":"The Unreasonable Ineffectiveness of the Deeper Layers","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG","stat.ML"],"primary_cat":"cs.CL","authors_text":"Andrey Gromov, Daniel A. Roberts, Hassan Shapourian, Kushal Tirumala, Paolo Glorioso","submitted_at":"2024-03-26T17:20:04Z","abstract_excerpt":"How is knowledge stored in an LLM's weights? We study this via layer pruning: if removing a certain layer does not affect model performance in common question-answering benchmarks, then the weights in that layer are not necessary for storing the knowledge needed to answer those questions. To find these unnecessary parameters, we identify the optimal block of layers to prune by considering similarity across layers; then, to \"heal\" the damage, we perform a small amount of finetuning. Surprisingly, with this method we find minimal degradation of performance until after a large fraction (up to hal"},"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":"2403.17887","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-03-26T17:20:04Z","cross_cats_sorted":["cs.LG","stat.ML"],"title_canon_sha256":"b2dccd928376da48da2ac31aed1fde3b73a643048bbc49b61c5e3478e65c9fd6","abstract_canon_sha256":"d9a6487532f887a1ee439b3120d276572168f471eab974789443bc247f849e02"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:22:45.128356Z","signature_b64":"nfW034rxymD1UViFJBERHwh99/zkrcE86vJalB34lDCxbsegXBuSObR920PkS/VaVw/wwq8j5k8JO7MHLxhTDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f723342152c27c2c45b0be56d5fe69fee84ef3d6443cab97931902a915612a45","last_reissued_at":"2026-07-05T10:22:45.127577Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:22:45.127577Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Unreasonable Ineffectiveness of the Deeper Layers","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG","stat.ML"],"primary_cat":"cs.CL","authors_text":"Andrey Gromov, Daniel A. Roberts, Hassan Shapourian, Kushal Tirumala, Paolo Glorioso","submitted_at":"2024-03-26T17:20:04Z","abstract_excerpt":"How is knowledge stored in an LLM's weights? We study this via layer pruning: if removing a certain layer does not affect model performance in common question-answering benchmarks, then the weights in that layer are not necessary for storing the knowledge needed to answer those questions. To find these unnecessary parameters, we identify the optimal block of layers to prune by considering similarity across layers; then, to \"heal\" the damage, we perform a small amount of finetuning. Surprisingly, with this method we find minimal degradation of performance until after a large fraction (up to hal"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2403.17887","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/2403.17887/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":"2403.17887","created_at":"2026-07-05T10:22:45.127669+00:00"},{"alias_kind":"arxiv_version","alias_value":"2403.17887v2","created_at":"2026-07-05T10:22:45.127669+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2403.17887","created_at":"2026-07-05T10:22:45.127669+00:00"},{"alias_kind":"pith_short_12","alias_value":"64RTIIKSYJ6C","created_at":"2026-07-05T10:22:45.127669+00:00"},{"alias_kind":"pith_short_16","alias_value":"64RTIIKSYJ6CYRNQ","created_at":"2026-07-05T10:22:45.127669+00:00"},{"alias_kind":"pith_short_8","alias_value":"64RTIIKS","created_at":"2026-07-05T10:22:45.127669+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":17,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.25008","citing_title":"Neural Scaling Universality: If Exponents Are Fixed, Time to Understand Coefficients","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2606.23670","citing_title":"Tapered Language Models","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2606.20246","citing_title":"Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think","ref_index":39,"is_internal_anchor":false},{"citing_arxiv_id":"2606.19150","citing_title":"Complementary Attention Head Pruning for Efficient Transformers","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2606.31796","citing_title":"CHERRY: Compressed Hierarchical Experts with Recurrent Representational Yield","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2606.30813","citing_title":"Gradient Smoothing: Coupling Layer-wise Updates for Improved Optimization","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2605.25344","citing_title":"A general tensor-structured compression scheme for efficient large language models","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26496","citing_title":"Dense2MoE: Pushing the Pareto Frontier of On-Device LLMs via Unified Pruning and Upcycling","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2606.26538","citing_title":"CascadeFormer: Depth-Tapered Transformers Motivated by Gradient Fan-in Asymmetry","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2412.14590","citing_title":"MixLLM: LLM Quantization with Global Mixed-precision between Output-features and Highly-efficient System Design","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2603.15031","citing_title":"Attention Residuals","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2602.01997","citing_title":"On the Limits of Layer Pruning for Generative Reasoning in Large Language Models","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2605.04971","citing_title":"Why Geometric Continuity Emerges in Deep Neural Networks: Residual Connections and Rotational Symmetry Breaking","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2604.20682","citing_title":"Variance Is Not Importance: Structural Analysis of Transformer Compressibility Across Model Scales","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2604.12426","citing_title":"Do Transformers Use their Depth Adaptively? Evidence from a Relational Reasoning Task","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07271","citing_title":"Understanding Performance Collapse in Layer-Pruned Large Language Models via Decision Representation Transitions","ref_index":52,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17224","citing_title":"LASER: Low-Rank Activation SVD for Efficient Recursion","ref_index":3,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/64RTIIKSYJ6CYRNQXZLNL7TJ73","json":"https://pith.science/pith/64RTIIKSYJ6CYRNQXZLNL7TJ73.json","graph_json":"https://pith.science/api/pith-number/64RTIIKSYJ6CYRNQXZLNL7TJ73/graph.json","events_json":"https://pith.science/api/pith-number/64RTIIKSYJ6CYRNQXZLNL7TJ73/events.json","paper":"https://pith.science/paper/64RTIIKS"},"agent_actions":{"view_html":"https://pith.science/pith/64RTIIKSYJ6CYRNQXZLNL7TJ73","download_json":"https://pith.science/pith/64RTIIKSYJ6CYRNQXZLNL7TJ73.json","view_paper":"https://pith.science/paper/64RTIIKS","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2403.17887&json=true","fetch_graph":"https://pith.science/api/pith-number/64RTIIKSYJ6CYRNQXZLNL7TJ73/graph.json","fetch_events":"https://pith.science/api/pith-number/64RTIIKSYJ6CYRNQXZLNL7TJ73/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/64RTIIKSYJ6CYRNQXZLNL7TJ73/action/timestamp_anchor","attest_storage":"https://pith.science/pith/64RTIIKSYJ6CYRNQXZLNL7TJ73/action/storage_attestation","attest_author":"https://pith.science/pith/64RTIIKSYJ6CYRNQXZLNL7TJ73/action/author_attestation","sign_citation":"https://pith.science/pith/64RTIIKSYJ6CYRNQXZLNL7TJ73/action/citation_signature","submit_replication":"https://pith.science/pith/64RTIIKSYJ6CYRNQXZLNL7TJ73/action/replication_record"}},"created_at":"2026-07-05T10:22:45.127669+00:00","updated_at":"2026-07-05T10:22:45.127669+00:00"}