Pith Number

pith:EPKXGYK2

pith:2026:EPKXGYK26Z27ZOFBTKJOBS52BK

not attested not anchored not stored refs resolved

Long Context Pre-Training with Lighthouse Attention

Bowen Peng, Jeffrey Quesnelle, Subho Ghosh

Lighthouse Attention enables faster pre-training of long-context transformers by using hierarchical compression for most training before a short full-attention recovery phase.

arxiv:2605.06554 v1 · 2026-05-07 · cs.CL

Open paper page JSON Open Graph Bundle Merged state Verified badge What is a Pith Number?

Add to your LaTeX paper

\usepackage{pith}
\pithnumber{EPKXGYK26Z27ZOFBTKJOBS52BK}

Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge

Record completeness

1 Bitcoin timestamp

2 Internet Archive

3 Author claim open · sign in to claim

4 Citations open

5 Replications open

✓ Portable graph bundle live · download bundle · merged state

The bundle contains the canonical record plus signed events. A mirror can host it anywhere and recompute the same current state with the deterministic merge algorithm.

Claims

C1strongest 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.

C2weakest 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.

C3one 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.

References

41 extracted · 21 resolved · 17 Pith anchors

[1] The Claude 3 model family, 2024 2024

[2] Zoology: Measuring and improving recall in efficient language models 2024

[3] Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation 2013 · arXiv:1308.3432

[4] 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. 2021

[5] FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning 2024 · arXiv:2307.08691

Receipt and verification

First computed	2026-05-18T15:04:06.605382Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

23d573615af675fcb8a19a92e0cbba0a8c62837d54ea709a9bd6f621d4e036e0

Aliases

arxiv: 2605.06554 · arxiv_version: 2605.06554v1 · doi: 10.48550/arxiv.2605.06554 · pith_short_12: EPKXGYK26Z27 · pith_short_16: EPKXGYK26Z27ZOFB · pith_short_8: EPKXGYK2

Agent API

Resolver JSON Graph JSON Events JSON Schema Signing key

Verify this Pith Number yourself

curl -sH 'Accept: application/ld+json' https://pith.science/pith/EPKXGYK26Z27ZOFBTKJOBS52BK \
  | jq -c '.canonical_record' \
  | python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: 23d573615af675fcb8a19a92e0cbba0a8c62837d54ea709a9bd6f621d4e036e0

Canonical record JSON

{
  "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
  }
}