{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2024:AXFU3DUTR7SPUGLPK667UTVIY5","short_pith_number":"pith:AXFU3DUT","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"},"canonical_sha256":"05cb4d8e938fe4fa196f57bdfa4ea8c7598877576765f04649a8f2cca081e4fa","source":{"kind":"arxiv","id":"2404.02258","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2404.02258","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"arxiv_version","alias_value":"2404.02258v1","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.02258","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"pith_short_12","alias_value":"AXFU3DUTR7SP","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"AXFU3DUTR7SPUGLP","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"AXFU3DUT","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2024:AXFU3DUTR7SPUGLPK667UTVIY5","target":"record","payload":{"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"},"canonical_sha256":"05cb4d8e938fe4fa196f57bdfa4ea8c7598877576765f04649a8f2cca081e4fa","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"},"source_kind":"arxiv","source_id":"2404.02258","source_version":1,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:15Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"bVmdaMVVijeNAakM44KUJiw7JobUQz58bn05oqsHRtU3zEhmK07iVXr5MLUs4blDQij2NkMWn3WCvj1H4N1zAA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-18T22:36:43.870386Z"},"content_sha256":"ced473b3eff80f934761c77b1a97db7310a92ac91d22c8f83f50c7343fbecd2c","schema_version":"1.0","event_id":"sha256:ced473b3eff80f934761c77b1a97db7310a92ac91d22c8f83f50c7343fbecd2c"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2024:AXFU3DUTR7SPUGLPK667UTVIY5","target":"graph","payload":{"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"},"verdict_id":"47fc0ddc-7aad-4707-9d3b-017bc1794454"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:15Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"XEs8e4FeGX6jcNwJbUWe3Yosjx26p/ClC0mWffvXyI8R9kgjGVEjHxgKNczQwzpCc410NL8i4cOBIpIXBHpyDw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-18T22:36:43.871176Z"},"content_sha256":"b3bb8e2c2258a28fdab4282da5ec9c1396dc72a8cc110708ee2b57e72c592ea3","schema_version":"1.0","event_id":"sha256:b3bb8e2c2258a28fdab4282da5ec9c1396dc72a8cc110708ee2b57e72c592ea3"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5/bundle.json","state_url":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/AXFU3DUTR7SPUGLPK667UTVIY5/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-05-18T22:36:43Z","links":{"resolver":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5","bundle":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5/bundle.json","state":"https://pith.science/pith/AXFU3DUTR7SPUGLPK667UTVIY5/state.json","well_known_bundle":"https://pith.science/.well-known/pith/AXFU3DUTR7SPUGLPK667UTVIY5/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2024:AXFU3DUTR7SPUGLPK667UTVIY5","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":"20196f2ef34c9194cb315bc7bc2d6b7e36cc76b23f47901f1dbbbc05b493a687","cross_cats_sorted":["cs.CL"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-04-02T19:28:11Z","title_canon_sha256":"ec4a74a99a0bfb9a2c6146a1ee7e2a6ddcd656c28232813ecf9f44bf5d9f3bf9"},"schema_version":"1.0","source":{"id":"2404.02258","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2404.02258","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"arxiv_version","alias_value":"2404.02258v1","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.02258","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"pith_short_12","alias_value":"AXFU3DUTR7SP","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"AXFU3DUTR7SPUGLP","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"AXFU3DUT","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:b3bb8e2c2258a28fdab4282da5ec9c1396dc72a8cc110708ee2b57e72c592ea3","target":"graph","created_at":"2026-05-17T23:38:15Z","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":"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."