{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:KDPB7COALUCQHG2XI4EMMWPISE","short_pith_number":"pith:KDPB7COA","schema_version":"1.0","canonical_sha256":"50de1f89c05d05039b574708c659e89120d4f19409f9ee45ac890b2f319f3913","source":{"kind":"arxiv","id":"2605.13190","version":1},"attestation_state":"computed","paper":{"title":"N-vium: Mixture-of-Exits Transformer for Accelerated Exact Generation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"N-vium mixture-of-exits transformers reach 57.9 percent wall-clock speedup at identical perplexity by routing tokens to multiple exit depths","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Aleksander Lorenc, Fr\\'ed\\'eric Berdoz, Jo\\\"el Mathys, Roger Wattenhofer","submitted_at":"2026-05-13T08:46:17Z","abstract_excerpt":"Improving the inference efficiency of autoregressive transformers typically means reducing FLOPs per token, usually through approximations that degrade model quality. We introduce N-vium, a mixture-of-exits transformer that partially parallelizes computation across depth on standard hardware, increasing effective FLOPs per second rather than minimizing compute per token. N-vium attaches prediction heads at multiple depths and defines the next-token distribution as a learned mixture over these exits, with token-adaptive routing. This formulation strictly generalizes the standard transformer, wh"},"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":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.13190","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T08:46:17Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"3d9415796db25435ae41ab4a4cd2e87e32e056dc649fabfd0cbc9e28ed2ec808","abstract_canon_sha256":"1e1ac88a055099cfdb6e12170cef87de7ecc32b4cb9e7972ff004652fd01905c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:08:56.146346Z","signature_b64":"ngM4BfP6KcrC0FU+tUKsRv9abivxsrJKd1tsgJnjZKTKUM1V+Fjt1OwsEv7jZ2fF24OHFIwVF1MgYcgntw4vDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"50de1f89c05d05039b574708c659e89120d4f19409f9ee45ac890b2f319f3913","last_reissued_at":"2026-05-18T03:08:56.145858Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:08:56.145858Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"N-vium: Mixture-of-Exits Transformer for Accelerated Exact Generation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"N-vium mixture-of-exits transformers reach 57.9 percent wall-clock speedup at identical perplexity by routing tokens to multiple exit depths","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Aleksander Lorenc, Fr\\'ed\\'eric Berdoz, Jo\\\"el Mathys, Roger Wattenhofer","submitted_at":"2026-05-13T08:46:17Z","abstract_excerpt":"Improving the inference efficiency of autoregressive transformers typically means reducing FLOPs per token, usually through approximations that degrade model quality. We introduce N-vium, a mixture-of-exits transformer that partially parallelizes computation across depth on standard hardware, increasing effective FLOPs per second rather than minimizing compute per token. N-vium attaches prediction heads at multiple depths and defines the next-token distribution as a learned mixture over these exits, with token-adaptive routing. This formulation strictly generalizes the standard transformer, wh"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our largest model reaches 57.9% wall-clock speedup over a parameter- and data-matched standard transformer at no perplexity cost.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a learned mixture over multiple exit heads can be trained to exactly match the perplexity of the full-depth model while enabling the described parallelization, cache deferral, and exact sampling without introducing any quality degradation or training instability.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"N-vium achieves 57.9% wall-clock speedup over matched standard transformers at no perplexity cost by mixing exact predictions from multiple model depths.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"N-vium mixture-of-exits transformers reach 57.9 percent wall-clock speedup at identical perplexity by routing tokens to multiple exit depths","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c6adb58ca35c25090443e66134e4f5214794d054ae10be646ac172f0f1dcf1f7"},"source":{"id":"2605.13190","kind":"arxiv","version":1},"verdict":{"id":"c566708f-2c54-4ff9-b578-4a0269d974e9","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:41:42.908621Z","strongest_claim":"Our largest model reaches 57.9% wall-clock speedup over a parameter- and data-matched standard transformer at no perplexity cost.","one_line_summary":"N-vium achieves 57.9% wall-clock speedup over matched standard transformers at no perplexity cost by mixing exact predictions from multiple model depths.