Institution profile

DeepSeek AI

Industry researchasia · cn
Official website
Research library18linked papers
Opportunities29open roles
Selected work

Representative Papers

Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models

Jan 12, 2026

This work addresses the limitation of conventional Transformers, which lack a native knowledge retrieval mechanism and rely on inefficient computation to simulate memory access, thereby constraining their performance on knowledge-intensive tasks. The authors propose “conditional memory” as a novel dimension of sparsity, introducing an Engram module based on modern N-gram embeddings to enable O(1) static knowledge lookup. This module is co-optimized with a Mixture-of-Experts (MoE) architecture to balance neural computation and memory invocation. For the first time, scalable static memory is integrated into sparse large language models, revealing a U-shaped scaling law between memory capacity and model performance. The memory module alleviates the burden on the backbone network, allowing it to focus on complex reasoning. The method significantly improves performance on knowledge benchmarks such as MMLU and CMMLU, yields even greater gains on reasoning and code tasks including BBH, ARC, HumanEval, and MATH, and boosts long-context retrieval accuracy from 84.2% to 97.0%.

10 citations3 influentialRead paper

DeepSeek-OCR 2: Visual Causal Flow

Jan 28, 2026

This work proposes DeepEncoder V2, a novel vision-language architecture that addresses the limitations of conventional models which process images using a fixed raster-scan order, thereby failing to emulate human-like, semantics-driven visual perception—particularly in complex layouts where causal perceptual sequencing is essential. To overcome this, DeepEncoder V2 introduces causal reasoning into visual token ordering for the first time, employing a causal-flow-driven dynamic reordering mechanism coupled with learnable semantic sequence modeling. The framework constructs a two-stage cascaded one-dimensional causal structure designed to approximate genuine two-dimensional reasoning. By moving beyond fixed positional encodings, the method significantly enhances cognitive consistency in interpreting complex layout images. The authors release both code and model weights, establishing a new architectural paradigm for vision-language models.

2 citationsRead paper

Security Assessment of DeepSeek Harness with A.I.G: Evaluating Resistance to Indirect Prompt Injection

Aug 17, 2026

This study addresses indirect prompt injection risks in DeepSeek Harness by establishing a systematic security evaluation framework based on A.I.G. Integrating controlled taint tracking with a dual rule-semantic judgment mechanism (RuleJudge/LLMJudge), the research precisely quantifies the attack surface across over 10,000 test samples. Results reveal a maximum attack success rate of 25.5% in text and file modes, identifying critical gaps in sensitive operation protection. Beyond validating the limitations of existing defenses, this work proposes targeted control strategies, providing empirical evidence and methodological support for securing large language model applications against emerging threats.

0 citationsRead paper
Recent publications

Latest Papers

Security Assessment of DeepSeek Harness with A.I.G: Evaluating Resistance to Indirect Prompt Injection

Aug 17, 2026

This study addresses indirect prompt injection risks in DeepSeek Harness by establishing a systematic security evaluation framework based on A.I.G. Integrating controlled taint tracking with a dual rule-semantic judgment mechanism (RuleJudge/LLMJudge), the research precisely quantifies the attack surface across over 10,000 test samples. Results reveal a maximum attack success rate of 25.5% in text and file modes, identifying critical gaps in sensitive operation protection. Beyond validating the limitations of existing defenses, this work proposes targeted control strategies, providing empirical evidence and methodological support for securing large language model applications against emerging threats.

0 citationsRead paper

DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

Jul 06, 2026

This work addresses key limitations in speculative decoding, where parallel draft generation suffers from rapidly decaying acceptance rates due to the lack of inter-token dependencies, and fixed-length verification wastes batch capacity, hindering high-throughput inference. The authors propose DSpark, a novel framework that employs a semi-autoregressive drafter to model local dependencies within drafts, mitigating suffix decay. Furthermore, DSpark introduces the first dynamic verification scheduling mechanism, which adaptively adjusts verification length based on prefix survival probability and engine throughput characteristics. Experiments demonstrate that DSpark substantially increases offline accepted sequence lengths; deployed in the DeepSeek-V4 production system, it achieves a 60%–85% improvement in user-side generation speed, attaining previously unattainable performance levels under stringent interactive latency constraints.

0 citationsRead paper

Pointer-CAD: Unifying B-Rep and Command Sequences via Pointer-based Edges & Faces Selection

Mar 04, 2026

This work proposes Pointer-CAD, a novel framework that introduces a pointer mechanism into large language model (LLM)-driven CAD generation to address critical limitations in existing command-sequence-based approaches. Current methods struggle to support explicit selection of geometric entities—such as faces and edges—and suffer from topological errors due to discretization of continuous variables, hindering complex editing operations. Pointer-CAD overcomes these challenges by decomposing the modeling process into stepwise operations, where each step jointly leverages textual instructions and the preceding B-Rep model to explicitly select geometric entities via pointers, thereby guiding precise command generation. This approach unifies B-Rep geometry representation with command sequences, enabling robust entity interaction and substantially reducing quantization-induced topological and segmentation errors. Evaluated on a dataset of 575,000 samples, Pointer-CAD demonstrates superior capability in efficiently generating complex CAD structures compared to state-of-the-art methods.

0 citationsRead paper