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Transformer language models can learn to dynamically allocate compute to select tokens at each layer."}],"snapshot_sha256":"2487a200c9eb271af365818e08fa621e4c53c12a8d37f2ee982a5b3b838e1b76"},"formal_canon":{"evidence_count":1,"snapshot_sha256":"740ed61cf7e8970d28f0d74fd67712b045d4b6754137d9ae6d7eec1875379ca6"},"paper":{"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,","authors_text":"Adam Santoro, Blake Richards, David Raposo, Peter Conway Humphreys, Sam Ritter, Timothy Lillicrap","cross_cats":["cs.CL"],"headline":"Transformer language models can learn to dynamically allocate compute to select tokens at each layer.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-04-02T19:28:11Z","title":"Mixture-of-Depths: Dynamically allocating compute in transformer-based language models"},"references":{"count":11,"internal_anchors":5,"resolved_work":11,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Controlling computation versus quality for neural sequence models","work_id":"21f2a3df-f950-438d-8553-6db49dc7afe8","year":2002},{"cited_arxiv_id":"1807.03819","doi":"","is_internal_anchor":true,"ref_index":3,"title":"Universal Transformers","work_id":"8e5baefe-d209-411c-aefc-5acaa9275c8a","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Depth-adaptive transformer","work_id":"6d994f34-ab13-423e-96ec-52b63f47a6d6","year":1910},{"cited_arxiv_id":"1603.08983","doi":"","is_internal_anchor":true,"ref_index":7,"title":"Adaptive Computation Time for Recurrent Neural Networks","work_id":"75565443-173e-479c-b0e7-d2464e7630be","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":8,"title":"Towards a unified view of parameter-efficient transfer learning","work_id":"e758988d-4ae2-44ff-94fc-4417c242b1e4","year":null}],"snapshot_sha256":"eb3b74aaf8e7de70ad3073dbc0bfbc87e51628e687adfc5c29387dbc73af8d74"},"source":{"id":"2404.02258","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-17T02:13:07.402130Z","id":"47fc0ddc-7aad-4707-9d3b-017bc1794454","model_set":{"reader":"grok-4.3"},"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","pith_extraction_headline":"Transformer language models can learn to dynamically allocate compute to select tokens at each layer.","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.","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."}},"verdict_id":"47fc0ddc-7aad-4707-9d3b-017bc1794454"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:ced473b3eff80f934761c77b1a97db7310a92ac91d22c8f83f50c7343fbecd2c","target":"record","created_at":"2026-05-17T23:38:15Z","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":"20196f2ef34c9194cb315bc7bc2d6b7e36cc76b23f47901f1dbbbc05b493a687","cross_cats_sorted":["cs.CL"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-04-02T19:28:11Z","title_canon_sha256":"ec4a74a99a0bfb9a2c6146a1ee7e2a6ddcd656c28232813ecf9f44bf5d9f3bf9"},"schema_version":"1.0","source":{"id":"2404.02258","kind":"arxiv","version":1}},"canonical_sha256":"05cb4d8e938fe4fa196f57bdfa4ea8c7598877576765f04649a8f2cca081e4fa","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"05cb4d8e938fe4fa196f57bdfa4ea8c7598877576765f04649a8f2cca081e4fa","first_computed_at":"2026-05-17T23:38:15.410613Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:15.410613Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"ItCiII7Q8FqCuUqiJ1JXdDFWooq6992J6vclKArOfUQ2XRKxuYghSFpANHgqa99LNAFHdWYj3JfR7H1n6muqCw==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:15.411217Z","signed_message":"canonical_sha256_bytes"},"source_id":"2404.02258","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:ced473b3eff80f934761c77b1a97db7310a92ac91d22c8f83f50c7343fbecd2c","sha256:b3bb8e2c2258a28fdab4282da5ec9c1396dc72a8cc110708ee2b57e72c592ea3"],"state_sha256":"e3c72af953077659e57ef667b3365429205c564941c39ebdbd8ca93f3c5c0bf2"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"NEydwDGz+MG9CiRfpuI6MrEVqT3biVWvDVUhO3Kfn4Zao1TjiXhefUBsmgssFt9MJb/mxj4qsSQpehGkqdcgCg==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-18T22:36:43.873985Z","bundle_sha256":"8a7ab9cf4c8757a674b415b62433ea4752befab47c1ff6b5c513051a881534b0"}}