Pith Number

pith:DACSCEA3

pith:2026:DACSCEA3NEK2F4Q6H6MLAPI6NM

not attested not anchored not stored refs resolved

WriteSAE: Sparse Autoencoders for Recurrent State

Jack Young

WriteSAE factors decoder atoms to match rank-1 cache writes so they can be swapped directly into recurrent state models.

arxiv:2605.12770 v2 · 2026-05-12 · cs.LG · cs.AI · 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{DACSCEA3NEK2F4Q6H6MLAPI6NM}

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

Atom substitution beats matched-norm ablation on 92.4% of n=4,851 firings at Qwen3.5-0.8B L9 H4, the 87-atom population test holds at 89.8%, the closed form predicts measured effects at R²=0.98, and Mamba-2-370M substitutes at 88.1% over 2,500 firings. Sustained three-position installs at 3× lift midrank target-in-continuation from 33.3% to 100% under greedy decoding.

C2weakest assumption

That atoms trained under matched Frobenius norm can be substituted into the live cache without unintended side effects on the model's recurrent dynamics, and that the closed-form logit shift remains accurate when atoms are installed in real forward passes.

C3one line summary

WriteSAE decomposes recurrent model cache writes into substitutable atoms with a closed-form logit shift, achieving high substitution success and targeted behavioral installs on models like Qwen3.5 and Mamba-2.

References

106 extracted · 106 resolved · 19 Pith anchors

[1] Transformer Circuits Thread , year=

[2] Sparse Autoencoders Find Highly Interpretable Features in Language Models · arXiv:2309.08600

[3] Scaling and evaluating sparse autoencoders · arXiv:2406.04093

[4] and McDougall, Callum and MacDiarmid, Monte and Freeman, C 2024

[5] Improving Dictionary Learning with Gated Sparse Autoencoders · arXiv:2404.16014

Formal links

2 machine-checked theorem links

Receipt and verification

First computed	2026-05-18T03:09:20.132619Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

180521101b6915a2f21e3f98b03d1e6b2483b32320028454f91d6ff82158ae4e

Aliases

arxiv: 2605.12770 · arxiv_version: 2605.12770v2 · doi: 10.48550/arxiv.2605.12770 · pith_short_12: DACSCEA3NEK2 · pith_short_16: DACSCEA3NEK2F4Q6 · pith_short_8: DACSCEA3

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/DACSCEA3NEK2F4Q6H6MLAPI6NM \
  | 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: 180521101b6915a2f21e3f98b03d1e6b2483b32320028454f91d6ff82158ae4e

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "6874e131c453332b54191b6002715149ebc9dc8f91796ff94e3105798ffa4398",
    "cross_cats_sorted": [
      "cs.AI",
      "cs.CL"
    ],
    "license": "http://creativecommons.org/licenses/by/4.0/",
    "primary_cat": "cs.LG",
    "submitted_at": "2026-05-12T21:32:45Z",
    "title_canon_sha256": "9ab77441e4d5e6cca37116e9f0fa2a8c9db0608d19962fac65052cf16109ce4f"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2605.12770",
    "kind": "arxiv",
    "version": 2
  }
}