DeepSeek：2026 DeepSeek-OCR 2技术报告（英文版）（15页）.pdf

1、DeepSeek-OCR 2:Visual Causal FlowHaoran Wei,Yaofeng Sun,Yukun LiDeepSeek-AIAbstractWe present DeepSeek-OCR 2 to investigate the feasibility of a novel encoderDeepEncoderV2capable of dynamically reordering visual tokens upon image semantics.Conventionalvision-language models(VLMs)invariably process v

2、isual tokens in a rigid raster-scan order(top-left to bottom-right)with fixed positional encoding when fed into LLMs.However,thiscontradicts human visual perception,which follows flexible yet semantically coherent scanningpatterns driven by inherent logical structures.Particularly for images with co

3、mplex layouts,human vision exhibits causally-informed sequential processing.Inspired by this cognitive mech-anism,DeepEncoder V2 is designed to endow the encoder with causal reasoning capabilities,enabling it to intelligently reorder visual tokens prior to LLM-based content interpretation.Thiswork e

4、xplores a novel paradigm:whether 2D image understanding can be effectively achievedthrough two-cascaded 1D causal reasoning structures,thereby offering a new architecturalapproach with the potential to achieve genuine 2D reasoning.Codes and model weights arepublicly accessible athttp:/ as Vision Enc

5、oderQwen2 500M.causallearnable query(new reading order).Tokenizer80M 16xnon-causalCLIP VIT 300M.DeepEncoderDeepEncoder V2Tokenizer80M 16xFigure 1|We substitute the CLIP component in DeepEncoder with an LLM-style architecture.By customizing the attention mask,visual tokens utilize bidirectional atten

6、tion while learnablequeries adopt causal attention.Each query token can thus attend to all visual tokens andpreceding queries,allowing progressive causal reordering over visual information.Contents1Introduction32Related Works42.1Parallelized Queries in Decoder.42.2Parallelized Queries in Projector.4