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a learned mixture over multiple exit heads can be trained to exactly match the perplexity of the full-depth model while enabling the described parallelization, cache deferral, and exact sampling without introducing any quality degradation or training instability.","pith_extraction_headline":"N-vium mixture-of-exits transformers reach 57.9 percent wall-clock speedup at identical perplexity by routing tokens to multiple exit depths"},"references":{"count":63,"sample":[{"doi":"","year":2023,"title":"J. Ainslie, J. Lee-Thorp, M. de Jong, Y. Zemlyanskiy, F. Lebron, and S. Sanghai. GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints. InProceedings of the 2023 Confere","work_id":"8d75ffc6-6b32-4f09-a0d7-622da944b51a","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"K. Alizadeh-Vahid, S. I. Mirzadeh, H. Shahrkokhi, D. Belenko, F. Sun, M. Cho, et al. Duo-LLM: A Framework for Studying Adaptive Computation in Large Language Models. InProceedings of The 4th NeurIPS E","work_id":"fb9d5c11-d742-45f8-b635-c30f370574bc","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"S. Bae, J. Ko, H. Song, and S.-Y. Yun. Fast and Robust Early-Exiting Framework for Autoregressive Language Models with Synchronized Parallel Decoding. InProceedings of the 2023 Conference on Empirical","work_id":"385ed1b9-f771-4a32-957d-96547fd14316","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"S. Bae, Y. Kim, R. Bayat, S. Kim, J. Ha, T. Schuster, et al. Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation. InAd- vances in Neural Information Processin","work_id":"a7a47611-841e-4e51-9b0d-0c5510e1ae2b","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"H. Bai, W. Zhang, L. Hou, L. Shang, J. Jin, X. Jiang, et al. BinaryBERT: Pushing the Limit of BERT Quantization. InProceedings of the 59th Annual Meeting of the Association for Computational Linguisti","work_id":"daa1eaf0-b504-40fb-9f5c-c4a19a18d551","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":63,"snapshot_sha256":"c089042d6bfef7ab1de35ed41cd38f81c875d1033d54b64bba10bc25fb3db27d","internal_anchors":11},"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":"2605.13190","created_at":"2026-05-18T03:08:56.145929+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.13190v1","created_at":"2026-05-18T03:08:56.145929+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13190","created_at":"2026-05-18T03:08:56.145929+00:00"},{"alias_kind":"pith_short_12","alias_value":"KDPB7COALUCQ","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"KDPB7COALUCQHG2X","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"KDPB7COA","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/KDPB7COALUCQHG2XI4EMMWPISE","json":"https://pith.science/pith/KDPB7COALUCQHG2XI4EMMWPISE.json","graph_json":"https://pith.science/api/pith-number/KDPB7COALUCQHG2XI4EMMWPISE/graph.json","events_json":"https://pith.science/api/pith-number/KDPB7COALUCQHG2XI4EMMWPISE/events.json","paper":"https://pith.science/paper/KDPB7COA"},"agent_actions":{"view_html":"https://pith.science/pith/KDPB7COALUCQHG2XI4EMMWPISE","download_json":"https://pith.science/pith/KDPB7COALUCQHG2XI4EMMWPISE.json","view_paper":"https://pith.science/paper/KDPB7COA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.13190&json=true","fetch_graph":"https://pith.science/api/pith-number/KDPB7COALUCQHG2XI4EMMWPISE/graph.json","fetch_events":"https://pith.science/api/pith-number/KDPB7COALUCQHG2XI4EMMWPISE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/KDPB7COALUCQHG2XI4EMMWPISE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/KDPB7COALUCQHG2XI4EMMWPISE/action/storage_attestation","attest_author":"https://pith.science/pith/KDPB7COALUCQHG2XI4EMMWPISE/action/author_attestation","sign_citation":"https://pith.science/pith/KDPB7COALUCQHG2XI4EMMWPISE/action/citation_signature","submit_replication":"https://pith.science/pith/KDPB7COALUCQHG2XI4EMMWPISE/action/replication_record"}},"created_at":"2026-05-18T03:08:56.145929+00:00","updated_at":"2026-05-18T03:08:56.145929+00:00"}