Pith Number

pith:CDPXLLXD

pith:2026:CDPXLLXDLVKC4L4DTXEESFZFVY

not attested not anchored not stored refs resolved

When Language Overwrites Vision: Over-Alignment and Geometric Debiasing in Vision-Language Models

Dianbo Liu, Harshvardhan Saini, Samyak Jha, Yiming Tang

Vision-language models hallucinate because they over-align visual embeddings to text, and removing a linguistic bias subspace fixes it.

arxiv:2605.08245 v3 · 2026-05-07 · cs.CV · cs.AI

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

Add to your LaTeX paper

\usepackage{pith}
\pithnumber{CDPXLLXDLVKC4L4DTXEESFZFVY}

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 trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with the text manifold, injecting a statistical linguistic bias that systematically overshadows fine-grained visual evidence.

C2weakest assumption

Linguistic bias concentrates in the top principal components of a universal, dataset-agnostic text subspace, so that explicitly projecting this subspace from visual representations removes the bias without discarding task-critical visual information.

C3one line summary

Decoder-based VLMs hallucinate because visual embeddings are over-aligned to a text manifold; projecting out the top principal components of a universal linguistic subspace reduces this bias and improves benchmark performance.

References

23 extracted · 23 resolved · 9 Pith anchors

[1] Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C

[2] Neeraj Anand, Samyak Jha, Udbhav Bamba, and Rahul Rahaman · arXiv:1505.00468

[3] Qwen3-VL Technical Report · arXiv:2511.21631

[4] David M Chan, Suzanne Petryk, Joseph E Gonzalez, Trevor Darrell, and John Canny 2023

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

Formal links

2 machine-checked theorem links

Receipt and verification

First computed	2026-05-20T00:00:41.432026Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

10df75aee35d542e2f839dc8491725ae186e4a73fabe8f8a042467bb4071a8b8

Aliases

arxiv: 2605.08245 · arxiv_version: 2605.08245v3 · doi: 10.48550/arxiv.2605.08245 · pith_short_12: CDPXLLXDLVKC · pith_short_16: CDPXLLXDLVKC4L4D · pith_short_8: CDPXLLXD

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/CDPXLLXDLVKC4L4DTXEESFZFVY \
  | 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: 10df75aee35d542e2f839dc8491725ae186e4a73fabe8f8a042467bb4071a8b8

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "dd62347be14405281fa47c7f4afb2a286d61049507361e4209e6cba3f86acabc",
    "cross_cats_sorted": [
      "cs.AI"
    ],
    "license": "http://creativecommons.org/licenses/by/4.0/",
    "primary_cat": "cs.CV",
    "submitted_at": "2026-05-07T10:09:18Z",
    "title_canon_sha256": "1f2d578158156bab9deba077bb72a4ceb7c2d42703adca3a0e2f2827023a0c89"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2605.08245",
    "kind": "arxiv",
    "version": 3
  }
}