{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:AXFU3DUTR7SPUGLPK667UTVIY5","short_pith_number":"pith:AXFU3DUT","schema_version":"1.0","canonical_sha256":"05cb4d8e938fe4fa196f57bdfa4ea8c7598877576765f04649a8f2cca081e4fa","source":{"kind":"arxiv","id":"2404.02258","version":1},"attestation_state":"computed","paper":{"title":"Mixture-of-Depths: Dynamically allocating compute in transformer-based language models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Transformer language models can learn to dynamically allocate compute to select tokens at each layer.","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Adam Santoro, Blake Richards, David Raposo, Peter Conway Humphreys, Sam Ritter, Timothy Lillicrap","submitted_at":"2024-04-02T19:28:11Z","abstract_excerpt":"Transformer-based language models spread FLOPs uniformly across input sequences. In this work we demonstrate that transformers can instead learn to dynamically allocate FLOPs (or compute) to specific positions in a sequence, optimising the allocation along the sequence for different layers across the model depth. Our method enforces a total compute budget by capping the number of tokens ($k$) that can participate in the self-attention and MLP computations at a given layer. The tokens to be processed are determined by the network using a top-$k$ routing mechanism. Since $k$ is defined a priori,"},"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":true},"canonical_record":{"source":{"id":"2404.02258","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-04-02T19:28:11Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"ec4a74a99a0bfb9a2c6146a1ee7e2a6ddcd656c28232813ecf9f44bf5d9f3bf9","abstract_canon_sha256":"20196f2ef34c9194cb315bc7bc2d6b7e36cc76b23f47901f1dbbbc05b493a687"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:15.411217Z","signature_b64":"ItCiII7Q8FqCuUqiJ1JXdDFWooq6992J6vclKArOfUQ2XRKxuYghSFpANHgqa99LNAFHdWYj3JfR7H1n6muqCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"05cb4d8e938fe4fa196f57bdfa4ea8c7598877576765f04649a8f2cca081e4fa","last_reissued_at":"2026-05-17T23:38:15.410613Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:15.410613Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mixture-of-Depths: Dynamically allocating compute in transformer-based language models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Transformer language models can learn to dynamically allocate compute to select tokens at each layer.","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Adam Santoro, Blake Richards, David Raposo, Peter Conway Humphreys, Sam Ritter, Timothy Lillicrap","submitted_at":"2024-04-02T19:28:11Z","abstract_excerpt":"Transformer-based language models spread FLOPs uniformly across input sequences. In this work we demonstrate that transformers can instead learn to dynamically allocate FLOPs (or compute) to specific positions in a sequence, optimising the allocation along the sequence for different layers across the model depth. Our method enforces a total compute budget by capping the number of tokens ($k$) that can participate in the self-attention and MLP computations at a given layer. The tokens to be processed are determined by the network using a top-$k$ routing mechanism. Since $k$ is defined a priori,"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Not only do models trained in this way learn to dynamically allocate compute, they do so efficiently. These models match baseline performance for equivalent FLOPS and wall-clock times to train, but require a fraction of the FLOPs per forward pass, and can be upwards of 50% faster to step during post-training sampling.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that a learned top-k router can reliably identify which tokens merit full processing at each layer without degrading overall model capacity or introducing training instabilities, and that this holds across model scales and tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Mixture-of-Depths enables transformers to dynamically allocate compute by routing only the top-k tokens through each layer's full computations, matching baseline performance with a fraction of the FLOPs per forward pass and up to 50% faster sampling.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Transformer language models can learn to dynamically allocate compute to select tokens at each layer.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2487a200c9eb271af365818e08fa621e4c53c12a8d37f2ee982a5b3b838e1b76"},"source":{"id":"2404.02258","kind":"arxiv","version":1},"verdict":{"id":"47fc0ddc-7aad-4707-9d3b-017bc1794454","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T02:13:07.402130Z","strongest_claim":"Not only do models trained in this way learn to dynamically allocate compute, they do so efficiently. These models match baseline performance for equivalent FLOPS and wall-clock times to train, but require a fraction of the FLOPs per forward pass, and can be upwards of 50% faster to step during post-training sampling.","one_line_summary":"Mixture-of-Depths enables transformers to dynamically allocate compute by routing only the top-k tokens through each layer's full computations, matching baseline performance with a fraction of the FLOPs per forward pass and up to 50% faster sampling.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that a learned top-k router can reliably identify which tokens merit full processing at each layer without degrading overall model capacity or introducing training instabilities, and that this holds across model scales and tasks.","pith_extraction_headline":"Transformer language models can learn to dynamically allocate compute to select tokens at each layer."