{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:EPKXGYK26Z27ZOFBTKJOBS52BK","short_pith_number":"pith:EPKXGYK2","schema_version":"1.0","canonical_sha256":"23d573615af675fcb8a19a92e0cbba0a8c62837d54ea709a9bd6f621d4e036e0","source":{"kind":"arxiv","id":"2605.06554","version":1},"attestation_state":"computed","paper":{"title":"Long Context Pre-Training with Lighthouse Attention","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Lighthouse Attention enables faster pre-training of long-context transformers by using hierarchical compression for most training before a short full-attention recovery phase.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Bowen Peng, Jeffrey Quesnelle, Subho Ghosh","submitted_at":"2026-05-07T16:49:28Z","abstract_excerpt":"Training causal transformers at extreme sequence lengths is bottlenecked by the quadratic time and memory of scaled dot-product attention (SDPA). In this work, we propose Lighthouse Attention, a training-only symmetrical selection-based hierarchical attention algorithm that wraps around ordinary SDPA and can be easily removed towards the end of the training. Our hierarchical selection is also gradient-free, which exempts us from dealing with a complicated and potentially inefficient backward pass kernel. Our contribution is three-fold: (i) A subquadratic hierarchical pre- and post-processing s"},"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":"2605.06554","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-07T16:49:28Z","cross_cats_sorted":[],"title_canon_sha256":"82764fb6596f7ad579d4d1ef4baf1d2549c112569b326df39b8f907d28562c30","abstract_canon_sha256":"2682881d5f125fa2fbecc6a2fc6613035188658a06a0a8690fcf19655d9e8a61"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T15:04:06.606101Z","signature_b64":"/c8W1CFH7sdrtDvUr84HjG6MlxCf5Oc3XhS10KqbgWhlggyAA275O5gH1X737UQ84IutwEhmTjSUTKROtmBtAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"23d573615af675fcb8a19a92e0cbba0a8c62837d54ea709a9bd6f621d4e036e0","last_reissued_at":"2026-05-18T15:04:06.605382Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T15:04:06.605382Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Long Context Pre-Training with Lighthouse Attention","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Lighthouse Attention enables faster pre-training of long-context transformers by using hierarchical compression for most training before a short full-attention recovery phase.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Bowen Peng, Jeffrey Quesnelle, Subho Ghosh","submitted_at":"2026-05-07T16:49:28Z","abstract_excerpt":"Training causal transformers at extreme sequence lengths is bottlenecked by the quadratic time and memory of scaled dot-product attention (SDPA). In this work, we propose Lighthouse Attention, a training-only symmetrical selection-based hierarchical attention algorithm that wraps around ordinary SDPA and can be easily removed towards the end of the training. Our hierarchical selection is also gradient-free, which exempts us from dealing with a complicated and potentially inefficient backward pass kernel. Our contribution is three-fold: (i) A subquadratic hierarchical pre- and post-processing s"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We run preliminary small scale LLM pre-training experiments that show the effectiveness of our method compared to full attention training with all other settings matched, where we achieve a faster total training time and lower final loss after the recovery phase.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the information lost during hierarchical compression can be reliably recovered in a short full-attention phase without introducing lasting biases or requiring extensive additional training, and that small-scale results will hold at larger model sizes and longer contexts.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Lighthouse Attention enables faster long-context pre-training via gradient-free symmetrical hierarchical compression of QKV while preserving causality, followed by a short full-attention recovery that yields lower loss than standard full-attention training.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Lighthouse Attention enables faster pre-training of long-context transformers by using hierarchical compression for most training before a short full-attention recovery phase.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"467436856f06fd37194e9112484b41e425e766fb42f396f583b0f4531f2f2702"},"source":{"id":"2605.06554","kind":"arxiv","version":1},"verdict":{"id":"ec25ce7b-4675-4b5f-9957-35f0102cc98e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T10:05:21.912416Z","strongest_claim":"We run preliminary small scale LLM pre-training experiments that show the effectiveness of our method compared to full attention training with all other settings matched, where we achieve a faster total training time and lower final loss after the recovery phase.","one_line_summary":"Lighthouse Attention enables faster long-context pre-training via gradient-free symmetrical hierarchical compression of QKV while preserving causality, followed by a short full-attention recovery that yields lower loss than standard full-attention training.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the information lost during hierarchical compression can be reliably recovered in a short full-attention phase without introducing lasting biases or requiring extensive additional training, and that small-scale results will hold at larger model sizes and longer contexts.","pith_extraction_headline":"Lighthouse Attention enables faster pre-training of long-context transformers by using hierarchical compression for most training before a short full-attention recovery phase."},"references":{"count":41,"sample":[{"doi":"","year":2024,"title":"The Claude 3 model family, 2024","work_id":"","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Zoology: Measuring and improving recall in efficient language models","work_id":"a51e0ddf-22e2-4d3c-9759-7201f7d9a699","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2013,"title":"Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation","work_id":"1fe8c7c8-aff7-4b94-9096-e549d7e60789","ref_index":3,"cited_arxiv_id":"1308.3432","is_internal_anchor":true},{"doi":"","year":2021,"title":"K. Choromanski, V . Likhosherstov, D. Dohan, X. Song, A. Gane, T. Sarlos, P. Hawkins, J. Davis, A. Mohiuddin, L. Kaiser, D. Belanger, L. Colwell, and A. Weller. Rethinking at- tention with Performers.","work_id":"","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning","work_id":"fff3953b-5efb-4753-bee4-002f59995810","ref_index":5,"cited_arxiv_id":"2307.08691","is_internal_anchor":true}],"resolved_work":21,"snapshot_sha256":"3c97813858b59ec23e36aba941b7210701fada9ad4a236841e2d04778a76dae3","internal_anchors":17},"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":"2605.06554","created_at":"2026-05-18T15:04:06.605489+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.06554v1","created_at":"2026-05-18T15:04:06.605489+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.06554","created_at":"2026-05-18T15:04:06.605489+00:00"},{"alias_kind":"pith_short_12","alias_value":"EPKXGYK26Z27","created_at":"2026-05-18T15:04:06.605489+00:00"},{"alias_kind":"pith_short_16","alias_value":"EPKXGYK26Z27ZOFB","created_at":"2026-05-18T15:04:06.605489+00:00"},{"alias_kind":"pith_short_8","alias_value":"EPKXGYK2","created_at":"2026-05-18T15:04:06.605489+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/EPKXGYK26Z27ZOFBTKJOBS52BK","json":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK.json","graph_json":"https://pith.science/api/pith-number/EPKXGYK26Z27ZOFBTKJOBS52BK/graph.json","events_json":"https://pith.science/api/pith-number/EPKXGYK26Z27ZOFBTKJOBS52BK/events.json","paper":"https://pith.science/paper/EPKXGYK2"},"agent_actions":{"view_html":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK","download_json":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK.json","view_paper":"https://pith.science/paper/EPKXGYK2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.06554&json=true","fetch_graph":"https://pith.science/api/pith-number/EPKXGYK26Z27ZOFBTKJOBS52BK/graph.json","fetch_events":"https://pith.science/api/pith-number/EPKXGYK26Z27ZOFBTKJOBS52BK/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/action/timestamp_anchor","attest_storage":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/action/storage_attestation","attest_author":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/action/author_attestation","sign_citation":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/action/citation_signature","submit_replication":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/action/replication_record"}},"created_at":"2026-05-18T15:04:06.605489+00:00","updated_at":"2026-05-18T15:04:06.605489+00:00"}