Recently, DeepSeek has embarked on a hiring spree, with openings across multiple departments including algorithm, R&D, product, operations, data engineering, and administrative functions.
At the same time, the official version of DeepSeek V4 is scheduled for release in mid-month. In the author list of the earlier DeepSeek V4 paper, we discovered the name of Gu Yuxian (Yuxian Gu), a Tsinghua University class of 2021 Ph.D. student and recipient of the 2025 Special Scholarship for Graduate Students.
To our knowledge, Gu Yuxian has formally joined DeepSeek.
Gu Yuxian has also received the 2025 Apple Ph.D. Fellowship and the Ant In-Tech Scholarship.
"Algorithmic innovation becomes the key to breaking computational bottlenecks when hardware resources are constrained," said Tsinghua alumnus Gu Yuxian. He is a graduating Ph.D. student in the Department of Computer Science at Tsinghua University, having also completed his undergraduate studies there.
His personal homepage shows that Gu Yuxian studied in the Conversational AI (CoAI) group at Tsinghua University under the supervision of Professor Huang Minlie.
Personal homepage address: https://t1101675.github.io/
His research primarily focuses on improving efficiency throughout the entire lifecycle of large language models, covering key stages such as pre-training, downstream adaptation, and inference. His recent work has mainly progressed in three directions:
Pre-training Data Selection: Dedicated to constructing theory and algorithms to optimize the data selection process for training large language models, thereby training more powerful and efficient models. Representative work includes PDS, Instruction Pre-training, and Learning Law.
Knowledge Distillation in Model Compression: Designing new methods to effectively transfer knowledge from large models to smaller, more deployable models. Representative achievements in this direction include MiniLLM and MiniPLM.
Efficient Model Architecture: Exploring and designing new model architectures that improve performance while reducing computational costs. Related work includes Jet-Nemotron.
On his Google Scholar homepage, Gu Yuxian's paper citation count has reached nearly 5000, with two papers exceeding 1000 citations: "Pre-trained models: Past, present and future" and "MiniLLM: Knowledge distillation of large language models".
As the first author, Gu Yuxian has published papers multiple times at top international AI academic conferences such as NeurIPS, ICLR, and ACL.
Last year, Jiqizhixin reported on "Jet-Nemotron", a new series of hybrid-architecture language models that achieved SOTA full-attention model accuracy while demonstrating outstanding efficiency.
The core innovations of Jet-Nemotron are primarily reflected in the following two points:
Post Neural Architecture Search (PostNAS): An efficient pipeline for post-training architecture exploration and adaptation, applicable to any pre-trained Transformer model.
JetBlock: A novel linear attention module whose performance significantly outperforms previous designs like Mamba2.
Paper address: https://arxiv.org/pdf/2508.15884
At that time, the 2B version of Jet-Nemotron's performance could rival the latest SOTA open-source full-attention language models like Qwen3, Qwen2.5, Gemma3, and Llama3.2, while achieving significant efficiency gains. On H100 GPUs, its generation throughput achieved a speedup of up to 53.6x (with a context length of 256K and maximum batch size).
On the MMLU and MMLU-Pro benchmarks, Jet-Nemotron's accuracy also surpassed some MoE full-attention models, such as DeepSeek-V3-Small and Moonlight, despite those models having larger parameter scales.
Earlier in 2024, Gu Yuxian and his collaborators proposed a knowledge distillation method for distilling large language models into smaller language models. They first utilized reverse Kullback-Leibler divergence (KLD) to replace the forward KLD objective in standard knowledge distillation methods, then derived an effective optimization method to learn this objective.
They named the resulting student model "MiniLLM". Extensive experiments in instruction-following scenarios showed that compared to baseline methods, MiniLLM could generate more precise answers with higher overall quality, while also having lower exposure bias, better calibration capability, and stronger long-text generation performance.
Leading open-source communities and industrial platforms like Google, Alibaba, and NVIDIA have adopted this method.
Paper address: https://arxiv.org/pdf/2306.08543
We also look forward to Gu Yuxian bringing more new achievements in the next chapter of his career at "DeepSeek".
This article is from the WeChat public account "Jiqizhixin" (ID: almosthuman2014), author: Jiqizhixin focusing on AI talent.