},"references":{"count":11,"sample":[{"doi":"","year":2002,"title":"Controlling computation versus quality for neural sequence models","work_id":"21f2a3df-f950-438d-8553-6db49dc7afe8","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Universal Transformers","work_id":"8e5baefe-d209-411c-aefc-5acaa9275c8a","ref_index":3,"cited_arxiv_id":"1807.03819","is_internal_anchor":true},{"doi":"","year":1910,"title":"Depth-adaptive transformer","work_id":"6d994f34-ab13-423e-96ec-52b63f47a6d6","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Adaptive Computation Time for Recurrent Neural Networks","work_id":"75565443-173e-479c-b0e7-d2464e7630be","ref_index":7,"cited_arxiv_id":"1603.08983","is_internal_anchor":true},{"doi":"","year":null,"title":"Towards a unified view of parameter-efficient transfer learning","work_id":"e758988d-4ae2-44ff-94fc-4417c242b1e4","ref_index":8,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":11,"snapshot_sha256":"eb3b74aaf8e7de70ad3073dbc0bfbc87e51628e687adfc5c29387dbc73af8d74","internal_anchors":5},"formal_canon":{"evidence_count":1,"snapshot_sha256":"740ed61cf7e8970d28f0d74fd67712b045d4b6754137d9ae6d7eec1875379ca6"},"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":"2404.02258","created_at":"2026-05-17T23:38:15.410725+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.02258v1","created_at":"2026-05-17T23:38:15.410725+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.02258","created_at":"2026-05-17T23:38:15.410725+00:00"},{"alias_kind":"pith_short_12","alias_value":"AXFU3DUTR7SP","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"AXFU3DUTR7SPUGLP","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"AXFU3DUT","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":19,"internal_anchor_count":19,"sample":[{"citing_arxiv_id":"2411.04996","citing_title":"Mixture-of-Transformers: A Sparse and Scalable Architecture for Multi-Modal Foundation Models","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2512.13751","citing_title":"MIDUS: Memory-Infused Depth Up-Scaling","ref_index":21,"is_internal_anchor":true},{"citing_arxiv_id":"2602.13215","citing_title":"When to Think Fast and Slow? AMOR: Adaptive Entropy Gate for Hybrid Models","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2603.18297","citing_title":"Path-Constrained Mixture-of-Experts","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13190","citing_title":"N-vium: Mixture-of-Exits Transformer for Accelerated Exact Generation","ref_index":40,"is_internal_anchor":true},{"citing_arxiv_id":"2605.03110","citing_title":"Cascade Token Selection for Transformer Attention Acceleration","ref_index":9,"is_internal_anchor":true},{"citing_arxiv_id":"2605.03109","citing_title":"Gated Subspace Inference for Transformer Acceleration","ref_index":18,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09630","citing_title":"Scratchpad Patching: Decoupling Compute from Patch Size in Byte-Level Language Models","ref_index":77,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09165","citing_title":"Sparse Layers are Critical to Scaling Looped Language Models","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2604.19835","citing_title":"Expert Upcycling: Shifting the Compute-Efficient Frontier of Mixture-of-Experts","ref_index":44,"is_internal_anchor":true},{"citing_arxiv_id":"2604.20470","citing_title":"DynamicRad: Content-Adaptive Sparse Attention for Long Video Diffusion","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18965","citing_title":"Transformer Architecture with Minimal Inference Latency for Multi-Modal Wireless Networks","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2604.12946","citing_title":"Parcae: Scaling Laws For Stable Looped Language Models","ref_index":65,"is_internal_anchor":true},{"citing_arxiv_id":"2604.14419","citing_title":"Equifinality in Mixture of Experts: Routing Topology Does Not Determine Language Modeling Quality","ref_index":33,"is_internal_anchor":true},{"citing_arxiv_id":"2604.16745","citing_title":"Why Training-Free Token Reduction Collapses: The Inherent Instability of Pairwise Scoring Signals","ref_index":45,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05222","citing_title":"Adaptive Computation Depth via Learned Token Routing in Transformers","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2604.17228","citing_title":"Revisiting Auxiliary Losses for Conditional Depth Routing: An Empirical Study","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2604.17286","citing_title":"Depth Adaptive Efficient Visual Autoregressive Modeling","ref_index":43,"is_internal_anchor":true},{"citing_arxiv_id":"2604.19835","citing_title":"Expert Upcycling: Shifting the Compute-Efficient Frontier of Mixture-of-Experts","ref_index":44,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":1,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5","json":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5.json","graph_json":"https://pith.science/api/pith-number/AXFU3DUTR7SPUGLPK667UTVIY5/graph.json","events_json":"https://pith.science/api/pith-number/AXFU3DUTR7SPUGLPK667UTVIY5/events.json","paper":"https://pith.science/paper/AXFU3DUT"},"agent_actions":{"view_html":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5","download_json":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5.json","view_paper":"https://pith.science/paper/AXFU3DUT","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.02258&json=true","fetch_graph":"https://pith.science/api/pith-number/AXFU3DUTR7SPUGLPK667UTVIY5/graph.json","fetch_events":"https://pith.science/api/pith-number/AXFU3DUTR7SPUGLPK667UTVIY5/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5/action/timestamp_anchor","attest_storage":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5/action/storage_attestation","attest_author":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5/action/author_attestation","sign_citation":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5/action/citation_signature","submit_replication":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5/action/replication_record"}},"created_at":"2026-05-17T23:38:15.410725+00:00","updated_at":"2026-05-17T23:38:15.410725+00:00"}