Pith Number

pith:KF37V5QU

pith:2025:KF37V5QUANOISNQBGSURYCLID6

not attested not anchored not stored refs pending

Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering

Dimitris N. Metaxas, Gemma E. Moran, Gerasimos Chatzoudis, Hao Wang, Zhuowei Li

Visual Sparse Steering extracts a steering vector from SAE features on unlabeled data to adapt CLIP models at test time and raise zero-shot accuracy by 1-4%.

arxiv:2506.01247 v3 · 2025-06-02 · cs.CV · cs.AI · cs.LG

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

Add to your LaTeX paper

\usepackage{pith}
\pithnumber{KF37V5QUANOISNQBGSURYCLID6}

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

Across CIFAR-100, CUB-200, and Tiny-ImageNet and two CLIP backbones, VS2 improves zero-shot top-1 accuracy by 3.45-4.12%, 0.93-1.08%, and 1.50-1.84% respectively, while remaining forward-only and adding minimal compute overhead.

C2weakest assumption

The assumption that sparse features extracted by an SAE trained on unlabeled in-domain activations contain task-relevant information that can be reliably turned into an effective steering vector without any labeled data or optimization at test time.

C3one line summary

VS2 constructs steering vectors from sparse SAE features on unlabeled in-domain activations to improve zero-shot accuracy of CLIP models by 0.93-4.12% on CIFAR-100, CUB-200, and Tiny-ImageNet while remaining forward-pass only.

Formal links

1 machine-checked theorem link

Cited by

2 papers in Pith

SwordBench: Evaluating Orthogonality of Steering Image Representations

Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision

Receipt and verification

First computed	2026-05-28T01:04:28.424360Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

5177faf614035c89360134a91c09681fa4f608e36a989d446fcb216cb2f8fd7d

Aliases

arxiv: 2506.01247 · arxiv_version: 2506.01247v3 · doi: 10.48550/arxiv.2506.01247 · pith_short_12: KF37V5QUANOI · pith_short_16: KF37V5QUANOISNQB · pith_short_8: KF37V5QU

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/KF37V5QUANOISNQBGSURYCLID6 \
  | 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: 5177faf614035c89360134a91c09681fa4f608e36a989d446fcb216cb2f8fd7d

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "6f5c4b69b71d0f8723bd4a66d3d3ecdce6b9a42bb9bf40880c2660c610d64bb2",
    "cross_cats_sorted": [
      "cs.AI",
      "cs.LG"
    ],
    "license": "http://creativecommons.org/licenses/by/4.0/",
    "primary_cat": "cs.CV",
    "submitted_at": "2025-06-02T01:51:20Z",
    "title_canon_sha256": "6569d1084f929db00478bca025773fc5878573e280f030c727caa9835d7d046f"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2506.01247",
    "kind": "arxiv",
    "version": 3
  }
}