{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:EPKXGYK26Z27ZOFBTKJOBS52BK","short_pith_number":"pith:EPKXGYK2","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"},"canonical_sha256":"23d573615af675fcb8a19a92e0cbba0a8c62837d54ea709a9bd6f621d4e036e0","source":{"kind":"arxiv","id":"2605.06554","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.06554","created_at":"2026-05-18T15:04:06Z"},{"alias_kind":"arxiv_version","alias_value":"2605.06554v1","created_at":"2026-05-18T15:04:06Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.06554","created_at":"2026-05-18T15:04:06Z"},{"alias_kind":"pith_short_12","alias_value":"EPKXGYK26Z27","created_at":"2026-05-18T15:04:06Z"},{"alias_kind":"pith_short_16","alias_value":"EPKXGYK26Z27ZOFB","created_at":"2026-05-18T15:04:06Z"},{"alias_kind":"pith_short_8","alias_value":"EPKXGYK2","created_at":"2026-05-18T15:04:06Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:EPKXGYK26Z27ZOFBTKJOBS52BK","target":"record","payload":{"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"},"canonical_sha256":"23d573615af675fcb8a19a92e0cbba0a8c62837d54ea709a9bd6f621d4e036e0","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"},"source_kind":"arxiv","source_id":"2605.06554","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":"2026-05-18T15:04:06Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"Zvs3WpTY26eW37FDHBsk8eLjK4ZkGWd6iKJNqB6Ni+YOwqWIJcUt7exZw1trEpaAEDeG9n9XIRxa2PSy9UNiAQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-27T11:49:40.745158Z"},"content_sha256":"b1e3c11542b0e82b41679f7b40b75be70097da196a359ede97b205e301d23333","schema_version":"1.0","event_id":"sha256:b1e3c11542b0e82b41679f7b40b75be70097da196a359ede97b205e301d23333"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:EPKXGYK26Z27ZOFBTKJOBS52BK","target":"graph","payload":{"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"},"verdict_id":"ec25ce7b-4675-4b5f-9957-35f0102cc98e"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T15:04:06Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"/f6swP6l8i749Qt3UDhnCzxJTNXWPBlXw4ICVb0EBeIaMMb7Pq2tPXFPbwgbKXowL9sd9fawwt+KFLPkJbg7Ag==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-27T11:49:40.746007Z"},"content_sha256":"c093113df7d1c1b64bbc14fb47e9df06b68aef9635cc5503badd278eef006422","schema_version":"1.0","event_id":"sha256:c093113df7d1c1b64bbc14fb47e9df06b68aef9635cc5503badd278eef006422"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/bundle.json","state_url":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/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":"2026-05-27T11:49:40Z","links":{"resolver":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK","bundle":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/bundle.json","state":"https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/state.json","well_known_bundle":"https://pith.science/.well-known/pith/EPKXGYK26Z27ZOFBTKJOBS52BK/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:EPKXGYK26Z27ZOFBTKJOBS52BK","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":"2682881d5f125fa2fbecc6a2fc6613035188658a06a0a8690fcf19655d9e8a61","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-07T16:49:28Z","title_canon_sha256":"82764fb6596f7ad579d4d1ef4baf1d2549c112569b326df39b8f907d28562c30"},"schema_version":"1.0","source":{"id":"2605.06554","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.06554","created_at":"2026-05-18T15:04:06Z"},{"alias_kind":"arxiv_version","alias_value":"2605.06554v1","created_at":"2026-05-18T15:04:06Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.06554","created_at":"2026-05-18T15:04:06Z"},{"alias_kind":"pith_short_12","alias_value":"EPKXGYK26Z27","created_at":"2026-05-18T15:04:06Z"},{"alias_kind":"pith_short_16","alias_value":"EPKXGYK26Z27ZOFB","created_at":"2026-05-18T15:04:06Z"},{"alias_kind":"pith_short_8","alias_value":"EPKXGYK2","created_at":"2026-05-18T15:04:06Z"}],"graph_snapshots":[{"event_id":"sha256:c093113df7d1c1b64bbc14fb47e9df06b68aef9635cc5503badd278eef006422","target":"graph","created_at":"2026-05-18T15:04:06Z","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":"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."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","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."}],"snapshot_sha256":"467436856f06fd37194e9112484b41e425e766fb42f396f583b0f4531f2f2702"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"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","authors_text":"Bowen Peng, Jeffrey Quesnelle, Subho Ghosh","cross_cats":[],"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.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-07T16:49:28Z","title":"Long Context Pre-Training with Lighthouse Attention"},"references":{"count":41,"internal_anchors":17,"resolved_work":21,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"The Claude 3 model family, 2024","work_id":"","year":2024},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Zoology: Measuring and improving recall in efficient language models","work_id":"a51e0ddf-22e2-4d3c-9759-7201f7d9a699","year":2024},{"cited_arxiv_id":"1308.3432","doi":"","is_internal_anchor":true,"ref_index":3,"title":"Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation","work_id":"1fe8c7c8-aff7-4b94-9096-e549d7e60789","year":2013},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"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":"","year":2021},{"cited_arxiv_id":"2307.08691","doi":"","is_internal_anchor":true,"ref_index":5,"title":"FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning","work_id":"fff3953b-5efb-4753-bee4-002f59995810","year":2024}],"snapshot_sha256":"3c97813858b59ec23e36aba941b7210701fada9ad4a236841e2d04778a76dae3"},"source":{"id":"2605.06554","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-08T10:05:21.912416Z","id":"ec25ce7b-4675-4b5f-9957-35f0102cc98e","model_set":{"reader":"grok-4.3"},"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","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.","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.","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."}},"verdict_id":"ec25ce7b-4675-4b5f-9957-35f0102cc98e"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:b1e3c11542b0e82b41679f7b40b75be70097da196a359ede97b205e301d23333","target":"record","created_at":"2026-05-18T15:04:06Z","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":"2682881d5f125fa2fbecc6a2fc6613035188658a06a0a8690fcf19655d9e8a61","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-07T16:49:28Z","title_canon_sha256":"82764fb6596f7ad579d4d1ef4baf1d2549c112569b326df39b8f907d28562c30"},"schema_version":"1.0","source":{"id":"2605.06554","kind":"arxiv","version":1}},"canonical_sha256":"23d573615af675fcb8a19a92e0cbba0a8c62837d54ea709a9bd6f621d4e036e0","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"23d573615af675fcb8a19a92e0cbba0a8c62837d54ea709a9bd6f621d4e036e0","first_computed_at":"2026-05-18T15:04:06.605382Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T15:04:06.605382Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"/c8W1CFH7sdrtDvUr84HjG6MlxCf5Oc3XhS10KqbgWhlggyAA275O5gH1X737UQ84IutwEhmTjSUTKROtmBtAA==","signature_status":"signed_v1","signed_at":"2026-05-18T15:04:06.606101Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.06554","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:b1e3c11542b0e82b41679f7b40b75be70097da196a359ede97b205e301d23333","sha256:c093113df7d1c1b64bbc14fb47e9df06b68aef9635cc5503badd278eef006422"],"state_sha256":"6ad38edf04fe0dda275a493c9f61b41bc4c3482688692331f393ce7e309e114e"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"CAjejHj4z+VvmQ4A/6pyBYOuBpwnIQxhN//RFWutYLfcHE615Qhu/UMwAEzGUGUL63X2aOdAsmvQkyD/UfOmBQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-27T11:49:40.749609Z","bundle_sha256":"62784cca661cff07ec39e70d0b24db3975a0d332dfa13878eb01993e23ab34c4"}}