Pith Number

pith:A5FG35S7

pith:2025:A5FG35S7CKS6D6ESOMRXLFI35P

not attested not anchored not stored refs resolved

Latent Visual Reasoning

Bangzheng Li, Emad Barsoum, Hao Chen, Jialian Wu, Jiang Liu, Muhao Chen, Xiaodong Yu, Ximeng Sun, Ze Wang, Zicheng Liu

Multimodal models can perform reasoning steps by autoregressively generating latent visual states that reconstruct key image tokens.

arxiv:2509.24251 v2 · 2025-09-29 · cs.CV · 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{A5FG35S7CKS6D6ESOMRXLFI35P}

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 introduce Latent Visual Reasoning (LVR), a new paradigm that enables autoregressive reasoning directly in the visual embedding space... By interleaving LVR with standard text generation, our model achieves substantial gains on perception-intensive visual question answering tasks.

C2weakest assumption

That generating latent states whose explicit goal is to reconstruct selected visual tokens constitutes genuine visual reasoning that improves downstream task performance beyond what language-only CoT or tool-based editing already achieves.

C3one line summary

Latent Visual Reasoning enables autoregressive generation of latent visual states that reconstruct critical image tokens, yielding gains on perception-heavy VQA benchmarks such as 71.67% on MMVP.

References

29 extracted · 29 resolved · 17 Pith anchors

[1] Qwen2.5-VL Technical Report · arXiv:2502.13923

[2] Diagnosing and mitigating modality interference in multimodal large language models

[3] SFT or RL? An Early Investigation into Training R1-Like Reasoning Large Vision-Language Models · arXiv:2504.11468

[4] Compressed Chain of Thought: Efficient Reasoning Through Dense Representations · arXiv:2412.13171

[5] v1: Learning to Point Visual Tokens for Multimodal Grounded Reasoning · arXiv:2505.18842

Formal links

2 machine-checked theorem links

Cited by

32 papers in Pith

4DThinker: Thinking with 4D Imagery for Dynamic Spatial Understanding

LatentOmni: Rethinking Omni-Modal Understanding via Unified Audio-Visual Latent Reasoning

S$^2$GR: Stepwise Semantic-Guided Reasoning in Latent Space for Generative Recommendation

Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models

What's Holding Back Latent Visual Reasoning?

Receipt and verification

First computed	2026-05-17T23:38:50.557867Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

074a6df65f12a5e1f892732375951bebd1a9aecc134fabdc66ec4e602b6bbda9

Aliases

arxiv: 2509.24251 · arxiv_version: 2509.24251v2 · doi: 10.48550/arxiv.2509.24251 · pith_short_12: A5FG35S7CKS6 · pith_short_16: A5FG35S7CKS6D6ES · pith_short_8: A5FG35S7

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/A5FG35S7CKS6D6ESOMRXLFI35P \
  | 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: 074a6df65f12a5e1f892732375951bebd1a9aecc134fabdc66ec4e602b6bbda9

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "6e5d09a53c02ec8eaea5cd02f2729d5217e9f565b9604f26438a0960398b3982",
    "cross_cats_sorted": [
      "cs.CL"
    ],
    "license": "http://creativecommons.org/licenses/by/4.0/",
    "primary_cat": "cs.CV",
    "submitted_at": "2025-09-29T03:52:01Z",
    "title_canon_sha256": "8020884362e361106ef159f775c8b48cb15e589341acafc13635e7b10be70580"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2509.24251",
    "kind": "arxiv",
    "version": 2
  }
}