Pith Number

pith:WP3BBLDL

pith:2026:WP3BBLDLKSDFOXPGWIHW7UNOJR

not attested not anchored not stored refs pending

Where Vision Becomes Text: Locating the OCR Routing Bottleneck in Vision-Language Models

Jonathan Steinberg, Oren Gal

OCR information routes into vision-language models at architecture-specific layers, forming a low-dimensional signal that transfers across datasets.

arxiv:2602.22918 v3 · 2026-02-26 · 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{WP3BBLDLKSDFOXPGWIHW7UNOJR}

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

The OCR signal is remarkably low-dimensional: PC1 captures 72.9% of variance. Crucially, principal component analysis (PCA) directions learned on one dataset transfer to others, demonstrating shared text-processing pathways. Surprisingly, in models with modular OCR circuits (notably Qwen3-VL-4B), OCR removal can improve counting performance (up to +6.9 percentage points).

C2weakest assumption

That inpainting text in images only affects the OCR pathway without introducing other unintended changes to the visual input that could confound the activation differences.

C3one line summary

Causal interventions identify architecture-specific OCR bottlenecks in VLMs at mid or early layers, with low-dimensional shared pathways and potential performance benefits from OCR removal.

Formal links

2 machine-checked theorem links

Cited by

1 paper in Pith

Don't Look at the Numbers: Visual Anchoring Bias and Layer-wise Representation in VLMs

Receipt and verification

First computed	2026-05-20T00:00:35.208847Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

b3f610ac6b5486575de6b20f6fd1ae4c4e9565b6e22d4dd46c45961fd0fddbd4

Aliases

arxiv: 2602.22918 · arxiv_version: 2602.22918v3 · doi: 10.48550/arxiv.2602.22918 · pith_short_12: WP3BBLDLKSDF · pith_short_16: WP3BBLDLKSDFOXPG · pith_short_8: WP3BBLDL

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/WP3BBLDLKSDFOXPGWIHW7UNOJR \
  | 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: b3f610ac6b5486575de6b20f6fd1ae4c4e9565b6e22d4dd46c45961fd0fddbd4

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "a7dc2563348038b7712030ef79d94e915b77d304f57d01eb0c8c5a525ff2c84c",
    "cross_cats_sorted": [],
    "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
    "primary_cat": "cs.CL",
    "submitted_at": "2026-02-26T12:06:02Z",
    "title_canon_sha256": "0e39d9d87f55beb0bb2e47a6db5c08075fbd593200c20273a65c7ec7d32c6516"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2602.22918",
    "kind": "arxiv",
    "version": 3
  }
}