{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:WWZUVLJIXCSBYPGXKDDHRSRWBH","short_pith_number":"pith:WWZUVLJI","schema_version":"1.0","canonical_sha256":"b5b34aad28b8a41c3cd750c678ca3609f3c79d77c3cbe6794e8a6967d3b4c11f","source":{"kind":"arxiv","id":"2402.02622","version":2},"attestation_state":"computed","paper":{"title":"DenseFormer: Enhancing Information Flow in Transformers via Depth Weighted Averaging","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Amirkeivan Mohtashami, Francois Fleuret, Martin Jaggi, Matteo Pagliardini","submitted_at":"2024-02-04T21:44:09Z","abstract_excerpt":"The transformer architecture by Vaswani et al. (2017) is now ubiquitous across application domains, from natural language processing to speech processing and image understanding. We propose DenseFormer, a simple modification to the standard architecture that improves the perplexity of the model without increasing its size -- adding a few thousand parameters for large-scale models in the 100B parameters range. Our approach relies on an additional averaging step after each transformer block, which computes a weighted average of current and past representations -- we refer to this operation as De"},"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":"2402.02622","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-02-04T21:44:09Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"4384b922d2aff2e4f63175de4baad71d627ba7c49863bdcc0a2b07dd37a81e33","abstract_canon_sha256":"ef659e5dc8b9658f73f50417b28e4657b8469675b5572ab66da2d60ca21e416a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:59:04.904723Z","signature_b64":"RXz65sjWuwrx+n3u2MSXDc61bGAjIWJCVthw+0pSRlUFATLlYyRgw7f2pDnBx5rvrf4W9exHXGrOpWT9GhnZDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b5b34aad28b8a41c3cd750c678ca3609f3c79d77c3cbe6794e8a6967d3b4c11f","last_reissued_at":"2026-07-05T07:59:04.904258Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:59:04.904258Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"DenseFormer: Enhancing Information Flow in Transformers via Depth Weighted Averaging","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Amirkeivan Mohtashami, Francois Fleuret, Martin Jaggi, Matteo Pagliardini","submitted_at":"2024-02-04T21:44:09Z","abstract_excerpt":"The transformer architecture by Vaswani et al. (2017) is now ubiquitous across application domains, from natural language processing to speech processing and image understanding. We propose DenseFormer, a simple modification to the standard architecture that improves the perplexity of the model without increasing its size -- adding a few thousand parameters for large-scale models in the 100B parameters range. Our approach relies on an additional averaging step after each transformer block, which computes a weighted average of current and past representations -- we refer to this operation as De"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.02622","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/2402.02622/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":"2402.02622","created_at":"2026-07-05T07:59:04.904322+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.02622v2","created_at":"2026-07-05T07:59:04.904322+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.02622","created_at":"2026-07-05T07:59:04.904322+00:00"},{"alias_kind":"pith_short_12","alias_value":"WWZUVLJIXCSB","created_at":"2026-07-05T07:59:04.904322+00:00"},{"alias_kind":"pith_short_16","alias_value":"WWZUVLJIXCSBYPGX","created_at":"2026-07-05T07:59:04.904322+00:00"},{"alias_kind":"pith_short_8","alias_value":"WWZUVLJI","created_at":"2026-07-05T07:59:04.904322+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.02780","citing_title":"Do Value Vectors in Deep Layers Need Context from the Residual Stream?","ref_index":89,"is_internal_anchor":false},{"citing_arxiv_id":"2603.15031","citing_title":"Attention Residuals","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20798","citing_title":"Most Transformer Modifications Still Do Not Transfer at 1-3B: A 2020-2026 Update to Narang et al. (2021) with Downstream Evaluation and a Noise Floor","ref_index":55,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18855","citing_title":"Delta Attention Residuals","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2604.21254","citing_title":"Hyperloop Transformers","ref_index":16,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/WWZUVLJIXCSBYPGXKDDHRSRWBH","json":"https://pith.science/pith/WWZUVLJIXCSBYPGXKDDHRSRWBH.json","graph_json":"https://pith.science/api/pith-number/WWZUVLJIXCSBYPGXKDDHRSRWBH/graph.json","events_json":"https://pith.science/api/pith-number/WWZUVLJIXCSBYPGXKDDHRSRWBH/events.json","paper":"https://pith.science/paper/WWZUVLJI"},"agent_actions":{"view_html":"https://pith.science/pith/WWZUVLJIXCSBYPGXKDDHRSRWBH","download_json":"https://pith.science/pith/WWZUVLJIXCSBYPGXKDDHRSRWBH.json","view_paper":"https://pith.science/paper/WWZUVLJI","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.02622&json=true","fetch_graph":"https://pith.science/api/pith-number/WWZUVLJIXCSBYPGXKDDHRSRWBH/graph.json","fetch_events":"https://pith.science/api/pith-number/WWZUVLJIXCSBYPGXKDDHRSRWBH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/WWZUVLJIXCSBYPGXKDDHRSRWBH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/WWZUVLJIXCSBYPGXKDDHRSRWBH/action/storage_attestation","attest_author":"https://pith.science/pith/WWZUVLJIXCSBYPGXKDDHRSRWBH/action/author_attestation","sign_citation":"https://pith.science/pith/WWZUVLJIXCSBYPGXKDDHRSRWBH/action/citation_signature","submit_replication":"https://pith.science/pith/WWZUVLJIXCSBYPGXKDDHRSRWBH/action/replication_record"}},"created_at":"2026-07-05T07:59:04.904322+00:00","updated_at":"2026-07-05T07:59:04.904322+00:00"}