---
title: "DeepSeek-OCR 2: Visual Causal Flow"
authors:
  - "Haoran Wei"
  - "Yaofeng Sun"
  - "Yukun Li"
year: 2026
venue: "arXiv:cs.CV"
arxiv: "2601.20552"
doi: null
source: "https://arxiv.org/abs/2601.20552"
topics:
  - "ocr"
  - "causal-flow"
  - "deepseek"
seed_rank: 998
seed_batch: "deepseek-lab-2026-08-13"
reviewed: "2026-08-13"
pool: "agents"
relevance_score: 7
lineage: deepseek
cites:
  - title: "DeepSeek-OCR: Contexts Optical Compression"
    url: "https://arxiv.org/abs/2510.18234"
    year: 2025
    arxiv: "2510.18234"
see:
  - "993-deepseek-ocr-contexts-optical-compression"
---

# DeepSeek-OCR 2: Visual Causal Flow

## One-sentence takeaway

DeepEncoder V2 reorders visual tokens by image semantics before the LLM reads them, turning raster-scan 2D into two cascaded 1D causal reasoners so layout (reading order, tables, UI flow) is a learned sequence rather than a fixed PE.

## Why it matters here

Follow-on to optical compression: the encoder itself becomes a causal reorderer. Layout-heavy agent perception — UI, docs, sheets — is exactly where raster order lies.

## Key ideas

- arXiv:2601.20552. Conventional VLMs feed vision tokens in raster order with fixed positional encodings; humans scan by structure.
- DeepEncoder V2 dynamically reorders visual tokens from image semantics, then the LLM reads that 1D causal stream.
- Thesis: 2D understanding via two cascaded 1D causal structures (encoder reorder, then language model).
- Exploratory encoder paper, not a general VLM replacement. Causal reorder is learned — adversarial layouts are an open failure mode.

## Caveats

## Links

- arXiv: [2601.20552](https://arxiv.org/abs/2601.20552)
- GitHub: https://github.com/deepseek-ai/DeepSeek-OCR-2