{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:KDPB7COALUCQHG2XI4EMMWPISE","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":"1e1ac88a055099cfdb6e12170cef87de7ecc32b4cb9e7972ff004652fd01905c","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T08:46:17Z","title_canon_sha256":"3d9415796db25435ae41ab4a4cd2e87e32e056dc649fabfd0cbc9e28ed2ec808"},"schema_version":"1.0","source":{"id":"2605.13190","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.13190","created_at":"2026-05-18T03:08:56Z"},{"alias_kind":"arxiv_version","alias_value":"2605.13190v1","created_at":"2026-05-18T03:08:56Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13190","created_at":"2026-05-18T03:08:56Z"},{"alias_kind":"pith_short_12","alias_value":"KDPB7COALUCQ","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"KDPB7COALUCQHG2X","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"KDPB7COA","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:4eff3f0dfc2e8c336c8afa97c23cfe1bc5cb82ef33ffec82d85a4d110f260f93","target":"graph","created_at":"2026-05-18T03:08:56Z","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 largest model reaches 57.9% wall-clock speedup over a parameter- and data-matched standard transformer at no perplexity cost."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"N-vium mixture-of-exits transformers reach 57.9 percent wall-clock speedup at identical perplexity by routing tokens to multiple exit depths"}],"snapshot_sha256":"c6adb58ca35c25090443e66134e4f5214794d054ae10be646ac172f0f1dcf1f7"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"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","authors_text":"Aleksander Lorenc, Fr\\'ed\\'eric Berdoz, Jo\\\"el Mathys, Roger Wattenhofer","cross_cats":["cs.AI"],"headline":"N-vium mixture-of-exits transformers reach 57.9 percent wall-clock speedup at identical perplexity by routing tokens to multiple exit depths","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T08:46:17Z","title":"N-vium: Mixture-of-Exits Transformer for Accelerated Exact Generation"},"references":{"count":63,"internal_anchors":11,"resolved_work":63,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"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","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"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","year":2024},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"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","year":2023},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"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","year":2025},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"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","year":2021}],"snapshot_sha256":"c089042d6bfef7ab1de35ed41cd38f81c875d1033d54b64bba10bc25fb3db27d"},"source":{"id":"2605.13190","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-14T20:41:42.908621Z","id":"c566708f-2c54-4ff9-b578-4a0269d974e9","model_set":{"reader":"grok-4.3"},"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","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","strongest_claim":"Our largest model reaches 57.9% wall-clock speedup over a parameter- and data-matched standard transformer at no perplexity cost.","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."}},"verdict_id":"c566708f-2c54-4ff9-b578-4a0269d974e9"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:3cff119fe9c49c77d62952d646581e7ef77c97c9ab16e9c6bf445431f49cd28a","target":"record","created_at":"2026-05-18T03:08:56Z","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":"1e1ac88a055099cfdb6e12170cef87de7ecc32b4cb9e7972ff004652fd01905c","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T08:46:17Z","title_canon_sha256":"3d9415796db25435ae41ab4a4cd2e87e32e056dc649fabfd0cbc9e28ed2ec808"},"schema_version":"1.0","source":{"id":"2605.13190","kind":"arxiv","version":1}},"canonical_sha256":"50de1f89c05d05039b574708c659e89120d4f19409f9ee45ac890b2f319f3913","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"50de1f89c05d05039b574708c659e89120d4f19409f9ee45ac890b2f319f3913","first_computed_at":"2026-05-18T03:08:56.145858Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T03:08:56.145858Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"ngM4BfP6KcrC0FU+tUKsRv9abivxsrJKd1tsgJnjZKTKUM1V+Fjt1OwsEv7jZ2fF24OHFIwVF1MgYcgntw4vDQ==","signature_status":"signed_v1","signed_at":"2026-05-18T03:08:56.146346Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.13190","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:3cff119fe9c49c77d62952d646581e7ef77c97c9ab16e9c6bf445431f49cd28a","sha256:4eff3f0dfc2e8c336c8afa97c23cfe1bc5cb82ef33ffec82d85a4d110f260f93"],"state_sha256":"572e316c71d8c828e3010e3c0fcd731ec30145d6ac2a135d70b85395cdc57a17"}